In [Two Cents #87 — Part I], we explored how the convergence of conversational AI and agentic code execution — embodied by architectures like OpenClaw — is producing something qualitatively new: an AI Concierge that can capture user intent, coordinate or create execution capabilities on the fly, and accumulate deep personalization over time. That was the what is happening.

In this Part II, I want to move from observation to projection. If a meaningful share of consumers adopts a real concierge agent — one that doesn’t just answer but operates — what changes in the structure of consumer software, distribution, and value capture? What existing profit pools get disrupted, and which new ones get created?

I’ll start with the structural framing that I believe matters most for understanding everything that follows (”The structure of the change ahead”), then will walk through three specific keywords — disruption of portals, app decomposition, device detachment, areas where the changes should be most consequential and where I’m focusing my investment attention. Then, the closing remarks on the future opportunity set for early‑stage founders and investors follows.

These are developing hypotheses, not declarations. My goal is not to fix a single “prediction” and defend it. It is to describe a set of structural forces shaping the surface on which the next generation of consumer AI products will be built—and to invite founders, operators, and fellow investors to refine or challenge these early theses.

I. The Structure of the Change Ahead

From “Attention Economy” to “Intent Economy”

As I discussed in [Two Cents #84], I believe there is a structural shift underway that deserves to be named clearly. For two decades, the consumer internet has operated on the logic of the Attention Economy. The entire ecosystem — Google Search, Facebook’s News Feed, TikTok’s algorithm, Amazon’s product rankings — was architected around a single fundamental constraint: we could never know exactly what a user wanted. We could only approximate it.

Website visits, clicks, dwell time, scroll depth, search keywords — all of these are “attention” signals that serve as the best available proxy for what the user actually intends to do. The entire $600B+ global digital advertising market, the SEO industry, the affiliate ecosystem — all of it was built on the logic of capturing eyeballs as an approximation of intent, then converting that approximated intent into economic value.

Hence, the “Attention Economy.”

AI changes this equation fundamentally. When a user tells an AI Assistant, “find me a family-friendly hotel in Kyoto for the first week of April, under $200 a night, with good vegetarian options nearby,” the assistant doesn’t need to approximate intent through click patterns. It has the intent — explicitly, precisely, in full context.

This is what I call the shift to the Intent Economy: AI-based environments introduce entities (ChatGPT, Claude, personal AI Assistants) that can explicitly capture user intent — not as a best approximation, but directly — and then execute on it. Cambridge researchers have described an emerging “intention economy” as a marketplace where human intentions become explicit commodities to be captured, interpreted, and acted upon in real time. Outlier Ventures calls it “a radical departure from our attention-centric internet,” where “platforms and services will now be designed to understand, interpret, and fulfill a user’s goals — often autonomously.”

This framing — Attention Economy → Intent Economy — is, I believe, the single most important lens for understanding every change I’ll describe below.

Who Captures Intent — and How?

If this framing is correct, then the most consequential question becomes: who captures the user’s intent, at what point, and through what mechanism?

Every major technology transition has been defined by the entity that seized the intent capture layer at the point of origin:

• Web era: Search engines captured intent at the query box (Google → $2T+)

• Mobile era: App stores captured intent at the download moment (Apple/Google)

• Social era: Feed algorithms captured intent through engagement proxies (Facebook/TikTok → ~$1T each)

In each case, the entity that controlled distribution — the chokepoint through which intent flowed on its way to fulfillment — accumulated the most economic value.

In the AI era, intent is expressed directly through prompts, voice commands, and contextual instructions. Whoever becomes the default interface where that intent first lands will control the new distribution.

This leads me to a set of questions I keep returning to:

First, if everyone has an AI Assistant like OpenClaw handling their needs end-to-end, do we still need portals and gateways? Do we still need Google as a default starting point? Amazon as a default shopping destination? The honest answer, I think, is: increasingly less so.

Second, if the AI Assistant itself becomes the primary intent capture point, where does value concentrate? Probably in the Assistant and the Context/Personalization Memory it manages. But here’s the nuance: that value may be distributed across millions of individual assistants, each holding its own user’s context. There’s no natural aggregation mechanism — which could mean we end up with a multipolar equilibrium rather than a single winner-take-all structure.

Third, when an AI Assistant captures intent, where does it route the request? Directly to category-specific providers? Or through a new intermediary? And how does the assistant choose — user specification, bidding, marketplace dynamics, algorithmic optimization? The answers to these questions will define the new economics of consumer distribution. Whoever processes this routing step sits at the most strategic chokepoint in the entire new value chain.

Two Candidate Structures for Intent Capture

When I pressure-test the future, I see two competing structural candidates:

Candidate 1: The AI Super App remains the portal. An entity like ChatGPT, Gemini, or a new entrant becomes the dominant portal — a WeChat-like super app that captures intent as the default destination. This remains possible. It was my first conclusion on AI Super App, until OpenClaw architecture emerged.

But I’d argue the emergence of the OpenClaw architecture has significantly reduced this probability, at least in Western markets. The open, decentralized nature of the concierge model makes it harder for any single app to monopolize the intent capture layer. A new WeChat-style AI Super App emerging from scratch gets especially less likely now.

Siri and Alexa are interesting edge cases — they have the advantage of being embedded in hardware ecosystems (iPhone for Siri, Amazon’s environment for Alexa). But if those ecosystems themselves diminish in relevance as AI assistants become more capable and autonomous, Siri and Alexa’s claim to Super App status weakens proportionally. The smartphone’s role in consumer intent capture remains genuinely unsettled — and how it evolves will determine whether device-native assistants hold their ground.

Candidate 2: The AI Assistant becomes the primary concierge — and therefore the primary intent capture entity. I find this more probable given OpenClaw-type architectures. In this model, the user’s personal AI Assistant captures intent and then finds other entities — agents, services, providers — to fulfill it through machine-to-machine interaction.

The modes of fulfillment of ‘users’ intent’ will likely be diverse: search (advertising economics), recommendation (affiliate economics), marketplace (agent-to-agent negotiation), bidding (auction economics). Multiple modes will probably coexist across different verticals and use cases. The key insight is that intent capture distributes rather than centralizes — many assistants, each with a user’s context, rather than one portal aggregating all traffic.

The Timeline: “Gradually, Then Suddenly”

I don’t expect this transition to happen overnight. Over the next few years, existing portals and AI Assistants will coexist. Google will defend with AI Overviews, already appearing on 30%+ of queries. Amazon will push features like Buy for Me. But user behavior is already shifting — OpenClaw has garnered over 145,000 GitHub stars, and Baidu recently embedded OpenClaw-type agents into its search app for 700 million users.

Within 5 years, the roles and value distribution among major players will have materially settled. And over a 10-year horizon, the traditional portal model may resemble BlackBerry — structurally sound, but generationally obsolete.

This is a typical “gradually, then suddenly” pattern. Google, Amazon, Facebook, and Roblox each needed 10–15 years to reach their current dominant positions. As an early-stage investor, the 5–10 year window is precisely where the power-law opportunities emerge — betting on new structural players before the market consensus recognizes the shift as inevitable.

With this structural backdrop in mind, let’s turn to three specific keywords I’m watching.

II. Three Keywords — What Might Happen?

Keyword 1: Portal Disruption and Value Capture Restructuring

Background

The first and perhaps most provocative question: if everyone has an AI Assistant, do consumers still need the portals they default to today? Google for information. Amazon for shopping. The App Store for software. Facebook and TikTok for social content.

The answer, I think, varies by the type of service — and the nuance matters considerably.

Transactional services (search, shopping, booking, utilities) are the most directly threatened. When a user can express their intent to an AI Assistant and have it executed end-to-end — comparing flight prices, booking hotels, purchasing goods — there is simply less reason to visit Google or Amazon as a starting point. Intent capture migrates from the portal to the assistant. The portal’s “traffic concentration” role erodes.

Personal data-based services (email, calendar, document management) represent the classic “personal data + tool → app” bundle. In a decomposed world, the personal data migrates to a user-owned personalization layer, and the tool function becomes something an AI Assistant handles directly. The key issue here is personal data portability — and the degree to which users can extricate their data from existing platforms.

Externality-data services — where the product is the externality data itself — are more durable. Social networks live squarely in this category: the social graph and the interactions around it are inherently collective, not easily replicated by a single user’s AI assistant. But even here, social graph portability could erode the moat over time. If your social graph and interaction history become exportable, the switching cost of moving to a new experience approaches zero — and the moats of Facebook and Instagram become structurally weaker.

What Might Happen: Value Capture Restructuring

The structural implication: value migrates from centralized portals toward a distributed ecosystem centered on AI Assistants and the personalization data they manage.[1]

The new value concentration points will likely be:

• The AI Assistant itself — as the primary entity interpreting and routing intent

• The Context/Personalization data layer — the memory and preferences that make the assistant increasingly useful over time (what Scott Belsky calls “personalization effects” as the new network effects)

• The Intent Routing layer — whoever intermediates the matching between user intent and service fulfillment

The players under the most direct pressure:

• Google (search advertising ~$200B): the most directly threatened, with an estimated $130B+ in purchase-intent-related revenue exposed to migration

• Amazon (search-based commerce): when AI agents begin to comparison-shop and purchase directly, the ~$55B in Amazon search advertising is at risk

• The SEO industry: facing the most fundamental disruption in its history. Discovery mechanisms are changing at the root — zero-click rates already exceed 65%, and the shift from SEO to GEO (Generative Engine Optimization) is underway.

• Existing Ad Networks: the foundational model shifts from Attention (impressions, clicks) to Intent (outcomes, fulfillment)

A few important caveats. I don’t think this collapses into a single new monopoly — particularly in the US market. The US consumer ecosystem is more fragmented than China’s. OS gatekeepers (Apple, Google, Microsoft) exert significant influence. Privacy and regulatory expectations are higher. The most likely outcome is a multipolar equilibrium: power-law outcomes within a multi-player structure rather than a single winner-take-all.

Underlying this entire restructuring is a question of state portability: can users take their data with them? Today, platform lock-in works because user state is trapped — your social graph in Instagram, your listening history in Spotify, your purchase patterns in Amazon. If Persistent State becomes separable (stored independently, exportable), then services can be re-implemented in real time by an AI Assistant using that state. The app stops being the locus of value; the state does. This is why DID (Decentralized Identity) and state portability infrastructure are not niche topics — they are the enabling technology for the entire value chain restructuring described above.

Investment Thesis Implications

• Intent Routing Layer infrastructure: the new strategic chokepoint, regardless of which macro scenario prevails. The entity that intermediates between intent and fulfillment captures distribution-level value.

• Personalization Infrastructure: personal data vaults, context stores, and state management protocols. “Personalization effects” replacing network effects as the primary retention driver.[8]

• DID/Identity infrastructure: the next-generation “login” standard — a foundational layer that enables everything from state portability to social graph export.

• GEO (Generative Engine Optimization): the successor to SEO — how services get discovered and selected by agents rather than humans.

• Most at risk: businesses structurally dependent on portal traffic — SEO agencies, affiliate networks, display ad-dependent publishers. The $300B+ value pool in Google’s search advertising and Amazon’s commerce search alone represents the magnitude of what’s in motion.

Keyword 2: “Death of App”, or Rather, App Decomposition, then Reassembly as Intent Capture & Distribution

Background

Today, an “app” is a bundled package: it captures your intent (through UI, search, navigation), delivers the function (features, integrations), and stores your personalization data (history, preferences, state) — all within a single unit.

In the Clawverse, this bundle breaks apart. An application decomposes into three independent layers:

In the old world, one app owned all three layers — and distribution was determined by two things: the app’s own delivery quality and the external discovery mechanism (Google search, App Store ranking). In the new world, each layer separates and can be owned by a different entity. The critical investment question becomes: which layer captures the most value?

My answer: the Intent Capture & Routing Layer. This is the new form of consumer app distribution — the AI-era equivalent of what Google Search and the App Store have been for the past two decades.

The precedent is consistent: in every technology transition, the entity that captured consumer intent at the point of origin accumulated the largest share of economic value. And the current market size at stake is staggering: Google’s advertising revenue (~$200B, of which roughly $130B relates to purchase intent), Amazon’s search-based advertising (~$55B), and App Store/Google Play revenue (~$200B). That’s a $400B+ market that gets structurally reorganized as intent migrates from search boxes and app stores to AI Assistants.

As Tom Tunguz has observed, the web is already flipping to agent-first design — “being findable” starts meaning “being usable by software acting for someone else,” not just ranking in human search.

What Might Happen: Three Directions for Delivery & Execution

The delivery layer — how user intent actually gets fulfilled — is where I see the most structural uncertainty. I expect a hybrid of three directions to coexist, with the relative weight of each still very much to be determined:

Direction 1: Direct routing. The AI Assistant identifies the optimal provider in each category and routes the request directly — essentially becoming an “AI Super App” that owns both intent capture and value accrual through its routing. If a new consumer platform emerges that hosts these assistants for mainstream users, it could function as a new kind of portal, reminiscent of the early web portal era.

Direction 2: Intermediated routing. A separate business may emerge to interpret and route intents — a new distribution layer analogous to what search was for the web and what app stores were for mobile. This is where I see the most interesting structural possibilities:

• 2-1: Single intermediary — one dominant router, like Google for search. Possible in theory, but I think unlikely to consolidate this way in the agent era.

• 2-2: Category-specific intermediaries — multiple routers, each serving a vertical. AI Assistants transact with these intermediaries through bidding or best-solution proposals. Value capture is distributed between the AI Assistant and multiple intermediaries. Google/Amazon-level concentration feels less likely; distributed value across numerous intermediaries feels more probable.

• 2-3: Headless providers — the AI Assistant, armed with personalization data, bypasses all intermediaries and goes directly to backend providers (e.g., for travel: bypassing Hotels.com and Kayak, going straight to GDS systems, hotel aggregators, local OTAs). Intermediary value capture vanishes almost entirely. This possibility was already demonstrated when GPT-4 placed orders directly with Instacart at launch. It will simply become widespread as AI assistants increasingly execute on users’ intent.

Direction 3: Direct execution via vibe coding. The AI Assistant doesn’t route at all — it builds the solution itself. Need a custom expense tracker? The assistant writes and runs the code. In this scenario, there is no third-party app capturing value whatsoever.

My expectation: Directions 1, 2, and 3 will all coexist in a hybrid structure. Direction 3 will handle simple, self-contained tasks. Direction 2 will handle complex, multi-party transactions. Direction 1 will emerge in verticals where a clear “best provider” exists. But the era of the app as a monolithic bundled unit is, I believe, structurally ending.

Investment Thesis Implications

• AI Concierge Platforms: consumer-facing platforms that make OpenClaw-style AI Assistants accessible to mainstream users. Multiple are already appearing, though I’m skeptical any single one evolves into a dominant Yahoo!/Google-style portal. More likely: multiple coexisting platforms with power-law distribution.

• Intent Routing / Marketplace Infrastructure (Directions 2-1, 2-2): the layer where agents find optimal services. I expect multiple mid-scale intermediaries rather than a single dominant player — more fragmented than today’s Google/Amazon duopoly, but still a collectively massive market.

• Existing transaction infrastructure rewrite: SEO → GEO, Ad Networks → Agent Bidding, affiliate networks → agent-to-agent negotiation. This points to a market that could match the scale of today’s directly exposed advertising revenue ($400B+) or potentially the global advertising market (~$1.5T, roughly 1.5% of global GDP which is the US historical average of the gross ad market to the GDP).

• Picks & Shovels: agent orchestration infrastructure, A2A (agent-to-agent) transaction rails — payment, trust, reputation, identity systems for the machine-to-machine economy. Tunguz predicts workers will manage ~50 agents daily; the coordination and transactional infrastructure this requires is a market unto itself.

• Most threatened: standalone apps and single-function SaaS (decomposed into Plug-ins/Skills, heavily commoditized); App Stores (role shrinks from primary distribution to a single category of intermediary — a “Skill Store” with far lower value capture); and portal-dependent businesses whose discovery mechanisms are being fundamentally rewritten.

Keyword 3: Device Detachment and the New “Consumer OS”

Background

Today, your digital life is bound to your device. iPhone means Apple’s ecosystem. Android means Google’s. This coupling — hardware + OS + defaults + distribution — is the foundation of incumbent power. And the degree to which it persists or dissolves is perhaps the single most important variable determining who wins Consumer AI.

The logic is straightforward: as the AI Assistant becomes the user’s primary interface for expressing and executing intent, the need for specific apps, files, and desktops diminishes. If you don’t need specific apps, you don’t strictly need a specific device. This opens the possibility of device detachment — the user’s primary digital relationship shifting from a device to an AI Assistant that lives in the cloud.

In a device-bound scenario, the smartphone remains the default AI touchpoint. Apple’s devices would become the default AI Assistant provider — Siri evolves into Apple Intelligence, and Apple controls intent capture through OS-level integration, background execution permissions, and default settings.

In a device-detached scenario, the user’s personal AI Assistant becomes the primary interface — accessible from any screen, anywhere. The AI becomes the de facto “Consumer OS,” and the device becomes a commodity viewport.

OpenClaw is an early instantiation of this device-detached vision: chatbot conversation + vibe coding + agent execution happening through messaging apps and cloud interfaces, independent of any particular hardware or OS. This aligns with Belsky’s argument that “markets will be won or lost at the interface layer, and the ultimate interfaces are ULTIMATELY controlled by the operating systems of our lives” — except that in the AI era, the “operating system of our lives” may no longer be iOS or Android. It may be the AI Assistant itself.

What Might Happen: Three Options for the User-AI Interface

I think about this along a spectrum:

Option 1: Device-bound (status quo trajectory). Users keep personal devices (smartphone, forthcoming AR glasses) as the primary AI interface. Voice + screen UX. Apple’s current devices become the default AI Assistant provider. This is the most easily predictable scenario and the one incumbents are betting on.

Option 2: Voice-primary hybrid (most likely in the 5–10 year window). Voice becomes the main interaction modality; the smartphone screen is consulted only when needed for visual confirmation or complex decisions. This is the scenario I find most probable over the medium term. Its implications are significant: the role of both the device and the OS shrinks. The AI Assistant becomes the user’s primary touchpoint — what I’d call the de facto “Consumer OS.” User loyalty begins migrating from device brand to AI Assistant brand.

Option 3: Full device detachment (long-term possibility). No personal device required. A personal/ambient microphone for input. Any available screen for output. Identity-based access (most likely, DID) replaces device-based access. The user walks up to any screen, authenticates via identity, and their full AI environment loads. The device is a “viewer” — the intelligence and state live in the cloud.

If device detachment progresses, the user’s personal data stack may reorganize into distinct layers:

• Layer 1 — Identity: who I am, portable and globally recognizable (DID). This becomes the access key that replaces device ownership.

• Layer 2 — Personal State: my files (email, calendar, messages, social graph), my transaction history (shopping, payments), my preferences and context (the personalization layer).

• Layer 3 — Dashboard: the real-time view of where my intents are being processed, their status and history. All my data on cloud/personal servers, accessible from any authenticated interface.

This is the dashboard-centric UX described in Part I — but with the critical addition that it is decoupled from any specific device or operating system.

Investment Thesis Implications

• Two-school bet: Device-bound (Apple ecosystem remains dominant, AI integrated at the OS level) and Device-detached (third-party AI Assistants win the relationship). I’d maintain exposure to both, with increasing allocation toward device-detached as signals of voice-primary interaction and cloud-native AI dashboards accumulate.

• “AI devices” — the new battleground: New form factors (AR/XR, ambient computing devices, wearables) where existing OS lock-in doesn’t apply. These represent greenfield territory — Meta’s AR explorations, startup hardware experiments, and the various “AI pin” attempts are all probing this space. As Belsky notes, “new approaches like Meta’s AR explorations” are emerging as alternatives to today’s reigning operating systems.

• Ambient Computing infrastructure: the infrastructure enabling device-agnostic access — cloud-native dashboards, cross-device seamless experiences, ambient authentication, and the identity layer that makes it all work.

• The key metric to watch: the ratio of AI Assistant usage time vs. native OS app usage time. When this ratio begins to tip decisively — when users spend more time in their AI Assistant than in traditional apps — the structural power shifts from device-maker to AI-provider. OpenClaw’s viral adoption through messaging apps (WhatsApp, Telegram, Signal) rather than through a dedicated native app is an early signal of this shift.

Closing Thoughts

These three keywords — Portal Disruption & Value Capture, App Decomposition & Reassembly, and Device Detachment & Consumer OS — are deeply interconnected and mutually reinforcing. Portal disruption creates space for new intent capture entities. App decomposition enables AI Assistants to assemble capabilities dynamically. Device detachment amplifies the AI Assistant’s role as the primary consumer interface. And identity/state portability is the enabling infrastructure that makes all of it possible.

I want to be honest about what I’m most uncertain about. The sequencing is unclear — does app decomposition drive device detachment, or vice versa? The speed is uncertain — will this take 5 years or 15? The market structure could go multiple ways — concentrated or distributed, platform-dominated or infrastructure-dominated.

What I feel more confident about is the direction. The value chain of consumer software is unbundling. Intent capture is becoming the new distribution. And the relationship between user and digital service is being mediated by an AI layer that barely existed two years ago. For an early-stage investor operating on a 5–10 year power-law horizon, the opportunity set here feels structurally as significant as the web-to-mobile transition — potentially more so, because it touches not just the application layer but the distribution, identity, and infrastructure layers simultaneously.

As mentioned earlier, these are developing hypotheses. I’d love to hear from founders building in these spaces, investors with competing frameworks, and anyone who spots the structural blind spots I’m surely missing.

The thinking gets sharper when it gets challenged.

Something is happening in the AI consumer space that I think deserves more attention than it’s getting. It’s not about a single product launch or a benchmark score. It’s about a structural shift in how consumers interact with software — and, potentially, in how the entire consumer internet gets reorganized.

I’ve been spending the past several weeks thinking through what it means, and I wanted to share some observations and a developing framework. I don’t claim to have all the answers — this is very much a work in progress — but I believe the pattern is significant enough to put out there and invite conversation.

Let me start with what triggered this line of thinking.

I. What’s Happening: Coding Agents Cross a Tipping Point

Coding agents have gotten powerful. That’s not news. But the degree of improvement we’ve seen in recent months has been, I think, genuinely underappreciated.

With the latest generation of frontier models and the maturation of agentic coding tools, we’ve crossed what feels like a meaningful threshold. Coding agents are no longer generating toy apps or stitching together fragile prototypes. They are building real software — CUDA-compatible packages for AMD, C compilers, complex full-stack applications like Claude Cowork — often in a single shot, without human intervention.

The term “vibe coding” was coined to capture a certain playful, experimental spirit of AI-assisted development. I’d argue the “vibe” part no longer applies. What these agents produce is production-grade. And if you extrapolate even modestly — as Elon Musk has suggested — you can envision a near future where coding agents bypass high-level languages entirely and generate executable binaries directly.

That, by itself, is a fascinating development for the developer ecosystem. But the thing that caught my attention is what OpenClaw represents.

OpenClaw takes this coding agent capability and wraps it in a conversational interface that anyone can use. It connects to messaging platforms consumers already live in. It runs continuously. It supports a growing ecosystem of community-created Skills. And it’s model-agnostic — working across frontier models from multiple providers.

Here’s what I find most notable: OpenClaw isn’t a developer tool. It’s a consumer agent. A non-technical user can, through ordinary conversation, instruct it to automate workflows, manage schedules, track expenses — and, critically, create entirely new software functions on the fly to fulfill whatever the user needs.

Think about what that means. Two capabilities that have historically existed in completely separate worlds are converging into a single interface:

• Conversational AI (the ChatGPT paradigm — analysis, writing, Q&A)

• Agentic code execution (the Claude Code paradigm — building and running software)

When you merge these two in a consumer-facing wrapper, you get something qualitatively new:

Full-fledged software creation capability, combined with concierge-level coordination, delivered directly to everyday users.

I would be calling this emerging world the Clawverse — not because OpenClaw is the winner (it may not be), but because it is one of the clearest early signals of where we’re heading.

II. OpenClaw, A New Architecture for “AI Concierge”

When I look at OpenClaw (and the architectures it represents), I don’t see “a better chatbot” or “another AI wrapper.” I see the emergence of what might be described as an AI Concierge — a persistent, intelligent layer that positions itself between the user and the entirety of the digital services landscape.

The structure matters. The AI Concierge serves as a central hub that performs three distinct functions:

1. Intent Capture

The concierge becomes the first place the user’s intention lands. Whether someone wants to book a trip, compare insurance options, track a package, or build a custom analytics dashboard — the intent is expressed to the concierge first. It acts as the portal reading all of the user’s intent — triaging what the user wants and deciding how to route it.

This is meaningfully different from how intent has historically been captured. A user’s intent has historically been captured as the best available approximation by analyzing signals of “user attention”—including website visits, clicks, dwell time, and search queries—within what has been termed the “Attention Economy.”

The concierge consolidates this by directly capturing the user’s intent.

2. Execution Coordination

Once it understands the intent, the concierge decides how to fulfill it. This is where it gets interesting. It can:

• Call an external service or API

• Invoke a pre-built Skill from the community ecosystem

• Chain multiple tools together

• Or — the genuinely novel part — generate a new function on the spot by writing and executing code

As Tom Tunguz captures this dynamic well: “Skills are programs written in English. They tell an agent how to accomplish a task: which APIs to call, what format to use, how to handle edge cases. A skill transforms an agent from a conversationalist into an operator.”

The concierge doesn’t just talk. It does.

3. Personalization & Memory

This is perhaps the most strategically consequential function, and the one I think is most underestimated. Every interaction — every decision, preference, transaction, and outcome — accumulates as a persistent data layer. Payment methods, shipping addresses, dietary restrictions, communication preferences, past transaction records and results.

Still limited today, but the future trajectory is clear: the concierge progressively becomes the repository of personalization data — not just what you asked for, but what you chose, what you rejected, and what worked.

The Strategic Implication

When you put these three functions together — intent capture, execution coordination, and personalization accumulation — the concierge begins to look like something quite powerful: the entity that controls the most value capture in the user’s digital life.

I’m not making a prediction that any single product will achieve this position. But architecturally, the pattern seems clear: an AI assistant that can read your intent, coordinate (or create) the capability to fulfill it, and learn from the outcome positions itself as the central hub of the consumer digital experience.

From the user’s perspective, the consequence is straightforward: why would I go looking for a specific app when the concierge already understands what I want and can either find or build the solution?

III. The “Death of Apps” — Or Rather, Their Reassembly

I want to resist the clickbait framing here. Apps aren’t going to vanish overnight. But I do think the concept of the app as the primary unit of consumer software is being structurally challenged in ways that are worth examining.

What Apps Actually Were

To see why, it helps to step back and think about what an “app” has always been: a bundled package of specific functionality and UX, designed to standardize a workflow in a particular domain. Order food. Edit photos. Manage email. Track fitness.

This bundling was a solution to a real problem. Users got efficiency and learnability. But the trade-off was: personalization was sacrificed for standardization, and user data became siloed inside each app. Your social graph lives in Instagram. Your purchase history lives in Amazon. Your learning progress lives in Duolingo. These data silos became the foundations of moats — lock-in, network effects, system-of-record gravity.

The UX Arc That Got Us Here

It’s worth tracing the evolution historically:

• PC era: The desktop metaphor. Files and folders were the primary objects; applications were tools you applied to those files.

• Web/Mobile era: Files disappeared. Apps became the first-class entities. Data moved inside each app and became invisible to users — siloed, proprietary, non-portable. Your phone became a grid of icons, each containing a walled garden of functionality and data.

This represented a reasonable equilibrium, as it freed users from having to build custom software to meet these needs, while requiring trade-offs in data portability and UX personalization.

But it was an equilibrium built on a specific constraint: the user had to find, download, learn, and manage each app individually. The app was the best available packaging unit for delivering digital value to consumers.

What Changes in the Clawverse

In a world where an AI Concierge can interpret your intent, assemble capabilities dynamically (through pre-built Skills, Plugins, or on-demand code generation), and access your personalization data in a portable layer — the app bundle may no longer be the optimal packaging unit.

The “death of apps” is less about software disappearing and more about the packaging unit of software value changing. A consumer might still use Uber (the service), but not “open Uber” (the app) as often. They might still pay for Spotify (the catalog + rights + recommendations), but discover and play through the concierge. The assistant becomes the default beginning of work.

IV. A New Architecture to Deliver “Consumer Intent”

If the app bundle is being pulled apart, what are the constituent pieces? This is where I’ve found it useful to think about a simple decomposition.

When you strip a consumer app down to its essentials, you get three components:

1. Intent Capture

How does the system learn what the user wants?

• Directly — through conversational UX, guided flows, dashboards

• Indirectly — through search/SEO (Google has historically monetized through the capture of user intent with the search keywords as its approximation.)

2. Delivery (Execution)

The actual work: purchasing, booking, scheduling, analyzing, messaging, paying. This is the core function of the app — the thing it does.

3. State & Memory

This is where things get nuanced. There are several layers of “state” that apps accumulate:

The key insight from this decomposition: different consumer services are defended by different types of state.

Transactional services — e-commerce, food delivery, travel booking — have relatively shallow state requirements. If a concierge can capture intent, compare options, and execute with your stored preferences, the branded app experience becomes… optional. The competitive axis shifts from “best app UX” to best supply, best price, fastest fulfillment.

Persistent-state services — social networks, education, therapy, long-running creative tools — are harder to disintermediate because the state is the product. Your social graph, your learning trajectory, your therapeutic history. These moats are deeper.

But here’s the question that keeps nagging at me: what happens when that state becomes portable? When a user can export their social graph, their learning progress, their preference history — and bring it to a new interface assembled by a concierge?

If state becomes portable, then even deep moats start to weaken. And the defensible asset shifts from “I hold your data” to “I consistently deliver the best outcome.”

I don’t think we’re there yet. But I think the direction is worth watching carefully.

V. Where the Moat Moves

If the above decomposition is directionally correct, it suggests some shifts in what “winning” means for consumer AI products — and I’ll share my current thinking, though I’d have to admit these are hypotheses, not conclusions.

The premium interface becomes intent-first. Historically, consumer products differentiated through UI polish, onboarding, and habit loops. In an intent-first world, the interface becomes less “screens you navigate” and more “intent you express,” plus a thin layer of confirmation and transparency. UX becomes more important, not less — but the target changes from feature discoverability to outcome reliability.

The moat may shift from UI to state + trust + access. If the concierge can route to multiple providers, then “being the front-end” alone is less defensible. Potential moats shift toward:

• State: depth and portability of personalization and history

• Trust: guarantees, safety, compliance, dispute resolution

• Access: exclusive supply, proprietary data rights, privileged integrations

Consumer AI winners may be the ones who combine execution quality with trusted custody of state, rather than simply shipping another beautiful app.

Disruption will come at different speeds. Transactional services face earlier exposure. Persistent-state services get disrupted differently, and potentially later. The key variable in both cases is state portability — the degree to which identity and context can move with the user rather than remaining captive to a single provider.

VI. A Closing Thought: From Attention to Intent

For the last twenty years, consumer internet economics have run on the Attention Economy: capture eyeballs, approximate intent from behavior, monetize through ads, ranking, or conversion.

In the Clawverse, the assistant can capture intent directly: “Find a flight.” “Order dinner.” “Plan my week.” “Help me learn this concept.” “Talk me through this decision.”

When intent is explicit, the economic locus shifts:

The system that captures intent becomes the new distribution layer.

This is why consumer AI assistants are not merely a product category — they are potentially a new platform layer. But I want to be careful: I don’t think we should assume a single “winner assistant.” Open-source approaches, model-agnostic architectures, and consumer preferences around privacy could lead to a more fragmented landscape than the mobile app era.

So my working questions are less “who wins?” and more:

• Where does intent land first?

• Where does state live?

• Who is trusted to execute?

• What becomes scarce when software is abundant?

The economics of the Intent Economy are still forming. And I’d genuinely love to hear how others are thinking about these dynamics — especially those who are building consumer products at the frontier, or who think this entire Clawverse framing is directionally wrong. The best insights on emerging platform shifts have always come from practitioners, and I’d welcome any thoughts, pushback, or pattern-matching you’re willing to share.

There will be Part II, which would look into what might happen in the near future with respect to a few keywords on the changes in the Consumer UX.

One of the most important battlegrounds in Consumer AI is quietly forming around the race to become the AI Super App.

This isn’t a contest for the best model, the smartest assistant, or the slickest UI. It’s a structural competition over where consumer intent is captured, interpreted, and executed in the AI era.

We’re already seeing the race unfold across multiple fronts: AI assistants (ChatGPT, Gemini, Claude, Perplexity), AI-native browsers (Atlas, Comet, Dia), agentic platforms (Manus, Genspark), OS-level intelligence from Apple and Google, and new entrants that may emerge with entirely new form factors. It’s still early, but the winner of this race has a real chance to become the dominant consumer platform of the AI era—what people will call the “Google of the AI age.”

China offers a useful precedent. Over the past decade, it already lived through a prolonged battle for mobile Super App dominance, which ultimately converged into a structure where a small number of platforms controlled most consumer time, transactions, and value creation. Today, China is entering a second platform war—this time for the AI Super App.

The U.S. will not replicate China’s outcome 1:1. But the economic logic, competitive behaviors, and distribution dynamics observed in China are highly instructive for understanding how the AI Super App race in the U.S. is likely to play out.

Why the AI Super App matters

The core function of an AI Super App is intent capture. That’s where the largest value capture tends to happen: the platform that understands intent most accurately—and responds most effectively—ends up owning the downstream flows.

History supports this. In every major computing shift, the platform that captured intent earliest captured disproportionate economic value:

• Web era: search engines became the starting point for information and commerce.

• Mobile era: app stores controlled distribution and monetization by controlling access to apps.

• Social era: feed-based recommendation systems (Facebook, TikTok) captured attention at scale—and each translated that control into ~$1T outcomes.

As discussed as the transition from the “Attention Economy” to the “Intent Economy” in [Two Cents #84] , the internet and mobile ecosystems historically struggled to understand user intent directly. The industry therefore relied on the best available proxy: the flow of user attention. Metrics like traffic, CTR, conversion, and A/B testing became the primitives, and the resulting businesses were ad networks, affiliate networks, and performance marketing—collectively, the Attention Economy.

AI changes the form factor of intent expression itself.

Users no longer communicate intent indirectly through keywords, dwell time, or click behavior. They state it directly through prompts and context: “Plan a five-day trip to X with a budget of Y,” or “Recommend a dress for a weekend party within $Z that matches this vibe.”

The interface that becomes the default starting point for these requests gains leverage over everything downstream—services, workflows, commerce, and monetization.

That’s why the AI Super App can become a $1T+ consumer opportunity. It sits above vertical services as a controlling layer: it intermediates across categories, routes execution, and effectively collects a toll on the value created along the way. China’s Super App history already shows that once this layer consolidates, value concentration accelerates quickly.

How China built the Super App and Mini-App ecosystem

China’s Super App ecosystem didn’t appear overnight from a single breakout product.

It emerged as platforms combined (1) massive consumer distribution, (2) durable consumer connection via social channels, and (3) integrated payments—then used those ingredients to internalize the entire “intent flow,” from intent expression to delivery, inside the Super App container.

WeChat is the canonical example. It began as a messaging product, then expanded step-by-step into content (Official Accounts), payments (WeChat Pay), and offline acceptance via QR codes. In 2017, Tencent formalized what had been forming organically by introducing Mini Programs—lightweight “no-install” apps running inside WeChat—turning these components into a coherent platform.

Precursors existed before Mini Programs: HTML5 pages in in-app browsers, QR-driven O2O flows, and service accounts that could interact with users. Mini Programs unified those pieces into a standardized platform, allowing WeChat to internalize the consumer intent flow—and capture most of the value created along that flow.

The result was unambiguous. As of 2024:

• ~936M MAU engaging with WeChat Mini Programs

• 1M+ Mini Programs live

• RMB 2T+ quarterly GMV flowing through the ecosystem (>$1T annualized)

Mini Programs didn’t just complement the app ecosystem; they redefined it. Many services no longer needed standalone apps. Customer acquisition costs dropped sharply. WeChat evolved from a portal into a massive closed-loop transaction infrastructure.

Other Chinese platforms—Alipay, Douyin (ByteDance), Meituan—replicated the model in their own domains. The end state was not a single monopoly, but a multi-polar oligopoly: a small number of Super Apps, each anchored in a core consumer habit, each internalizing the full intent flow within its sphere.

Common success factors among China’s Super App winners

China’s Super App winners didn’t win by packing in more features. They won by positioning themselves at the earliest point of consumer intent, then pulling the entire downstream intent flow inside the platform. This created a compounding structural advantage: value creation and value capture strengthened over time.

1. Owning distribution

Chinese Super Apps became the default entry points for everyday actions—messaging, paying, scanning, searching, consuming content—effectively controlling the top of the funnel across categories. The critical point wasn’t any single feature; it was establishing a default behavior: “If I want to do something, I open this app first.”

WeChat illustrates this clearly. By stacking payments, content, offline QR flows, and service access on top of messaging, it reduced the need for users to leave the platform. Once “the service lives inside WeChat” became internalized, WeChat’s coverage and value capture expanded rapidly.

At the ecosystem level, this created a structure where new services could not bypass the Super App. Consumer access costs (CAC) converged inside the container, and standalone apps became increasingly disadvantaged in marketing and distribution. The Super App became the market’s gate.

2. Embedding trust anchors

The decisive factor that allowed Super Apps to move beyond discovery into execution was embedding trust infrastructure—especially payments and identity. Payments aren’t just a feature; they are the anchor that connects intent to transaction with minimal friction.

WeChat Pay and Alipay tied together real-name identity, financial accounts, and social context. Users could transact immediately without needing to rebuild trust for each new service. Conversion speed inside the Super App accelerated dramatically.

This collapsed the boundary between “finding” and “doing.” Search, recommendation, and comparison increasingly led directly to payment. The Super App stopped being a traffic intermediary and became the transaction infrastructure itself—leaving far less room for value to remain outside the container.

3. Minimizing developer and merchant friction

Mini Programs were the ecosystem’s key structural innovation. The point wasn’t just new functionality; it was removing friction across distribution, installation, login, and payment.

Developers and merchants no longer needed to fight app-store review processes, downloads, updates, or re-engagement challenges. They could ship instantly inside WeChat’s runtime and leverage built-in distribution through search, QR codes, sharing, and location-based discovery.

This created an ecosystem where long-tail services could exist economically. Many services that couldn’t justify standalone apps could thrive inside the Super App, while the platform strengthened its value capture by becoming the infrastructure layer for all activity.

4. Maximizing habit stacking

Perhaps the most powerful weapon was that Super Apps did not try to invent new behaviors. They layered minimal additional actions on top of existing habits.

Messaging flowed into payment; payment flowed into booking; booking flowed into content—naturally, without users feeling like they were learning “new products.” The Super App expanded the user’s behavioral surface area with very little incremental cognitive cost.

Over time, the Super App became less like an application and more like an OS for daily life. The lock-in was not to a single service, but to the platform itself—and that lock-in strengthened as the ecosystem grew.

The outcome: a power-law oligopoly, not a single monopoly

These forces did not produce a single winner-take-all monopoly. They produced a power-law oligopoly, where a small number of Super Apps (WeChat, Alipay, Douyin, Meituan) captured most consumer value, each anchored around a core habit.

The key lesson is that the battle was not feature-driven. The winners were the platforms that owned the intent entry point and internalized the entire intent flow—creating structural advantages that compounded over time. The same framework applies directly to the AI Super App era.

Act II: China’s AI Super App war

China is now replaying this dynamic in the AI era.

The dominant Super Apps are moving aggressively to become AI Super Apps as well. Incumbents like Tencent are embedding AI directly into WeChat’s core surfaces—search, chat, and workflows—sometimes integrating third-party models. It’s the classic incumbent play: defend the container, modernize the surface.

At the same time, challengers such as ByteDance are attacking more aggressively. Doubao is leveraging Douyin’s attention graph and creator ecosystem to open a new AI front door, while expanding into devices to capture consumer intent more directly. Competition is already turning hostile—e.g., ByteDance’s mobile AI agent app reportedly faced rapid blocking by platforms like WeChat and Taobao within days of launch.

Once again, the same principle holds: the AI Super App race is being driven less by “AI capability” and more by control of distribution.

The AI Super App race in the U.S.

What’s similar to China

Despite surface differences, the underlying economics are similar. In both markets, the central prize is intent capture at scale. If you control intent, you sit upstream of decisions and actions—and gain leverage over everything downstream. That leverage is what translates into outsized value capture.

Both markets also tend to converge toward oligopoly at the consumer layer. Even if many services exist beneath, durable moats are built at the top—through ecosystems that form around the dominant container.

For that reason, the AI Super App race in Consumer AI is plausibly a $1T competition, comparable in scale to what Google represented in the web era.

The structural differences that matter

The U.S., however, is structurally different in ways that materially shape how the market can form.

China entered the AI era with entrenched Super Apps controlling messaging, payments, and identity. The U.S. consumer ecosystem is more fragmented—payments, messaging, and commerce are split across many players.

As a result, OS-level gatekeepers (Apple, Google, Microsoft) exert far more influence. Defaults, permissions, and background execution are powerful constraints on how an AI Super App can emerge. Apple’s App Store, Siri, and default search placements are obvious choke points. Even Safari’s default search placement alone represents meaningful economic value (on the order of tens of billions in revenue-equivalent economics over time).

Privacy and regulatory expectations are also higher. That can slow automation, but it increases the value of platforms that can provide trust, auditability, and governance.

These factors make a single WeChat-style monopoly unlikely in the U.S. But they do not eliminate power-law outcomes. Instead, they push the market toward a multipolar structure, where dominance is distributed across contexts: OS surfaces (Siri), default search boxes, app stores, AI assistants, AI browsers, and more.

Signals the U.S. is entering its “mini-app moment”

Recent moves suggest the U.S. is entering its own version of a mini-app moment.

With the introduction of an App SDK and app-store-like distribution, ChatGPT is evolving from an assistant into a platform. Natural-language intent can be routed into third-party execution layers, while discovery and monetization remain controlled upstream by the platform.

The expansion of skills and agent libraries in systems like Claude and ChatGPT points in the same direction. Over time, these modules could function like Mini Programs—allowing AI platforms to aggregate and monetize specialized tasks.

Separately, AI-native browsers (Atlas, Comet, Dia) and agentic platforms (Manus, Genspark) are attempting to capture intent at the moment of navigation. By embedding agents into browsing workflows, they challenge the browser as the primary unit of interaction.

Across these efforts, the direction is consistent:

Own the layer where intent is expressed and action begins.

Potential scenarios for the U.S. AI Super App market

The U.S. market may evolve through several paths:

• AI assistant–first platforms (ChatGPT, etc.): become the default interface and capture value via intent routing and app marketplaces

• OS-native AI (Apple, Google): dominate through defaults and system permissions and extract rents from AI workflows

• Social AI (Meta): capture intent through social context inside feeds and messaging, monetizing via creators and commerce

• Enterprise AI spillover (Microsoft): control high-ARPU intent through workflows, identity, and governance

Below is a deeper look at each.

1. Assistant-first platforms (ChatGPT, etc.)

The core value capture is intent interpretation + intent routing.

Natural-language requests become the transaction starting point, and the layer that decides which service/agent/tool fulfills that intent gains economic power. Where the App Store controlled app distribution, the AI assistant controls intent distribution.

If this path becomes dominant, the market forms as an AI-native agent/app marketplace. Vertical agents and tools compete for discovery and execution inside the assistant. The platform captures take rates through recommendation logic, distribution control, payments, and identity. Structurally, it resembles a unified layer combining search, app distribution, and payments.

Strengths: fast iteration, model-UX coupling, and relative platform neutrality.

Weaknesses: distribution fragility—lack of direct control over OS/browser defaults. The biggest difference vs. China is that payments, identity, and daily-life infrastructure are not pre-integrated, so the product may look more like a “Super Gateway” than a true WeChat-style Super App.

If this scenario leads, Consumer AI becomes agent-first and API-first. Standalone consumer brands weaken, and execution agents capture value downstream while the assistant controls routing upstream.

Key contenders: OpenAI (ChatGPT), Google (Gemini), Anthropic (Claude). The most likely leader is OpenAI, but OS dependence remains a persistent risk.

2. OS-native AI (Apple, Google)

The value capture here is defaults + system permissions + background execution.

Intent is detected and acted on at the OS layer, often without explicit user initiation. The economic outcome is less about ads and more about extracting workflow rents—similar in spirit to app-store economics.

In this world, the “Super App” is not a separate app; the OS becomes the Super App. Siri/Gemini are interfaces, but the real advantage is control over notifications, permissions, sensors, payments, and identity. The market forms as an AI-enabled workflow layer over the existing app ecosystem.

Strengths: unmatched distribution and lock-in.

Weaknesses: slower innovation and constrained openness; stronger regulatory and privacy pressure than China. Apple, in particular, may face limitations on how directly it can push into commerce-like value capture.

If this scenario dominates, Consumer AI becomes OS-centric and incumbent-friendly. New entrants become extensions, plug-ins, or infrastructure providers rather than top-layer interfaces.

Key contenders: Apple, Google. Long-term, Apple may have the strongest position—but the outcome is closer to an “AI-powered OS” than a classic AI Super App.

3. Social AI (Meta, TikTok)

The value capture is intent capture through social context.

Many consumer desires are socially formed: taste, identity, validation, and trend-following. Platforms that sit inside social graphs can capture intent earlier and more richly than systems that only see explicit prompts.

An AI Super App here likely manifests as ambient AI inside feeds and messaging: the system proactively suggests, curates, and routes commerce through creators and advertising. The market becomes a unified arena where creators, brands, and AI agents compete inside the same attention surface.

Strengths: massive social graphs and proven ad monetization engines.

Weaknesses: trust and execution. China’s Douyin integrated deeper into payments, logistics, and O2O; Meta is stronger in discovery and influence than end-to-end execution. Regulatory exposure is also significant.

If this scenario leads, Consumer AI evolves into influence-driven commerce + AI recommendation, strongest in low- to mid-consideration decisions rather than high-stakes, high-complexity tasks.

Key contenders: Meta, TikTok. This is less a “Super App” than an “AI-augmented social OS.”

4. Enterprise AI spillover (Microsoft)

The value capture is workflow, identity, and governance.

This path captures intent not from consumer leisure but from work. Copilot may appear consumer-like, but enterprise contracts remain the economic center.

The “Super App” here becomes a work-life crossover platform. Habits formed in Outlook, Teams, and Windows extend into personal contexts, sustaining a high-ARPU structure. The market forms as a SaaS + consumer hybrid.

Strengths: durable revenue, enterprise lock-in, and strong governance/audit capabilities.

Weaknesses: weaker consumer-native distribution and cultural gravity. China has no true analogue here—this is a distinctly U.S.-shaped pathway.

If this scenario strengthens, Consumer AI becomes a two-tier market: a high-ARPU enterprise-driven layer and a separate, lower-ARPU pure consumer layer.

Key contender and likely winner: Microsoft—though it’s unlikely to become the universal consumer interface.

Multipolar equilibrium most likely

The most likely end state is not a single universal winner but a multipolar equilibrium, where power-law dominance emerges within contexts.

ChatGPT could own intent routing; Apple/Google could control execution infrastructure; Meta could dominate social discovery; Microsoft could control high-value workflows. The “AI Super App” outcome may be a set of super layers rather than one app.

The AI Super App race: a new $1T opportunity

The AI Super App race is not fundamentally about building “better AI.” It’s about owning the default interface where intent is captured and converted into execution. This is not a product competition—it’s a market-structure competition that rewires where traffic and revenue flow.

The winners won’t just dominate a single vertical. They will control the top layer that interprets intent, routes execution, and captures economics across categories. In that sense, the AI Super App is less a category and more a defining layer that will shape how value accrues in Consumer AI over the next decade.

China’s Super App history makes one point clear: once a platform becomes the default destination for consumer intent, discovery, payments, and supply all get pulled into the container—and value capture consolidates faster than most expect.

The U.S. is unlikely to converge into a single WeChat-like monopoly. But the same economic gravity still applies: intent capture drives value capture. The structure may differ, but the rule remains.

The key implication is that founders and investors must focus less on surface-level products and more on market structure. The right question is not “Who has the best demo?” but “Where will the choke points and toll booths form in the intent flow?”

For founders, the opportunity is not necessarily to build the AI Super App directly, but to find the indispensable positions in the stack—where intent must pass, where execution must be verified, where trust must be anchored. For investors, the bet is less about short-term traction and more about the layers that compound value over time.

Every major technology wave includes a moment where control of the consumer layer is renegotiated. Web search produced Google. The AI Super App race is that moment for AI—and it will define where the next $1T of consumer value is created and captured.

We are still in the phase where standards and rules are being written. The goal isn’t to declare a single “answer,” but to understand the direction of change and explore new opportunities along that direction.

I hope this framework helps as a starting point for imagining—and defining—meaningful opportunities in the next Consumer AI cycle.

Human needs don’t fundamentally change. What changes is how those needs are served—the delivery mechanism, the form factor, the defaults, and the surrounding environment. When the delivery mechanism shifts, consumer behavior shifts with it—often gradually, then suddenly. And when consumer behavior rewires, business models and market structure follow.

There’s still debate about how large AI’s impact will be and where it will land. But in Consumer AI, the likely outcome isn’t just “the market gets a few times bigger.” The more consequential possibility is a structural rewrite: new consumer flows, new distribution, and a reallocation of value away from today’s incumbents.

In [Two Cents #84], I looked top-down—how the value chain is being reassembled and what that implies for market structure. Here, I’ll go bottom-up: what becomes newly possible because of AI, what new consumer behaviors emerge, and what second-order effects those behaviors create.

I. “YouTube-ification of Everything”

The first pattern worth naming is what I’ll call the “YouTube-ification of everything.”

By “YouTube-ification,” I mean this: tasks that used to be inaccessible to everyday users—because they required expertise, time, tooling, or scarce resources—become broadly doable thanks to a new technology. Over time, that creates entirely new consumer behaviors, not just new features.

YouTube’s 2005 breakthrough wasn’t simply “online video exists.” It was the removal of friction: digitize, upload, and instantly watch without wrestling with formats, codecs, or downloads. Once watching and sharing became trivial, video consumption and distribution commoditized—and an entire behavioral and economic layer emerged on top of that. “YouTube-ification” is that broader phenomenon: a technology makes something newly possible at scale, and the behavior that follows becomes mainstream.

AI-driven YouTube-ification is already showing up across more domains than most people expected—and it’s spreading fast.

It started with image and video generation, which was the first “wow” moment for the mainstream. But it quickly expanded to music, 3D model generation, 3D animation, and now even real-time 3D world models you can navigate. (Examples: Google Genie3, World Labs.)

And it’s not limited to “creative content.” If you zoom out, “vibe coding” is part of the same arc: making software creation accessible to non-experts. At the extreme end, you now see platforms that turn software building into a social activity—closer to writing and sharing than traditional development (Wabi)—and even game creation via vibe coding (Verse8).

If everything—from images and video to games and software—gets YouTube-ified, what happens to consumer behavior? Even if you restrict the thought experiment to “content creation,” the scale feels not 2–3x, but 100–1,000x relative to prior creative revolutions (think the invention of the camera). And at that scale, you don’t just get more volume—you get qualitative behavioral shifts.

A few plausible first-order and second-order effects:

• Professional content moves upmarket toward stronger narrative and higher production quality

  • Similar to how photography pushed painting into new forms and branches.

• Consumer-generated content becomes less about the artifact, more about the behavior around it

  • Snapchat: disappearing media enabled deeply personal, conversation-native social behavior.

  • Early TikTok: “music video battles” mattered less as content and more as social interaction.

• Real-time, personalized generation and navigation of 3D models / virtual worlds / games

  • AI-assisted “spatialization” as a new default—raising the question: what new behaviors emerge in AI-shaped spaces?

• Commoditized software creation → software as a utility (more below)

What other second-order behaviors become possible once these first-order shifts go mainstream?

Note: The examples above mostly stay at the first-order level. That’s a limitation of imagination, not importance. If first-order shifts create near-term opportunities, second-order shifts are often the next wave—likely emerging 2–3 years later, built on top of the first.

Software as a Utility

The most extreme expression of this worldview is something I say informally: software becomes a utility—like electricity or water.

Today we take electricity for granted: plug in anywhere and it works. But if you look at the early history of electrification—even just ~100 years ago—using electricity in a neighborhood required custom wiring, local generation, and bespoke infrastructure shaped by building layouts and density. The transition to modern electricity—standardized voltage, national-scale distribution, metering, and “plug-and-play” access—required decades of technical standardization, industry consolidation, and infrastructure consolidation, catalyzed by leaders like Samuel Insull.

Software’s evolution since commercial adoption in the 1960s is structurally similar:

Hardware- and task-specific programs (ENIAC, Apollo 11-era code) → vendor-led hardware standardization (IBM) → OS/platform consolidation by hardware scale (IBM, DEC) → off-the-shelf OS standardization (Unix, Microsoft) → shrink-wrapped software → client-server → cloud/SaaS → vibe coding → and eventually, software as a utility (still ahead of us).

If you buy this arc, the next phase is “software as utility,” or the mainstreaming of “one-time-use software.” Vibe coding is an early signal. Platforms like Wabi - built around sharing vibe-coded software - look like the beginning of “software’s YouTube moment.”

A few plausible behavioral implications (again, mostly first-order):

• First, we may stop thinking of software as an “object you build” and start treating it as a disposable “tool you summon” to accomplish a job. As tools are created and shared—and as remixing becomes normal regardless of whether the output looks like an app, a SaaS endpoint, or an agent

  • The first consumer behavior shift is straightforward:

    people will try to find tools before they try to build them. This resembles early “GPTs” or today’s prompt libraries, but likely becomes far more consumer-native.

  • If that happens, the place where people browse, buy, and exchange tools may look less like an App Store or Zapier-style marketplace, and more like a consumer retail experience—think Amazon or Coupang, optimized for discovery, trust, and conversion rather than developer packaging.

  • A second behavior is also likely: rent-and-discard becomes normal—using software once, then moving on.

  • Over time, the “tool-sharing” space could evolve beyond tools into an arena for best practices, workflows, and mental models—ultimately becoming a social commons for “how to work” and “how to live.” Instagram didn’t stay a “pretty photo app”; it turned into identity, then marketing, then commerce. A tool commons could follow a similar path.

• Second, as software fades into the background, consumer attention shifts from tools to workflows.

  • Many of today’s habits—email, calendars, to-do lists—aren’t “natural”; they’re behaviors shaped by the tools that existed (not the other way around).

  • If tools become as abundant and disposable as electricity, “using a tool” stops being a conscious act. The focus moves to what you’re trying to accomplish with the tool—like cleaning a room or watching TV, not “using electricity.”

  • The analogy is the internet itself: it used to be a high-friction activity (“doing the internet”); now it’s ambient.

What additional second-order behaviors emerge once this becomes a default?

AI Scientist

This isn’t directly consumer behavior, but if you extend the “YouTube-ification” thought experiment to its logical endpoint, one destination is frontier R&D.

One of the most interesting AI-native categories emerging at the SOTA frontier is “AI scientists.”

About a year ago, there were announcements of agent-only research teams discovering new Covid-19 vaccine candidates without human intervention. Now, teams composed primarily of AI agents are becoming more common. It’s plausible that “vibe coding for science” becomes real—at least for professional researchers—sooner than we think.

As foundation models improve, the bottleneck shifts from raw modeling to domain knowledge and problem framing. That suggests the first breakout domain could be drug discovery and life sciences, where value-add is enormous and the data/constraints are deeply technical. We’re already seeing “AI scientists” emerge in life science in a way that resembles early startup formation.

Push that far enough, and you can at least entertain a bigger question: could this be the practical starting point for an AGI/ASI-driven “intelligence explosion” that many visionaries speculate about?

II. “No App, No OS” — The End of UI/UX as We Know It

If “YouTube-ification of everything” is the first shock, the second is what it implies about interfaces.

Two implications stand out:

1. The abstractions we’ve taken for granted—OS, software, apps—and the businesses built around them may be dismantled and recomposed.

2. The resulting consumer behavior shift could be far larger than we’re currently modeling.

“AI is being baked into keyboards, browsers, and operating systems. If your assistant is everywhere you type, why open a separate app?” — 2025: The State of Consumer AI | Menlo Ventures

Elon Musk recently claimed that within 5–6 years, OS and apps will become unnecessary; users will talk to an AI-enabled device (likely not a smartphone), and most content people see will be generated in real time as needed.

Musk is often controversial, but his first-principles instincts—setting a direction, then building toward it—are hard to dismiss. I’m not sure about his exact timeline, but I think the direction is broadly right: the concept of “software” and the way humans interact with it—including the OS layer—may change at a foundational level.

It’s also worth remembering how young the “mobile app” artifact really is: ~20 years since modern smartphones, perhaps 30–40 if you go back to Palm Pilot-era apps. Apps exist because the smartphone existed. If the smartphone itself becomes less central over the next few decades, it would be odd to assume the “app” remains the dominant interface primitive.

One useful framing: “This multiplayer AI moment will create entire categories of software we don’t have names for yet, because the structure of collaboration is about to change from ‘people using apps’ to ‘groups interacting with intelligence’” — Greg Isenberg on X

The UI/UX Reset: From Interface-Heavy to Intent-Driven

This transition implies a deep reset in how humans and systems connect—both at the UI touchpoint and across the broader user experience.

As discussed in [Two Cents #76] UI, UX, the interface layer is likely to change along several dimensions. This won’t fully happen in the next three years, but it also doesn’t feel like a 10+ year story.

1) The end of “request–response UX”

The most fundamental shift is the decline of the classic loop: the user requests something through a UI, the system responds, the user requests the next step, and so on. That interaction model will be replaced in multiple ways.

Consider an ambient agent that proactively intervenes. Instead of waking up and scanning email and calendar, a 24/7 assistant summarizes your day over breakfast: priorities, context, trade-offs, and suggested actions—then asks lightweight questions like “Should I proceed?” If users get comfortable with this, why would anyone care which email client, calendar app, or to-do list product they use?

2) Where is the primary UI surface?

If intent becomes the primitive, what becomes the default surface? The smartphone again? The desktop? Or ambient environments—cars, offices, homes, public transit, streets—filled with cameras, microphones, and speakers? Or new form factors like pins, AR glasses, or wrist-worn devices?

3) What is the primary interaction modality?

Is it still screen-first—typing and uploading—just with smarter AI? Or voice as the default? If voice, who is the counterpart: Siri-like assistants, ambient devices like Alexa, or a dedicated AI app? If the interaction is ambient, does a cloud agent triage continuously?

4) Who becomes the primary interaction “owner” before the new world arrives?

During the transition, what becomes the primary orchestrator in today’s device environment? Does Siri become the front door and route everything else? Do ChatGPT/Claude become a new super-app—and effectively a new app store? Do users navigate “AI URLs” or standalone AI apps like they do today? Or does the browser reassert itself?

In aggregate, this points to a behavioral shift: consumers move from interface-heavy UI/UX to intent-driven workflows.

Some near- and mid-term implications:

• In the short term, the platform-based web/app environment begins to fragment, and apps/services/mini apps reorganize around new platform layers.

• Early signals include the OpenAI App SDK and Apple’s newly announced Mini Apps Partners Program. (My related notes posted on X: App SDK, Apple Mini Apps.)

• The “new platform layer” could settle at iOS/App Store, an AI browser, ChatGPT (an “AI super app”), or a distributed mix. Where it lands depends on technical trajectories and consumer adoption.

• Combined with “software as a utility”—and the broader idea that “SaaS is dead, at least as we know it” (a topic for later)—the market may simultaneously:

  1. aggregate into a smaller number of headless suppliers, and

  2. fragment into an ecosystem of countless micro/mini apps (often single-purpose), reminiscent of mini programs inside WeChat.

• This reconfiguration will reshape the front-end interface layer and the back-end headless supplier layer—likely over a long time horizon (10+ years).

Travel is a concrete example. Today’s structure is dominated by consumer-facing aggregators like Booking.com and Hotels.com. A plausible future structure is: an AI super app or agent becomes the consumer-facing front-end, while a small set of headless booking suppliers—air (GDS like APOLLO, AMADEUS) and hotels—become the back-end infrastructure.

III. Second-Order Effects on Consumer Behavior Shifts

Once AI moves beyond “automation for efficiency” and “content generation” into agent-led judgment and execution, the basic unit of consumer behavior stops being apps, pages, or features. Users begin with intent (“what am I trying to accomplish?”), and the system handles search, comparison, execution, and follow-through as one continuous flow.

This is less a UI/UX evolution and more a shift in how consumers conceptualize digital systems.

Reframed as second-order behavioral changes:

Mass creation: consumers don’t just consume outputs—they consume participation

The most direct consequence of “YouTube-ification” is that creation is no longer reserved for experts. As the cost of producing images, video, music, 3D, code, and games collapses, consumers naturally move across the boundary between producer and consumer. The key isn’t that outputs get better; it’s that making becomes normal.

The second-order effect is a shift in attention. Consumers increasingly value the live-ness of the process, the ability to participate, and the social context of remix and feedback more than the finished artifact. As with early TikTok battles or Snapchat’s disappearing media, the “work” isn’t the point—the relationships and interactions around the work are. Consumers act less like audiences and more like co-creators and curators.

And this isn’t limited to media. Software itself is becoming a consumer creation surface, and platforms like Wabi are early examples of consumer-native creation and sharing around vibe-coded software.

The end of search and browsing: “finding” fades; “delegating” becomes default

A large fraction of the legacy internet was built on search and exploration. To solve a problem, you formed queries, opened links, compared pages, and made decisions yourself. As AI replaces that workflow (search → summary), consumers stop gathering information and instead provide constraints and decision criteria—then accept or reject the system’s output.

That’s a subtle but foundational change. Cognitive load moves from “how much did I read?” to “how much do I trust this system?” The important question becomes not what was chosen, but why it was chosen—and whether that choice is reversible. The end of search UX is not just the disappearance of a search box; it’s the transition from exploration-first thinking to intent-first thinking.

Software as a utility: consumers don’t “use tools”—they accomplish work

When software behaves like a utility (electricity/water), consumers stop thinking of themselves as “people who use tools.” Much like we no longer say “I’m using the internet,” the act of using software becomes invisible. In a world where capabilities are summoned on demand and vanish when the job is done, installing and learning apps becomes the exception.

As a result, behavior reorganizes around workflows. The focus is no longer which email client, calendar, or to-do app you use, but how you run your day and how you sequence tasks. Software stops being an identity or productivity object and becomes background infrastructure—sometimes visible, often not—that enables outcomes.

The collapse of request–response UX: users don’t operate systems—they approve flows

Traditional UI/UX is built on a repetitive request–response loop. But when agents maintain context and take responsibility for end-to-end workflows, that loop weakens. Users stop asking “what should I click?” and start approving, rejecting, or adjusting the system’s proposed plan.

In that world, UX is less about interface complexity and more about the design of intent transmission and decision processes—delegation, human-in-the-loop checkpoints, frequency of intervention, escalation rules, and personalization. “Good UX” becomes: ask the minimum number of questions at the right moments, and help the user decide—or decide on their behalf within clearly defined bounds.

The core “cost” of consumption shifts: from money to intent, trust, and permission

Finally, the nature of what consumers “pay” changes. As time and attention spent inside apps declines, users delegate more to systems. The scarce resource becomes not dollars, but trust—and the permission to act.

In that context, ad-driven feeds and click-optimized funnels weaken. Consumers express intent; agentic systems focus on delivery. The economy shifts from what I described in [Two Cents #84] as the “Attention Economy” toward an “Intent Economy.”

Premium value accrues not to content, clicks, or conversion mechanics, but to the quality of delegation: how confidently a consumer can trust the system, how decisions are made (HITL, delegation), and how failures are audited, explained, and reversed. Consumers will do less themselves and delegate more—and the design of that delegation will define the next generation of consumer experience.

IV. Market Structure Shifts and New Startup Opportunities

Market structure: from “app-centric internet” to “agent-centric economy”

As AI takes ownership not only of generation but execution, the foundational unit of the consumer internet moves from apps to intent. Instead of searching for functionality, installing software, and learning workflows, users state goals (explicitly or implicitly) and rely on systems to decompose and orchestrate actions. Value accrues less to UI polish or feature breadth and more to who can correctly interpret intent, delegate execution effectively, and design the decision boundary: who decides what, and how.

In practice, the “driver” of this process is less the human user and more a multi-level swarm of agents coordinating tasks. The service providers behind the scenes increasingly appear as headless infrastructure—agents, micro-services, and capability endpoints—rather than consumer-facing apps.

This naturally polarizes the market. On one end, standardized supply consolidates into a small number of headless suppliers/aggregators (travel is the canonical example via GDS). On the other end, micro-capabilities explode: skills, mini apps, tools, single-purpose endpoints. Where “big apps with many features” once dominated, we shift toward workflows that dynamically call whatever capability is needed. The user experience is no longer defined by cross-app switching costs, but by end-to-end “intent → execution” quality.

Player reconfiguration: where does the new platform live?

In this world, “platform” is no longer primarily about app distribution (like an app store). It’s about being the first surface where intent is expressed. Layers that sit at the everyday gateways—keyboard, browser, OS, notifications, payments, identity, contacts, calendar—have the most natural access to user context and permissions. That positions them to evolve from “hosting apps” to routing intent: deciding which model, service, or API to call, when. As the act of opening apps declines, that layer becomes the real distribution channel.

At the same time, AI super apps (ChatGPT-like) or AI browsers can become the new portal—except the portal no longer routes to links; it completes tasks. The competitive moat shifts from response quality to orchestration reliability: persistent understanding of user context, stable integration across external tools/suppliers, and the ability to reverse outcomes when something goes wrong (control, auditability, guarantees). This becomes less a fight over “who owns the app ecosystem” and more a fight over “who owns permissions + context,” and therefore who controls execution flow.

Startup opportunities: not “another app,” but a better execution layer for intent

The largest opportunity here is not building yet another app, but building the layer that executes intent better. Most real-world goals are multi-step and constraint-heavy (e.g., “optimize a family trip to Jeju this weekend”). Products that can decompose intent, connect data and tools, validate outputs, and handle exceptions will win. The product isn’t UI. The product is progress.

A second major opportunity is the trust and permission layer—delegation by design. When AI can pay, book, message, and submit documents, consumers don’t want more eloquent answers. They want accountability: who is responsible when it’s wrong, what evidence drove the decision, how far permissions extend, and how quickly delegation can be revoked. That creates demand for provenance, audit logs, policy-based permissions, rollback mechanisms, and even guarantees/insurance-like constructs. This is not “feature work”; it’s foundational infrastructure that makes delegation psychologically safe.

Finally, as capabilities atomize, the market converges on discovery and composition. Consumers won’t browse endless micro tools. They will expect systems to curate and bundle capabilities around goals. The resulting opportunity looks less like “App Store 2.0” and more like workflow-native commerce and distribution. Examples that feel natural in this direction:

• Goal-based packaging: input “interview prep,” “moving,” or “kid’s birthday party,” and get tools, content, checklists, and execution bundled end-to-end

• Agent-native commerce infrastructure: standardized product/service metadata, real-time price/inventory/policy APIs, and automated post-purchase customer support

• Personal data vaults: secure storage of personal context (calendar, relationships, preferences, finances) with policy-based partial delegation to agents as a personalization layer

In short, the weakening of the app-centric ecosystem doesn’t mean “products disappear.” It means the center of gravity shifts toward intent execution flows and trust-by-design systems. In this transition, the most promising wedge points are: (1) intent routing and orchestration, (2) trust/permission layers, and (3) distribution, bundling, and discoverability for micro-capabilities—areas where new tier-one consumer and platform companies can emerge.

Key takeaways

The core implication is simple: many assumptions that shaped internet playbooks for the past 10–30 years are no longer stable. Business models built on those assumptions may become obsolete faster than people expect.

There’s a recurring failure mode in tech transitions: even after a new paradigm arrives, old business models keep growing due to inertia, which leads people to underweight the probability of structural change.

BlackBerry is a clean example. In retrospect, its collapse looks inevitable after the iPhone. But historically, BlackBerry actually grew for years after the iPhone (2008) and App Store (2009). It peaked around 2011—roughly 4x larger than it was in 2008.

The lesson: a business can be growing—strongly—while still moving in the wrong direction relative to the next market structure. And a decade later, it becomes the new BlackBerry.

If you’re building now, don’t be the “BlackBerry” of the Intent Economy.

Human needs don’t fundamentally change. What changes is how those needs are served—the delivery mechanism, the form factor, the defaults, and the surrounding environment. When the delivery mechanism shifts, consumer behavior shifts with it—often gradually, then suddenly. And when consumer behavior rewires, business models and market structure follow.

There’s still debate about how large AI’s impact will be and where it will land. But in Consumer AI, the likely outcome isn’t just “the market gets a few times bigger.” The more consequential possibility is a structural rewrite: new consumer flows, new distribution, and a reallocation of value away from today’s incumbents.

This post is an attempt to map the shape of that business-model transition.

Where $1T of value capture is headed

As AI assistants—ChatGPT, Gemini, Claude, Perplexity (I’ll shorthand this whole class as “ChatGPT”)—cross the billion-user threshold, we’re starting to see early signals of a behavioral shift in how consumers use the internet.

When consumer behavior changes, business models change. And when business models change at internet scale, you’re not talking about a niche opportunity—you’re talking about new $1T outcomes (measured in terms of the market cap in the same league as Google and Amazon).

From “search” to “summary”

By 2025, estimates suggest that ChatGPT-driven query volume has reached ~10% of Google’s search query volume, and the “starting point” for information discovery is rapidly moving from Google to ChatGPT. Google took more than a decade to reach its current scale; ChatGPT reached comparable behavioral gravity in roughly two years.

More importantly, the function is shifting. AI assistants are moving from being a starting point (“search”) to becoming the endpoint (“summary”)—where users expect the answer directly, without needing to click out and browse.

That behavioral change matters because it reroutes the entire flow of traffic—what the internet has historically monetized as “attention.” The classic path was:

Google Search → SEO → ad networks → destination sites (commerce, services, news, info)

If the user journey increasingly ends inside an AI assistant, that traffic doesn’t follow the old route. And if traffic doesn’t follow the old route, revenue pools tied to that route become mobile.

If roughly ~50% of Google’s search traffic is “informational intent,” then a shift from search → summary implies that a meaningful portion of Google’s revenue is exposed. Against Google’s 2024 ad revenue base of roughly $265B (Search + YouTube + ad network), a back-of-the-envelope framing is that ~$130B of value may be vulnerable to redistribution into new channels.

The “purchase journey”: from “search & browse” to “intent & recommendation”

The same re-routing is happening for purchase-intent behavior—shopping, local, and professional services.

The first battleground is already visible: platforms are competing to own more of the purchase journey, and that competition is expanding beyond the boundaries of any single platform.

Historically, commerce platforms competed within their own walls: capture search, keep the user browsing, close at checkout. What’s changing now is that platforms are attempting to extend their control outward—effectively running search → browse → checkout on behalf of the user across other surfaces.

That dynamic is already creating friction between incumbents.

• Amazon has pushed forward with “Buy for me”, where an agent handles discovery, comparison, and checkout from other platforms inside Amazon’s ecosystem.

• In response, platforms are beginning to define new policy boundaries around agent access to product data and checkout—with examples like Amazon restricting agent access entirely, and Shopify allowing access to product pages while limiting checkout pathways.

The next step is the AI assistant itself absorbing the full purchase flow. OpenAI’s “Instant Checkout” directionally signals a future where a user can complete search, browse, and checkout without leaving ChatGPT. That’s the start of a new competitive frontier: ChatGPT vs. commerce platforms for ownership of the consumer purchase journey.

If you broaden the lens, this goes beyond buying a single item. AI assistants are increasingly providing “full-stack intent fulfillment” for multi-step tasks that require planning, research, and curation—e.g., “Plan a trip to X with Y budget”—including vendor discovery and evaluation.

This is exactly the experience commerce and booking services have wanted for decades, but couldn’t deliver due to technical constraints.

And these behaviors are no longer hypothetical.

In my own case, when my family planned an overseas trip in summer 2025, an in-chat agent (specifically Manus in this case) did the research: proposed an itinerary, surfaced hotels, and identified local travel agencies. We then completed bookings through the relevant providers—but critically, we did not need Google search at any point for planning or discovery.

Full agent execution—booking the hotel end-to-end, negotiating with a local operator via voice and messenger, etc.—is not universally productized yet. But the components are already real:

• hotel booking automation is already feasible in agentic workflows

• outbound/inbound calls via AI have been demonstrated for years (Google’s Duplex demo in 2018 was the early proof point), and AI-assisted calling is increasingly normal in specific verticals (reservations, outbound sales), at least in the US market

What’s still missing is not capability—it’s packaging, incentives, and rails (including the affiliate layer that routes demand to sellers in a way sellers will accept).

At a technical level, agents can execute the entire transaction pipeline:

• for ecommerce: search → SEO → ecommerce front-end → product page → checkout

• for service providers: affiliate network → service provider → checkout

Which implies a deeper substitution: search → browse → checkout gets replaced by task → recommendations → selection.

When that happens, traffic—and therefore monetization—no longer needs to flow through Google, ad networks, affiliate networks, or merchant/service-provider front-ends. The value capture shifts upstream to wherever intent is interpreted, executed, and monetized.

Considering roughly 50% of Google’s traffic carries purchase intent, ~$130B of Google revenue is in the blast radius. And Amazon’s ~$55B annual search advertising revenue is directly exposed as well.

“Attention Economy” → “Intent Economy”

The Intent Economy

The cleanest framing is that the internet is reorganizing from an “attention-first” system to an “intent-first” system.

The Attention Economy was built around approximating intent indirectly—through traffic patterns and engagement signals. The primitives looked like this:

1. user arrives (eyeballs)

2. the system infers intent from signals (queries, dwell time, likes/comments)

3. the system optimizes response to drive reaction (click)

4. the user browses tools/interfaces toward decision (conversion)

5. revenue accrues to the destination and intermediaries (sales/subscription/ads)

User flow: eyeballs → attention → click → conversion → revenue / gross margin

If you reverse it from the money side (simplified for commerce):

• merchant generates ~30% gross margin

• ~half of that (10–15% of the purchase price) goes to sales/marketing/brand spend

• that spend gets allocated across branding, performance marketing, affiliate, ad networks

• and ultimately lands with platforms like Google, Meta, TikTok/Instagram, Amazon, and media pages

Value accrual: GM (30%) → marketing budget (10–15%) → affiliate/ad layers (~5%) → platforms

This entire architecture of Attention Economy in the current system exists because true intent was hard to grasp, so the industry got very good at modeling it from proxies: CTR, conversion funnels, A/B testing, attribution.

The Intent Economy flips the order.

If the system can capture intent directly (through prompts, context, or memory inside the AI Assistants) and execute the exploration process on the user’s behalf, then the core loop becomes:

• understand intent

• run search/browse/analysis through agents

• deliver the best matched recommendation

• execute (with or without human approval)

A simplified future state would be like:

• User flow: intent/context → “concierge agent” → multi-agent swarm → delivery (product/service/content)

• Value accrual: GM (30%) → brand/marketing/sales agent → recommendation agent → intermediary agent swarm → “concierge agent” → delivered to user

This isn’t a solid structure yet. It’s highly fluid, and multiple versions will be tested in the market. Even if stabilization takes a decade+, we should expect intense competition over the next 2–3 years, and the beginnings of a durable role/value map within ~5 years—similar to how Web 1.0 markets formed and consolidated.

Note on the “concierge agent”

The concierge agent is the layer that owns the user relationship and orchestrates intent fulfillment. Long-term, it likely becomes the most important choke point for value capture—especially once it is powered by a real personalization layer (agent form factor vs. data layer form factor remains unclear). The current “AI Super App” race is, at its core, a race to own this layer. I’ll cover that separately.

How the intent flow changes across the funnel

The simplest description: from “user-driven exploration” to “agent-driven delivery and execution.”

In the Attention Economy, the consumer journey is typically framed as: Discovery → Engagement → Conversion → Loyalty/Retention

Mapped to the Intent Economy, each stage changes meaning as follows:

I. Discovery

In the Attention Economy, discovery is about capturing attention through search, SEO, ads, content, and branding.

In the Intent Economy, discovery becomes: understanding intent and initiating execution.

• fades: attention capture (search/SEO), affiliate marketing, ads, social discovery

• emerges: intent interpretation, personalization, transaction initiation, agent competition/swarming, results comparison, personalized selection

Core components:

• Concierge agent: interprets intent, triggers transactions, routes results, requests approval (human-in-the-loop) or executes autonomously

• Personalization layer: filters/selects outputs based on user context (embedded in the agent or accessed as a data layer)

• Infra: delegation permissions, rights management, privacy

II. Engagement

In the Attention Economy, engagement is about keeping users in the funnel: UX optimization, reviews/social proof, recommendations, metrics.

In the Intent Economy, the focus shifts from engagement to delivery: agent negotiation, execution, and verifiable outputs.

• fades: UI personalization as funnel optimization, A/B testing for clicks, review-driven persuasion as primary tool

• emerges: agent-to-agent negotiation, machine-readable execution artifacts, micro-transactions/financial rails for agent commerce

III. Conversion

In the Attention Economy, conversion is optimized through checkout design, cart recovery, upsell/cross-sell.

In the Intent Economy, conversion becomes: execution + closing—either via human approval or direct agent action.

• fades: checkout-flow optimization, cart abandonment tactics, upsell/cross-sell mechanics

• emerges: human-in-the-loop confirmation, direct agent execution, payment rails designed for agent finalization

IV. Loyalty & retention

In the Attention Economy, loyalty is maintained through branding, loyalty programs, retargeting, and repeated re-entry into discovery/engagement loops.

In the Intent Economy (outside of certain categories like luxury), loyalty mechanics likely weaken. Branding and loyalty won’t disappear, but their marginal impact should compress because the system optimizes for intent-match delivery, not for attention capture.

• fades: branding as primary retention lever, loyalty programs, retargeting

• emerges: continuous competition over intent outcomes; new strategies to influence the personalization layer

What changes should we expect?

Much of this remains speculative. But directionally:

• A rewrite of B2C infrastructure: SEO, ad networks, affiliate networks, creator-economy tooling

  • The roles may not vanish, but the value propositions and handoffs across the chain will change.

  • Current incumbents likely might have a ~5-year window to adapt; growth-stage players must either start natively in the new structure or continuously re-architect in real time.

• Commerce likely would change first and most violently. It’s where intent is easiest to monetize and where the funnel is most directly measurable.

• **Marketplaces will be rebuilt. “**Discovery and connection” mechanics are changing at the protocol level. The next-generation marketplace likely won’t look like the current one—even if much current commentary focuses on how incumbents defend today’s structure rather than imagining a new one.

Key takeaways

The core implication is simple: many assumptions that shaped internet playbooks for the past 10–30 years are no longer stable. Business models built on those assumptions may become obsolete faster than people expect.

There’s a recurring failure mode in tech transitions: even after a new paradigm arrives, old business models keep growing due to inertia, which leads people to underweight the probability of structural change.

BlackBerry is a clean example. In retrospect, its collapse looks inevitable after the iPhone. But historically, BlackBerry actually grew for years after the iPhone (2008) and App Store (2009). It peaked around 2011—roughly 4x larger than it was in 2008.

The lesson: a business can be growing—strongly—while still moving in the wrong direction relative to the next market structure. And a decade later, it becomes the new BlackBerry.

If you’re building now, don’t be the “BlackBerry” of the Intent Economy.

I’ve been trying to map Consumer + AI by walking through individual verticals and listing “potential opportunities” (e.g., education in [Two Cents #81]). I’ve come to believe that’s not the best way to do this.

Founders—not investors—are the ones who discover the real wedge. They live inside the problem long enough to see what outsiders can’t: the hidden constraints, the behavioral friction, the second-order effects. Even when a product looks “obvious” in hindsight (Uber, TikTok, Snapchat), getting to PMF—and then turning early-adopter behavior into mainstream habit—is an exhausting, non-linear craft. It’s hard to appreciate from the outside.

And when a casual observer lays out “the opportunities” too early, it can actually narrow founder imagination. It can keep the conversation shallow, or worse, anchor a team to a premature framing. That’s an easy form of investor arrogance to fall into. I’ve fallen into it before.

So I’m changing the lens: instead of enumerating startups-by-vertical, I want to focus on insights and value props—the kinds of shifts that create new behavior. Let founders own the path from thesis to product.

Going forward, I’ll share a set of “starting points” that might help founders find concrete problems inside Consumer + AI. And when helpful, I’d like to be a thinking partner as they go from macro shift → specific wedge.

Tech shifts trigger consumer behavior shifts

Human needs don’t fundamentally change. But when technology changes the form factor, distribution, and cost structure of how needs are served, behavior changes—slowly, then suddenly. And when behavior changes, business structures eventually follow.

Early automobile history is the canonical example. If you asked consumers what they wanted before cars existed, they might have described “a faster horse.” They couldn’t predict highways, suburbs, logistics networks, road trips, or the re-architecture of cities. The invention wasn’t the story—the second- and third-order consequences were.

AI is that kind of enabling technology.

Where generational companies come from

From a consumer investor’s perspective, generational companies are born where tech shifts intersect with consumer behavior shifts—where a new capability unlocks a new default habit.

Web 1.0 produced Google and Amazon. Mobile produced Uber, WeChat, Coupang. These weren’t just better products; they shaped new behavior at scale.

An early-stage VC’s job is to find these generational opportunities before they’re obvious. Over the last 30 years, examples include Google, Roblox, Facebook, Uber, Tencent (WeChat), ByteDance (TikTok). In Korea: Naver, NCSoft, Nexon, Kakao. In the AI era, the goal is the same: identify the winners early, earn allocation, and help them compound.

Tech shifts vs. behavior shifts: causality usually runs tech → behavior

Tech and behavior influence each other, but in most cases the direction is simple: new tech makes previously impossible things possible, and behavior follows.

Mobile didn’t just improve the internet—it introduced new primitives: location tracking, always-connected usage, a camera in everyone’s pocket. Those primitives enabled new behaviors, and new companies followed (Uber, Instagram, TikTok).

It’s also useful to think in orders of magnitude, not binaries. A 2x improvement often reshapes incumbents. A 100x–1000x improvement creates behaviors that were effectively impossible before.

Generative media is a clean example. When image/video/music creation becomes dramatically cheaper and easier, creation expands from “experts” to “everyone,” and new forms of creativity emerge. At the same time, professionals don’t disappear—they move to a different frontier of craft, just as painters evolved after photography. Photography didn’t kill art; it created a new medium and forced differentiation.

A closer example: when video distribution approached near-zero marginal cost, entertainment shifted from film-and-broadcast constraints to personalized, on-demand streaming. The medium changed; the consumer habit changed.

On the time window of change

These transitions play out over long arcs. Mobile took ~10+ years to fully reshape behavior. The web took ~30 years.

The timeframe depends on how “new” the behavior is. The web’s core shift—democratized information distribution—created fundamentally new social patterns. Mobile, in many ways, was a massive form-factor upgrade on existing behaviors (more continuous, more accessible).

By that framing, AI looks closer to the web than to mobile: it introduces genuinely new capabilities, not just a better interface. A 30–50 year arc seems more realistic than a 5–10 year arc—while acknowledging that the most investable company formation happens in the early chapters of that arc.

So if we want to reason about consumer behavior change, the practical method is:

1. identify what tech now makes possible that previously wasn’t, and then

2. model the first-order and second-order behavioral consequences.

That’s the framework I’ll use in this “consumer behavior shifts” series.

Two broad arenas of consumer behavior change

You can over-intellectualize this with Maslow, but a simpler split is useful when evaluating new consumer services—especially if you’re building for a 3–10 year horizon.

1) “Survival” needs: products & services

Products: commerce, logistics, physical goods distribution.

Services: everything else that supports life and productivity—education, work, health, finance, transportation, travel.

If I list early keywords for how AI changes this arena:

white-glove AI concierge, hyper-personalization, lifetime companion, near-zero marginal cost.

If you can translate a combination of these into a crisp value proposition that feels inevitable, you’re usually staring at a real opportunity.

2) “Play” needs: the Spectrum of Play

When tech shifts happen, “play” tends to react first—and monetize first. Entertainment pushes new primitives to their limits.

Historically: adult content seeded early web monetization; then games and broader entertainment followed. If you widen the lens, sports betting and even prediction markets can be seen as extensions of this “play” impulse—though I’m still forming my view on what that implies.

My prior is simple: the play economy remains a huge share of consumer spend and attention, and it will be among the first places AI-native behavior becomes mainstream.

Is this the opening of Consumer + AI?

For the last three years, especially in the U.S., the AI industry has been enterprise-first. Consumer experimentation existed (image/video generation, lightweight GPT wrappers, AI companions), but the deeper question—what new consumer behavior AI enables—has only recently started getting serious attention. (In Silicon Valley, I only began hearing “Consumer AI” as a mainstream phrase around 2025 Q3–Q4.)

Why enterprise moved first

Enterprise had the perfect conditions:

• immediate willingness to pay for productivity

• proven ROI

• existing IT budgets that could be reallocated

• relatively low price sensitivity versus consumers

When a coding agent can plausibly pay for itself, a $200/month price point isn’t a deal-breaker. That makes token-cost economics less visible to the buyer.

I expect enterprise adoption to stay strong for a long time—expanding steadily, more linear than exponential, because organizations and spending patterns evolve slower than consumer habits.

Why consumer will be slower—but larger

Consumer scale requires behavior change. That means:

• the value proposition must feel personally obvious

• new habits must form

• many products need a critical mass before the flywheel kicks in

In prior cycles, abundant VC liquidity helped companies bridge the gap from “new tech” to “new habit.” This time, the macro environment is different, and token costs can become a real constraint in price-sensitive consumer markets.

However, AI also has one major advantage versus prior cycles: the distribution substrate already exists. We don’t need broadband penetration (web) or smartphone adoption (mobile) as a precondition. The initial critical mass is already here—though the dominant form factor and UX may look very different 10 years from now.

Net: consumer may take longer to re-architect, but the upside is larger if you become the default surface—an “AI-era Google”-scale outcome.

And as a separate point: Korea has historically been unusually strong at inventing new consumer behaviors during platform transitions (Cyworld-era social, Lineage-era gaming, etc.). That creative edge can matter disproportionately in consumer.

Tech keywords that could trigger new consumer behavior

At this stage, a few “trigger” themes stand out:

• Near-zero marginal cost content generation → the “YouTube-ification of everything”

• Lifelong personalization with infinite options (personalization as a compounding asset)

• Entertainment gets re-imagined end-to-end

• Agents as a new Consumer OS

• Rebuilding online social infrastructure: ads, discovery, connection—what we used to call marketplaces

I’ll unpack these as first-order and second-order effects, and connect them to startup opportunity spaces.

These triggers will also interact with broader consumer environment shifts—demographics, labor disruption, geopolitics, and the macro shift away from abundant cheap capital. The causality is rarely one-way; it’s a dynamic system.

A second lens: consumer environment changes

Some investors anchor more heavily on macro and social data—demographics, wealth transfer, job market turbulence, de-globalization, and capital cycles—then infer what consumers will do next.

That’s a valid approach, and there are great data-driven references (e.g., government social trend reports, or investor research like Digital Native’s chart-based analysis).

In Two Cents, I’ll take a slightly different stance: I’ll focus less on describing what’s already happening, and more on what AI changes in the environment itself—and what behaviors that newly enables.

What comes next

From here, I’ll go keyword by keyword—what it means, what consumer behavior it could unlock (first-order and second-order), and what kinds of startups could emerge.

This won’t be a 10–20 year grand forecast. It’s a snapshot meant to be useful for founders building now, centered on what may plausibly shift over the next 1–3 years—the window where new behaviors start to form and early category winners begin to appear.

It’s becoming clear that the market’s “readiness” for Consumer AI has crossed a tipping point.

What we need now is to get far more concrete about how AI-driven market change will unfold—the direction, the mechanisms, and the implications for industry structure, competitive dynamics, and the economics between participants.

For founders, the job is to identify those opportunities a little earlier and move first. For investors, the job is to recognize those early moves quickly and support them aggressively.

This series—my “Flights of Thought”—is an attempt to share how I’m thinking through what will happen, what it will unlock, and what kinds of ideas are likely to matter.

This time I want to focus on what I suspect will become the most important “hidden platform” in consumer AI: the personalization layer.

I originally wanted to lay this out before diving deeper into vertical opportunities, but personalization is harder to write about cleanly—because the market is still early, the architecture is unsettled, and most “obvious” answers don’t work yet. My thinking here is not final, but it’s mature enough to be worth publishing as a baseline.

What “personalization” actually means

GenAI made two things broadly feasible for the first time:

1. Reasoning and generation over massive knowledge—at a scale no individual can internalize (transformers).

2. Infinitely personalized generation—content, outputs, experiences—at near-zero marginal cost (diffusion for images/video; transformers for code, UX, and increasingly “vibe coding”).

We’ve already seen this arc move from images → video → and now into games, 3D/world models, and software creation. When this capability extends beyond “content” and into how consumers interact with systems—the interface, the workflow, the ongoing relationship—then personalization becomes a true platform layer.

A concrete example is the “hyper-personalized, infinite-choice, lifelong learning” model I discussed in [Two Cents #81] for education.

Even today, we can see early, crude personalization in mainstream products:

• The more you use ChatGPT, the more it carries forward prior conversational context—what you’re working on, what trip you’re on, what you tend to ask for—and it starts to behave differently without being explicitly instructed every time.

• As that accumulates, switching costs rise. This is the current “LLM chatbot personalization” lock-in dynamic.

If you inspect ChatGPT’s memory, it’s still primitive: it extracts a few pieces of context it believes are relevant and reuses them as prompt context. Yet even this lightweight approach already creates meaningful stickiness.

My read is that OpenAI’s consumer strategy is essentially: build personalization infrastructure, wrap it in an account, and turn it into lock-in. Today that starts with memory/context. Long-term it likely expands into a more explicit personalization data layer—which becomes core to making ChatGPT feel like a consumer “AI super app.”

If personalization expands to the whole AI “system,” what happens?

By “system,” I mean the broader AI environment: services, apps, agents, infrastructure, and eventually ambient computing contexts.

Once personalization becomes system-level, a few characteristics define it:

• It understands your context (scope and depth vary).

• It infers your intent—either reactively (from what you say) or proactively (even when you don’t ask).

• It delivers a personalized UX that adapts over time.

• It produces outcomes that feel tailored—because they are.

Importantly, this isn’t the same personalization we’ve had for the last 30 years. The depth and compounding effect are fundamentally different.

The evolution: favorites → recommendations → delegation → proactive/prescriptive

Most “personalization” historically looked like favorites—explicit saving (playlists, bookmarks).

Then it became recommendations—systems predicting what you might like based on behavior and similarity graphs.

AI-driven personalization now splits into two major modes:

1) Delegation / autonomy (reactive personalization)

You delegate tasks to a system that “knows your context.”

• “Find a dress for this weekend’s event that fits my style.”

• “Plan and book a 2-night family trip, including my youngest.”

The key is that you’re no longer selecting from menus; you’re delegating outcomes.

2) Proactive / prescriptive personalization

An agent that knows your context begins to surface tasks before you ask.

• “Next weekend your youngest has a school event. You’ll need X items—should I order them now?”

• “Your calendar suggests travel next week; do you want me to adjust your routine purchases and deliveries?”

This is where personalization starts to look like an ambient “butler agent,” not a recommendation widget.

What form does personalization take?

I expect personalization to manifest through multiple building blocks:

Personalized agents

• Per-system personalization: each service remembers your interaction history.

• Per-person personalization: one “lifetime partner” agent carries your context across services, delegates work outward, and nudges you proactively.

“Conformative software”

Software that adapts its behavior and UX over time based on how you interact with it. Today’s ChatGPT personalization is an early, minimal version of this.

On-device vs. cloud-based

• Some of your personalization may live on-device (privacy, low latency, continuous capture).

• Some will live in the cloud (ambient agents, cross-service execution).

Collective memory

Personalization won’t stop at the individual. We’ll build shared context layers for:

• families

• teams

• organizations

  This becomes “collective memory” that enables group-level personalization.

Why personalization becomes the strongest lock-in mechanism

As personalization accumulates, lock-in grows in two directions:

1. Users will prefer systems that remember them.

   You’ll never want to be a stranger to a new agent or AI app.

2. Personalization compounds.

   Every decision you make becomes training data for the system’s future usefulness. The benefits increase with time, which strengthens retention loops.

In this world, personalization effects can become as powerful as network effects—sometimes more powerful, because they attach directly to the individual’s daily workflow and identity.

The data required for personalization

The data spectrum is broad, but not necessarily exotic. A rough stack:

• Level 1: email, calendar, documents/files

• Level 2: social graphs + messenger history

• Level 3: commerce + delivery + banking + card transactions

• Level 4: real-time location trails (e.g., maps history)

• Level 5: health data (e.g., iCloud/Apple Health)

• Level 6: real-time conversations/calls + screen/tap/click streams

• Level 7: TBD

The bottleneck is not imagining the data—it’s acquiring and using it:

• How does a service legally and practically access these layers (especially Level 4+ in real time, which is extremely difficult for third parties)?

• How do you merge and transform the data into something that creates felt personalization value?

• How do you do this while preserving privacy, trust, and user control?

Where the industry is, technically

A useful simplification of agent systems: state (memory/context), model, actions (tools/servers).

Most products today are still largely:

• stateless agents, or

• one-shot workflows.

The frontier is building stateful agents—systems that can manage memory over long horizons, maintain consistency, and improve over time.

This is why “sleep-time compute” matters: agents doing background work—organizing memories, analyzing patterns, preparing proactive suggestions—without explicit user prompts.

A key idea here is: the agent can outlive the model. Models will be upgraded frequently; the user’s long-lived agent identity, memory, and personalization layer cannot reset every time.

To make true stateful agents work, we likely need something closer to a context OS—the “context engineering” direction Andrej Karpathy has pointed at.

Early attempts and what they teach us

The next “personalization data layer” architecture is not settled. Everyone is experimenting with different workflows, UX patterns, and data-layer models.

A few examples from recent history:

• Rewind tried capturing screen-level activity (“digital time machine”), ran into privacy + unclear everyday value, and pivoted (now Limitless) toward conversation/meeting capture.

• mem0 is pursuing a memory layer that records digital activity and builds a personal knowledge base.

• Letta is building tooling to create agents with “memory.”

• Poke.com is experimenting with building a personal context layer and delivering a consumer-facing personalized assistant.

A pattern emerges: building a horizontal personalization data layer is a classic early chicken-and-egg problem. Before you have enough data, it’s hard to deliver a product users truly feel. Without that felt value, you can’t scale data collection. Without scale, you can’t become a platform.

That likely explains why “data layer first” attempts often stall.

The more promising strategy may be the inverse: start with a sharp, high-value product where users willingly contribute data, reach critical mass, then gradually broaden the personalization layer horizontally. Poke.com hints at this path.

This resembles how massive networks like Facebook or Kakao built dominance: not by starting as “infrastructure,” but by winning with a single wedge product and then compounding data advantages over time.

“Next Google”

If someone does build a true consumer-grade personalization data layer—horizontal enough to sit beneath many services—its power could exceed what Google achieved in Web 1.0.

In an agent-native world, the ratio of what consumers do themselves vs. what they delegate could flip dramatically—from something like 90:10 today toward 20:80 (or even 10:90).

The personalization data layer becomes the enabling infrastructure for:

• the best “personal concierge” agent, and

• the deepest, compounding lock-in loop.

The end-state looks like the cultural archetype: Her’s Samantha, or the classic “butler” who understands your intent better than you can articulate it—and executes reliably.

Whoever owns that layer could become the primary consumer entry point, with lock-in dynamics that compound more strongly than search ever did.

That’s why I believe the personalization data layer / agent category could produce a dominant consumer platform on the scale of Google—or potentially larger. The exact form will evolve dynamically over the next decade, but the strategic direction feels clear.

Call for Startups

The purpose of sharing this thinking is straightforward. As an early-stage investor focused on Consumer + AI, I hope this series helps existing startups better leverage AI-driven shifts—and helps new founders reduce trial-and-error as they search for meaningful opportunities.

In that sense, this is Two Cents’ version of a Call for Startups.

If you are an early-stage founder or startup in Consumer + AI and believe you are onto something, my inbox is always open. Feel free to reach out via DM or email:

hur at hanriverpartners dot com

It’s becoming clear that the market’s “readiness” for Consumer AI has crossed a tipping point.

What we need now is to get far more concrete about how AI-driven market change will unfold—the direction, the mechanisms, and the implications for industry structure, competitive dynamics, and the economics between participants.

For founders, the job is to identify those opportunities a little earlier and move first. For investors, the job is to recognize those early moves quickly and support them aggressively.

This series—my “Flights of Thought”—is an attempt to share how I’m thinking through what will happen, what it will unlock, and what kinds of ideas are likely to matter.

Now it’s time to move from macro themes to concrete opportunity spaces by vertical.

If there’s one consumer category where AI’s value proposition is immediately legible, it’s education and personal development.

Education is structurally constrained by scarce human labor (great teachers, great coaches) and fundamentally heterogeneous demand (every learner is different). AI attacks both at once.

I. Market Forces

I believe education is among the sectors AI can reshape most deeply.

Multiple long-running forces—personalization, lifelong learning, and global distribution—are now colliding. AI doesn’t just improve existing tools; it changes the underlying economics and the default learning experience. That creates both opportunity and friction.

Here’s the direction of travel.

Personalization

K–12 and most private education were built for “one-size-fits-all”: the same curriculum, the same pace, delivered to a cohort. That structure was never optimal—it was simply the only scalable option.

But learners differ across pace, interest, cognitive style, and motivation. AI can observe learning behavior in real time, adjust difficulty, re-teach weak concepts, accelerate strengths, and change the explanation style dynamically. In effect, it enables mass-market, 1:1 tutoring with economies of scale.

What used to be a premium experience accessible to a small minority via expensive private tutoring can now become broadly available.

Consumer behavior is already shifting

Parents and students are already conditioned to digital learning tools—Duolingo, Khan Academy, Quizlet—and the pandemic normalized remote learning workflows at scale.

AI raises expectations again. The baseline is no longer “content that’s accessible anywhere.” The new baseline is content that adapts to me—my level, my interests, and my learning context. The same math concept might be taught through real-world examples for one student and visual animations for another.

This is the shift from “education going digital” to education becoming individualized.

Workforce training is becoming lifelong

The labor market is changing fast enough that a single linear career is increasingly fragile. Workers need to reskill, move laterally across industries, and in many cases retrain mid-career or later.

We’re already seeing this in the expansion of job training content on platforms like Coursera. But AI takes it further: beyond courses into personalized career coaching, adaptive learning pathways, and “career transition consulting” at scale.

Workforce education becomes less like a one-time phase and more like an always-on layer across the lifespan.

Global access is accelerating

In emerging markets where high-quality teachers and resources are scarce, AI tutoring can be delivered at low cost, anywhere. This is not only about serving underserved regions—it has the potential to reduce global education inequality.

If the internet democratized access to information, AI can democratize access to personalized instruction—especially via affordable smartphones and AI-native learning apps.

How AI reshapes education, by category

Education has always been a proving ground for new technologies. The internet expanded access, mobile put learning in your pocket, and the cloud enabled distribution and analytics.

Yet even with all that progress, education remained largely static and broadcast-like: more content, but not meaningfully adaptive to the learner’s context.

AI can change the core assumption.

Learning shifts from content delivery to a co-pilot model: the system understands the learner, gives real-time feedback, remembers long-term progress, and evolves with the student. Education becomes a continuous, personalized journey across life—not a sequence of static courses.

Below is how this plays out across major segments.

1) K–12: Public education support

Tools like Google Classroom, Khan Academy, and Quizlet helped teachers distribute assignments, automate basic grading, and give students access to content. But structurally they’re still broadcast systems—same materials pushed to many students.

AI changes the starting point: it can adapt difficulty per student, reinforce weak concepts, and explain using the learner’s interests. The same math concept might be taught via music rhythm for one student and via animation for another.

It also frees teacher time by automating grading, essay feedback, and lesson planning. And it can make individualized education plans (IEPs) and special-needs support more scalable—ADHD, dyslexia, autism spectrum, and more.

In public education, AI is not about replacing teachers. It’s about buying back teacher time and individualizing student experience at scale—reducing burden, narrowing gaps, and helping students grow at their own pace.

2) K–12: Private tutoring and alternative education

Online tutoring expanded rapidly over the last decade—VIPKid connected native teachers to Asian students; Outschool built interest-based small group classes; test-prep apps offered practice banks and dashboards.

But the bottleneck remained human time and cost, and the content scope stayed relatively fixed.

AI can remove the bottleneck through always-available, 1:1 tutoring that adapts to the child’s interests (horses, soccer, comics), supports passion-driven acceleration, and shifts learning tactics when engagement drops.

It can also enable new forms of micro-schooling and homeschooling by auto-generating curricula, assignments, and assessment—making it easier for parents and educators to run small alternative programs.

Net: AI expands access to high-touch education that used to be elite and expensive, increasing diversity of educational choices and unlocking new ecosystems.

3) Language learning

Apps like Duolingo and Rosetta Stone mainstreamed language learning through gamified vocabulary and grammar. But they still skew toward repetition and limited real conversational immersion.

AI can turn language learning into a real-time, interactive, context-rich experience.

It can generate endless scenarios—restaurant reservations while traveling, workplace presentations, negotiation calls—while adapting to a learner’s professional domain (medical, finance, legal). A voice-based AI tutor becomes a constant conversational partner: always available, always responsive.

The shift is from “memorizing words and grammar” to speaking and understanding in context—language as a life skill, not a curriculum artifact.

4) Workforce reskilling and career transitions

MOOCs democratized knowledge; bootcamps increased employability through project-based learning. But mentorship and coaching remain expensive, and career transitions don’t scale well without human support.

AI can transform reskilling into personalized career coaching:

• diagnose skill gaps against target roles

• generate customized learning paths

• simulate role-specific practice (coding interviews, clinical scenarios, sales pitches)

• move assessment toward performance-based outcomes

• connect learning to labor market demand in real time

Instead of “content delivery,” AI becomes a practical partner for changing careers. Over time, education begins to blur into labor-market infrastructure.

5) Lifelong learning and personal development

Traditional lifelong learning existed (community colleges, MOOCs), but participation and retention were low. Brain training apps had moments of popularity, but personalization and evidence were often weak.

AI can make lifelong learning an adaptive journey tied to interests and life stage. It can help retirees repackage experience into coaching or freelance income, support hobbies with personalized tutoring (music, art, gardening), and blend learning with cognitive health and mental wellbeing.

A critical piece is psychological: AI can reduce loneliness by being a consistent learning companion. Lifelong learning becomes less “extra education” and more an ongoing part of identity and quality of life.

6) Higher education and university alternatives

MOOCs and certificate platforms democratized access, but didn’t fully earn employer trust. Universities still largely own credentialing.

AI can weaken the degree’s monopoly by enabling:

• curated micro-credentials aligned to career goals

• scalable mentorship via AI

• “college-in-a-box” models that deliver full curricula globally

• reliable evaluation via projects, writing, and simulations tied to hiring platforms

Universities may not disappear, but they can be repositioned as one of multiple credentialing and learning platforms. AI-native alternatives can emerge with credible, job-linked assessment.

7) Early childhood and Pre-K

Early childhood tools (ABCmouse, PBS Kids) taught basic literacy and numeracy, but personalization was limited and content repetitive.

AI enables adaptive play companions that adjust story, difficulty, and interaction in real time—using a child’s interests (e.g., dinosaurs) to teach math, or using conversation to practice social skills. For parents, AI can recommend activities, track development indicators, and support early intervention.

Pre-K can move from “edutainment” toward truly personalized learning aligned with a child’s developmental curve.

The macro shift: Access → Context & Personalization

The last generation of edtech was mostly about access—free lectures, practice problems, and content distribution. That democratized availability, but the learning experience remained largely static and broadcast-like.

AI-native education is different. The center of gravity shifts to context and personalization:

• it adapts pace and explanation style

• it responds to motivation and psychology

• it remembers long-term progression

• it behaves like a co-pilot, not a library

This unlocks markets that were structurally impossible before: large-scale 1:1 tutoring, career transition guidance, retirement learning, credible alternative credentials, and more.

From an investment perspective, this isn’t a “better app” cycle. It’s a rewrite of the education value chain and economics. Over the next 5–10 years, I expect some of the most compelling Consumer + AI opportunities to come from this category.

II. Startup opportunities

At this point the key question becomes: what kinds of startups get created, where do they wedge in, and how do they differentiate? The opportunity set is broad, but it’s not uniform—each segment has different buyers, adoption cycles, and distribution constraints.

1) K–12 public education: platforms that prove outcomes

Public education is at a structural breaking point: teacher shortages, widened learning gaps post-pandemic, administrative overload, and rigid curricula.

AI’s wedge is concrete:

• teacher copilots that save daily time (grading, planning, reporting, parent comms)

• IEP and special-needs tooling that becomes scalable

• classroom-level personalization systems that make mixed-ability teaching feasible

• policy-aligned copilots that map to standards and testing frameworks

Business model is typically B2B2G (district procurement), which makes compliance, integration, and proof of outcomes essential. Winning products will deliver obvious daily value inside the classroom while fitting regulatory reality.

2) Private tutoring and alternative education: DTC subscription at scale

This segment adopts faster: clear willingness to pay, lower regulatory friction, and a massive existing spend base.

AI-native tutoring platforms can:

• deliver always-on 1:1 tutoring

• personalize by interest and engagement

• enable micro-schools and homeschool copilots

• create immersive “learn-as-play” experiences beyond static gamification

This is where premium DTC models, international expansion (especially Asia), and category-specific champions can emerge quickly.

3) Language learning: the next “AI-native Duolingo”

Language is one of the cleanest AI product surfaces because conversation is the product.

Opportunities include:

• voice-first conversational tutors

• domain-specific language coaching (medical, legal, finance)

• always-on “speaking companions”

• infrastructure APIs embedding language practice into travel, commerce, and communication products

The market supports subscription, and AI makes differentiation immediate via multimodal capability.

4) Workforce reskilling: education becomes labor infrastructure

The most promising startups here won’t look like course libraries. They’ll look like systems:

• AI career navigators that map skills to roles

• simulation-first training environments

• vertical “AI trade schools” for industries

• B2B2E reskilling platforms sponsored by employers, connected to internal mobility

If education links directly to hiring and job performance, the budget unlocks fast.

5) Lifelong learning: wellness + learning convergence

This category expands the frame from “education” to quality of life:

• hobby tutors

• cognitive fitness platforms

• retiree-to-coach marketplaces

• social learning communities where AI supports group retention and wellbeing

This is where new behavior can be created, not just captured.

6) Higher education alternatives: stackable credentials + trusted assessment

Two paths:

• direct-to-consumer “degree alternatives” with job-aligned outcomes

• B2B tooling that helps institutions build AI-native curricula, assessment, and mentorship

Long-term, degrees likely lose exclusivity as performance-based, job-linked credentials mature.

7) Pre-K: AI-native play tutors and parent copilots

Opportunities span:

• interactive AI toys (hardware + subscription software)

• personalized storytelling and language immersion

• parental copilots for activity planning, development tracking, and early alerts

Parents pay for trust, and early childhood is a high-leverage point for lifelong outcomes.

8) Specialized edge cases: niches that add up

Some of the most monetizable markets are “edge cases” that are actually massive in aggregate:

• high-stakes test prep (SAT/LSAT/GRE/CSAT/Gaokao)

• neurodiverse learning support (ADHD/dyslexia/autism)

• corporate compliance and onboarding in regulated industries

These markets often have high willingness to pay or mandatory spend, and AI’s value is tangible.

The big picture: a portfolio market, not a single-winner market

Education + AI won’t produce one platform that eats everything. It’s more likely a portfolio market where category champions emerge across segments—because each segment has different buyers, payers, regulatory constraints, and cultural norms.

The common weapons are the same: personalization, accessibility, and cost efficiency. But distribution and adoption dynamics differ dramatically across K–12 districts vs. parents vs. enterprises vs. global consumers.

The strongest startups won’t just ship tools. They’ll design systems that connect learning, assessment, credentialing, and trust—end-to-end.

III. Market sizing and analysis

1) The overall edtech market

Global edtech is estimated at roughly $163B in 2024, projected to reach ~$348B by 2030 (low-teens CAGR). Some forecasts are materially higher, extending into $800B+ or even >$1T in the early 2030s under more aggressive assumptions.

The direction is clear: edtech is already a large and fast-growing global category. AI likely accelerates that growth by expanding the addressable market beyond content distribution into personalization-led systems.

2) K–12

K–12 estimates vary widely depending on definitions (public vs. private, content vs. tooling, domestic vs. global). Reported ranges span from single-digit billions in narrow definitions to hundreds of billions in broader ones.

The variance matters less than the underlying point: AI can drive adoption at the classroom, teacher, and district level by delivering measurable outcomes and clear ROI. Public adoption is slower but sticky; private adoption is faster and consumer-driven.

3) Language learning

Language learning is one of the largest consumer education markets globally. Estimates commonly place it in the tens of billions today, with large growth forecasts over the next decade. Online language learning alone is often cited in the $20B+ range with strong growth rates.

AI creates immediate product differentiation (voice + context + personalization), and subscription behavior is already proven.

4) Workforce reskilling and corporate learning

Corporate learning is one of the largest spending pools in education, with estimates in the hundreds of billions globally. The structural driver is unavoidable: a large share of the workforce will require reskilling by 2030 as technology reshapes jobs.

This category is budget-rich, but distribution requires enterprise-grade procurement and integration. The winners will tie learning tightly to performance and internal mobility.

5) Lifelong learning and personal development

This market is under-measured, but signals are strong: adult online learning participation has grown sharply since 2019, and aging societies (Korea, Japan, Europe) are creating demand at the intersection of learning, cognition, and wellbeing.

AI can create new markets here—not only capture existing spend.

Closing

Over the last decade, edtech mostly expanded access. Over the next decade, AI will shift education from broadcast to co-pilot.

That transition rewrites the economics of tutoring, career transitions, lifelong learning, credentialing, and early childhood development. It doesn’t just make education more efficient—it makes entirely new experiences feasible.

From a VC lens, Education + AI is not one bet. It’s a layered opportunity set where multiple category-defining companies can emerge—each with different distribution, business models, and adoption curves.

And the tailwind is structural: personalization, lifelong learning, and global access are now converging, with AI as the unlock.

Call for Startups

The purpose of sharing this thinking is straightforward. As an early-stage investor focused on Consumer + AI, I hope this series helps existing startups better leverage AI-driven shifts—and helps new founders reduce trial-and-error as they search for meaningful opportunities.

In that sense, this is Two Cents’ version of a Call for Startups.

If you are an early-stage founder or startup in Consumer + AI and believe you are onto something, my inbox is always open. Feel free to reach out via DM or email:

hur at hanriverpartners dot com

It’s becoming clear that the market’s “readiness” for Consumer AI has crossed a tipping point.

What we need now is to get far more concrete about how AI-driven market change will unfold—the direction, the mechanisms, and the implications for industry structure, competitive dynamics, and the economics between participants.

For founders, the job is to identify those opportunities a little earlier and move first. For investors, the job is to recognize those early moves quickly and support them aggressively.

This series—my “Flights of Thought”—is an attempt to share how I’m thinking through what will happen, what it will unlock, and what kinds of ideas are likely to matter.

Now it’s time to move from macro themes to concrete opportunity spaces by vertical.

I’ll start with commerce, because it’s arguably the clearest category to model—at least on paper. (That confidence may prove wrong, but it’s a good place to begin.)

The beginning of the end

For the last two decades, online commerce has followed a surprisingly stable customer journey. A consumer searches Google a handful of times, visits a dozen sites to compare information, and eventually checks out on one specific store. (Fashion is the exception—more browsing-oriented, closer to scrolling Instagram than “research then buy.”)

In that world, the monetization stack is clear: Google captures search ads, Amazon captures sponsored listings, and creators capture affiliate economics.

AI is now breaking that loop.

First, the purchase journey is starting to close inside AI chat surfaces—ChatGPT, Perplexity, and others—where discovery, recommendations, browsing/comparison, and even checkout can happen end-to-end in a single interface.

Second, if agentic commerce (discussed in [Two Cents #79]) becomes more common, the legacy funnel that begins with Google search will compress further—often into agent-to-agent interactions that don’t require a “search step” at all. The mechanics of that shift tie back to the agent framework in [Two Cents #77].

In the near term, the market impact is showing up through two distinct pathways:

1. the rise of GEO, and

2. the rise of headless commerce.

The rise of GEO

The first phase keeps the current shopping platform structure largely intact—but changes how intent-driven traffic reaches those platforms.

The practical shift is simple: the consumer-brand interface moves from web pages to AI conversations. Marketing strategy has to move with it—from “drive traffic to our site” to “make sure ChatGPT recommends us first.”

That optimization game is GEO (Generative Engine Optimization). It’s the evolution of SEO from keyword/backlink mechanics tuned for Google’s ranking algorithm to structures and signals that AI chatbots and AI agents can understand, trust, and recommend.

You can already see the early motion:

• Google itself is reshaping discoverability via AI experiences.

• A first wave of GEO startups (e.g., Profound and others) is emerging to help brands adapt.

The next phase will look different. As agentic commerce becomes more common, “GEO for chatbots” won’t be enough. We should expect a more agent-native version of GEO—optimized for how agents request data, compare options, and transact.

These two GEO arcs—chatbot-focused vs. agent-focused—will evolve differently. Which one dominates depends on where most consumer purchase journeys end up:

• closed-loop “chat commerce,” or

• delegated “agent commerce.”

My working expectation: chatbot GEO leads in the short term, because it’s already here; agent GEO dominates in the long term, because I’m skeptical that a CLI-like chatbot interface remains the primary consumer UX forever (as discussed in [Two Cents #76]).

The rise of headless commerce

In parallel, there’s a second structural change: if chatbots/agents route consumers directly to product pages (or bypass front-end browsing entirely), the traditional ecommerce front-end becomes less important. Commerce becomes headless.

We already saw the shape of this in early GPT-4 demos—function calling that completes an Instacart order or books travel via Kayak. The interface is no longer “the store.” The interface is the AI. The store becomes an API.

In headless commerce:

• storefront UX matters less,

• product data and checkout APIs matter more.

This is why incumbents are taking different stances on how much access external agents get—product data, inventory, pricing, checkout. As argued in [Two Cents #79], platforms with “default destination” power (Amazon-like) see headless access primarily as a threat, while platforms that compete for distribution (Shopify-like) can see it as opportunity.

What changes, concretely?

The GEO pipeline will likely split into multiple lines:

• We move from optimizing metadata/keywords/backlinks for Google to creating structures that:

  1. make it easy for LLMs to include products in generated answers (and surface them during AI-assisted web search), and

  2. support discoverability and retrieval patterns for agents acting on behalf of intent-driven buyers (agentic-commerce GEO).

• The affiliate/creator layer—historically an intermediary between search and product discovery—will need a major reset in role and economics.

• New transaction networks will emerge to enable commerce discovery and placement inside AI interfaces—effectively ad networks and deal rails built for chat/agent surfaces.

In a headless flow, product data (price, inventory, options) and checkout are exposed via API or A2A interaction, while the consumer completes the journey inside an AI interface. The center of gravity moves away from “shopping UX” and toward “data + permissions + execution.”

Where the market structure can go

If you extend this logic, a few futures become plausible:

1) A “super-default destination” outcome

A small number of default destinations (Amazon-class) evolve into super-platforms by using agents to reach other platforms’ inventories—consolidating distribution even further. Call this the “unification” path.

2) A “democratized product access” outcome

Access to product information across Amazon/Shopify/long-tail merchants becomes more open and standardized. In that world, a new kind of default destination could emerge—one built specifically for agentic commerce.

Both paths create demand for new middle layers:

• Product data aggregation layers that normalize access across fragmented platforms—Plaid-like infrastructure for commerce catalogs and availability.

• Checkout layers that support delegated purchase flows: agent-initiated payments, user-confirmation payments, and agent-to-agent microtransactions—Stripe-like infrastructure for agentic commerce.

• Tooling layers that connect, optimize, and measure the new system: data sync, ranking optimization, analytics, attribution, and policy enforcement.

Startup opportunities

This is still early, and the opportunity set will expand. But even now, a few categories are visible.

1) A2A ad/marketing networks + a new affiliate/attribution layer

The old model was simple: pay Google to buy position. In an agent world, new transaction patterns become possible:

• Buyer-paid research: consumers (or their agents) pay a fee for high-quality information and comparisons—especially for high-consideration categories (lifestyle, cars, homes, “life purchases”). This can generalize.

• Seller-to-buyer-agent incentives: seller/brand/affiliate agents bid—potentially via reverse auction—by explicitly offering incentives to the buyer’s agent to be preferred.

Generalize this and you get:

• “Google Ads for agents”: preferred placement in agent decision loops, and

• new attribution protocols: tracking “who influenced the agent’s purchase decision,” not just who got the last click.

2) Agent platforms for merchants — “Shopify 2.0 for agents/headless”

Merchants may run their own brand agents directly, rather than relying on platform storefronts.

Opportunities include:

• Shopify accelerating into this and winning by default, or

• a new “Shopify for agents/headless” competitor emerging, and

• marketplaces that aggregate headless merchants and expose them to AI agents as the primary discovery channel.

3) Creator economy 2.0

Creators historically sat between search and product pages via affiliate networks. That intermediary role will be redefined.

Creators might:

• be partially replaced by agents,

• become agents themselves (brand-agent-like), or

• operate “creator agents” that carry their voice, taste, and credibility into AI surfaces.

Possible platforms:

• “Linktree for agents” (creators operating their own agent endpoints), and

• discovery platforms for creator/agent entities (another “Shopify for agents/creators” layer).

4) Agent-optimized product information infrastructure + a new commerce aggregator/platform

We’re already watching platforms diverge on agent access policies. If product information becomes meaningfully unbundled from incumbent storefronts, the path from consumer to product changes fundamentally.

A useful analogy: the web unbundled access to information and democratized distribution. A similar unbundling could happen for commerce product information.

If that dynamic accelerates, the prize is enormous:

• a new “head” (new default destination) built for agent commerce that aggregates product databases across platforms, and

• the enabling infrastructure: “Plaid for Agentic Commerce.”

This is the category where the long-term map of commerce could look different a decade from now.

Market size and why this matters

Google is still the front door for a large fraction of online shopping journeys. In 2024, Google’s advertising revenue was ~$274B, and shopping-intent categories (e-commerce, retail, travel, local) are often estimated at ~40–60% of that. That implies $100B–$170B of annual ad spend tied directly to product discovery and comparison.

Amazon also dominates a massive “shopping search ads” pool. In 2024 it recorded roughly $56B in ad revenue, largely from sponsored listings closely linked to purchase intent.

Beyond traditional SEO and ads, affiliate/creator commerce has become a meaningful market:

• global affiliate marketing spend in 2023: roughly $15B, and

• influencer-driven affiliate commerce: roughly $20B (estimated).

Add those together and you’re looking at $180B–$250B per year of commerce-related SEO/ads/affiliate/creator economics that could migrate toward AI-native surfaces and GEO-native rails.

That alone is a “next Google Ads / next Amazon Ads” scale opportunity.

More speculatively (but directionally important): if these shifts change the balance of power among commerce platforms themselves—not just SEO → GEO—then the restructuring could reach much deeper into industry structure across an ecommerce market that’s ~$6T today, with long-term TAM discussions often in the $20T–$30T range. Full restructuring would likely take 10–20 years, but the early phases can move fast.

Closing

SEO defined the fate of online businesses for a generation. Ranking meant revenue. Entire ecosystems of tools and agencies grew around it.

Over the next decade, that center of gravity shifts toward AI-native GEO. In parallel, GEO + agentic commerce can change not only marketing economics, but the competitive dynamics between commerce platforms themselves.

And whenever an industry’s distribution and monetization rails get rewritten, startups get an unusually large window to build category-defining companies.

Call for Startups

The purpose of sharing this thinking is straightforward. As an early-stage investor focused on Consumer + AI, I hope this series helps existing startups better leverage AI-driven shifts—and helps new founders reduce trial-and-error as they search for meaningful opportunities.

In that sense, this is Two Cents’ version of a Call for Startups.

If you are an early-stage founder or startup in Consumer + AI and believe you are onto something, my inbox is always open. Feel free to reach out via DM or email:

hur at hanriverpartners dot com

It’s becoming clear that the market’s “readiness” for Consumer AI has crossed a tipping point.

What we need now is to get far more concrete about how AI-driven market change will unfold—the direction, the mechanisms, and the implications for industry structure, competitive dynamics, and the economics between participants.

For founders, the job is to identify those opportunities a little earlier and move first. For investors, the job is to recognize those early moves quickly and support them aggressively.

This series—my “Flights of Thought”—is an attempt to share how I’m thinking through what will happen, what it will unlock, and what kinds of ideas are likely to matter.

Now it’s time to move from macro themes to concrete opportunity spaces by vertical.

I’ll start with commerce, because it’s arguably the clearest category to model—at least on paper. (That confidence may prove wrong, but it’s a good place to begin.)

Commerce + AI as a new “frontier”

The most important premise: most sellable products are already online. There are still offline pockets, but they likely won’t be the decisive part of the market.

That matters because the AI-driven transformation in commerce is unlikely to be “a new wave of products going online.” It’s not primarily a new Amazon/Coupang—i.e., a new product-page database.

The real battlefield is access and distribution: who controls how consumers discover, select, and buy products that are already digitized and purchasable.

Under that premise, the first major shift is obvious: today’s purchase journey—

Google search → product page (Amazon/Shopify/etc.) → browse/compare → checkout—

will compress into agent-led, increasingly autonomous flows.

In other words: Agentic Commerce moves from concept to default.

Structural shifts driven by Agentic Commerce

The core change is not “better tech.” It’s a change in the actor: the buyer becomes an agent.

Consumers won’t browse individual stores, compare products, and check out manually. They’ll delegate.

“Keep my weekly grocery spend under $100 and restock what I need.”

The agent executes the plan and returns an outcome—either fully purchased, or pre-filled carts waiting for final confirmation.

Once agents own discovery → selection → checkout, the economics of the entire stack shifts:

• Search (SEO / ads)

• Commerce platforms

• Payments

• Retargeting, cart ads, affiliate loops

• Merchant tooling and attribution

In the old world, intent flowed from search → ads → product discovery → checkout via payment processors. In the agent world, intent is captured once—then the agent executes end-to-end.

That means value gets redistributed. Some incumbents lose leverage. New intermediaries appear. Gross margin pools move.

That is where startups get paid.

Early on, we’ll see demand- or vertical-specific agents that lock in user habits. The winners can expand into broader platforms. And we’ll need new infrastructure “rails” to make any of this work: product data access, multi-merchant aggregation, automated checkout, and persistent shopping memory/personalization.

Incumbents are already taking different positions

Even today, large commerce players are reacting very differently—because each one is defending a different moat.

• Amazon is effectively trying to restrict external agent access to its product data—its most strategic asset—while building its own agent layer (“Shop for Me”-style direction). The goal is simple: keep the consumer surface, block the disintermediation.

• Shopify is more nuanced: allow product discovery, but restrict checkout. That’s consistent with Shopify’s incentives—payments and fintech economics are central. Open the top of funnel to developers, keep the bottom of funnel in-house.

• Walmart has been comparatively more open: allow access, let agents drive traffic and sales. That can be rational if you believe openness increases volume—or if you’re fighting from a less dominant position and can’t afford to shut off new distribution.

This divergence creates whitespace. As SEO and search ads lose leverage, the premium shifts to:

• agent-friendly APIs

• multi-merchant data layers

• automated checkout rails

• identity, permissions, and trust primitives

• personalization/memory systems

How the consumer purchase journey changes by category

As a16z has framed it, consumer purchases can be grouped into five broad types—each with different decision dynamics. In an AI-native commerce world, each will evolve differently:

Impulse purchases

From checkout-line candy → TikTok/Instagram “buy in 5 seconds.”

AI’s role is less “research” and more “conversion + friction removal.”

Routine essentials

Detergent, toilet paper, pet food → agent-managed replenishment.

Not subscription-by-calendar, but replenishment-by-state: the agent infers when you’ll run out.

Lifestyle purchases

Beauty, fashion, home—highly taste-driven.

Requires a persistent personalization layer that tracks your history, your preferences, and the market’s shifting trends. This may naturally become an ambient agent behavior.

Functional high-consideration purchases

Laptop, bike, sofa.

The agent becomes a research analyst and procurement advisor: market scan, shortlist, trade-off explanation, and negotiation—often with a brand-side agent on the other end.

Life purchases

Home, car, wedding, college—large, multi-step decisions.

This starts to look like a “butler agent”: research, staged consultation with you, coordination, and even financing solutions.

The implication is straightforward: agentic commerce is not one product category. It is five different problems, with five different UX and monetization shapes—hence five different startup opportunity maps.

What “fully autonomous agentic shopping” can look like

Given those different journeys, autonomous shopping will likely emerge in multiple structures:

1) Personal agents (consumer-facing)

Automate repeat demand end-to-end.

Example: “Under $120/week, restock groceries aligned with my diet.”

Monetization could come from subscriptions, affiliate economics, and recommendation-driven upsells.

2) Category-specific agents

Specialists for high-consideration categories (electronics, travel, furniture).

Beyond checkout automation: research, negotiation, warranties, and human-in-the-loop moments—delivered as a new buying workflow.

This is where startups can own a vertical wedge early—until horizontal platforms (Amazon, ChatGPT-level super-apps) attempt to absorb the behavior.

3) Enterprise / brand agents

Brands and retailers will increasingly put forward their own agents as the frontline sales rep.

These agents will upsell, cross-sell, and personalize in real time. That creates a clear “agent-as-a-service” SaaS opportunity for merchants—structurally similar to the way Klaviyo/Attentive rode the marketing automation wave, except the agent becomes the primary interface, not just the automation layer.

4) Marketplace / platform agents

A more integrated model: a platform agent that coordinates across sellers and brand agents to deliver the entire journey.

Here the AI dynamically composes bundles across SKUs, merchants, budgets, and constraints—effectively acting as a concierge that “just handles it.” At that point, the distinction between “agent” and “platform” starts to blur.

Unlike Amazon/Google’s centralized model, an AI-native aggregator—or a network of brand agents that represent supply—could evolve into a new kind of marketplace.

Startup opportunities that fall out of this shift

These are early sketches—examples of what becomes possible if the structure shifts the way we expect. The real opportunity set will expand as the market evolves.

1) Infrastructure (agent “picks & shovels”)

The deepest opportunities are often the rails.

To enable agents to transact, commerce needs new primitives: data access, negotiation, identity, checkout, memory. Whoever builds these layers becomes foundational.

Examples:

• Multi-merchant aggregation APIs / agents

  Connect to Amazon, Shopify, and long-tail merchants, normalize catalog access, and enable agent-to-agent negotiation. Think “Plaid-like connectivity,” but for commerce inventory and purchasing flows. The exact shape depends on how A2A economies and agentic commerce converge.

• Agent-native payments and checkout

  M2M payments, delegated checkout, programmable permissions. A large portion likely shifts toward stablecoin-like rails or at least non-human-first assumptions. The mental model is “Stripe for agentic commerce.”

• Memory + personalization engines

  Persistent shopping constraints and context: “Only vegan under $50,” “auto-reorder staples when my state indicates I’m low,” preferred suppliers, substitution logic. A “Twilio for consumer context”—a layer other apps build on.

These companies can become the rails that the rest of the ecosystem depends on—like Twilio (communications), Plaid (fintech data), and Stripe (payments) did in prior eras.

2) Consumer apps (autonomous shopping agents)

This is the new “search bar.” Whoever owns the delegation interface owns the consumer commerce surface.

Examples:

• Personal shopping agents for staples, budget control, price tracking, calendar-aware replenishment.

• Cross-platform agents that route across Amazon + Shopify + direct brands—generalizing “Shop for Me” into a multi-platform aggregator.

In past cycles: Honey captured coupons, Instacart captured grocery execution. In this cycle: the winner captures the first consumer touchpoint for buying.

3) Vertical AI shopping agents

The clearest wedge opportunities are verticals where Amazon is weaker and consumers need expert guidance—typically higher-margin, high-consideration categories.

Examples:

• Travel agents that bundle flights/lodging/activities with personalization + real-time price tracking

• Health & wellness agents that manage OTC/supplements linked to personal health context

• Luxury & fashion agents that curate and extend into resale/circular consumption

• Financial product agents that optimize cards/insurance/refinancing with continuous monitoring

Historically, Expedia (travel), Oscar (insurance), and Farfetch (luxury) built large businesses through vertical focus. This time, the “expert layer” becomes AI-native and can replace or redefine the UX entirely.

4) Merchant tools / brand agents (agent-as-a-service)

Brands won’t just run Shopify stores. They’ll run AI agents that interact 1:1 with customers.

Examples:

• Agent storefronts: conversational, personalized first sales rep

• Conversion optimization agents: real-time upsell/cross-sell at checkout

• Post-purchase care agents: warranty, support, replenishment, retention

This mirrors prior merchant SaaS stack land-grabs (support, CRM, marketing automation). The difference is that the agent becomes the front door, not just the workflow layer.

Closing

Agentic commerce is not “ecommerce getting smarter.” It’s a restructuring of the commerce value chain.

As search ads, SEO, marketplace ads, and payment models that powered the last 20 years get rewired, large new profit pools open—and new companies will be built to capture them.

Shopping will move from “search → compare → checkout” performed by humans to delegated, increasingly autonomous execution by AI.

Just as mobile transitions produced mobile-native winners (KakaoTalk, Baemin), this transition will produce agent-native winners—and likely the next wave of consumer unicorns.

This is the starting line.

Call for Startups

The purpose of sharing this thinking is straightforward. As an early-stage investor focused on Consumer + AI, I hope this series helps existing startups better leverage AI-driven shifts—and helps new founders reduce trial-and-error as they search for meaningful opportunities.

In that sense, this is Two Cents’ version of a Call for Startups.

If you are an early-stage founder or startup in Consumer + AI and believe you are onto something, my inbox is always open. Feel free to reach out via DM or email:

hur at hanriverpartners dot com

It’s becoming clear that the market’s “readiness” for Consumer AI has crossed a tipping point.

What we need now is to get far more concrete about how AI-driven market change will unfold—the direction, the mechanisms, and the implications for industry structure, competitive dynamics, and the economics between participants.

For founders, the job is to identify those opportunities a little earlier and move first. For investors, the job is to recognize those early moves quickly and support them aggressively.

This series—my “Flights of Thought”—is an attempt to share how I’m thinking through what will happen, what it will unlock, and what kinds of ideas are likely to matter.

First things first — taking a quick breather after GPT-5.

GPT-5 vs. Consumer + AI

Plenty of people are already going deep on GPT-5’s benchmark gains, its step-up in multi-shot reasoning, and its genuinely strong one-shot coding. I won’t repeat that here.

Instead, I want to focus on the parts that matter most for Consumer + AI—the changes that could reshape product surfaces, distribution, and ultimately consumer behavior.

Two things stood out to me:

1. Mixture of Models (MoM) + a Model Router

2. Tools as “generalized agents”

1) Mixture of Models (MoM) + Model Router

MoE vs. MoM

With GPT-4, scale mattered, but I’d call that mostly incremental. The bigger conceptual shift was that Mixture of Experts (MoE) became broadly understood and adopted as a pattern.

In MoE, the “experts” are structurally similar subsystems with roughly comparable capacity, each specialized by domain. The router’s primary purpose is largely engineering optimization—only activating the necessary experts to reduce inference cost and latency. In that sense, the heart of MoE is “efficient compute.”

GPT-5’s Mixture of Models (MoM) feels meaningfully different. Here, the system is a collection of sub-models that can be quite different from each other—different sizes, different capabilities, potentially different modes of reasoning. The router’s core job is no longer just efficiency. It becomes an orchestrator: deciding how deep the system needs to think, handing work to the right model(s), evaluating outputs, and then delegating follow-on work—sometimes to the same model, sometimes to a different one.

If you compare it to how GPT-5 seems to process internally, it’s effectively multi-shot by design, which naturally supports orchestration across “thinking steps.”

Router as an orchestrator

This is structurally similar to a multi-agent system: break a task apart, delegate pieces to different agents, collect results, and iterate until the desired outcome is achieved.

In other words, the MoM router looks like a generalized version of what today’s agentic orchestration frameworks try to do—except applied across models with different “strength profiles.”

If you generalize further: GPT-5 is implicitly proposing a new default interface for how we will serve consumers (and enterprises) as the number of available models explodes. The world won’t be “one model.” It will be many models—small vs. frontier, reasoning vs. non-reasoning, on-device vs. cloud, multimodal specialists, world models, etc.—composed into a system. The router becomes the single point of entry and the unified UI/UX abstraction for that system—and, importantly, it shapes the economics.

A small but telling observation from using GPT-5 ChatGPT: it often “over-reasons.” For tasks where a lightweight summary would be enough, it sometimes goes full consultant—writing Python code and producing an overbuilt solution. It feels like asking a new hire to pull a quick datapoint and getting a 30-page report back the next morning.

That suggests a few implications:

1. We may need something like a “reasoning temperature”—a consumer-facing control that tunes the depth of reasoning the router chooses (analogous to creativity temperature).

2. Since developers can access individual GPT-5 models via API, it’s likely we’ll soon see alternative UX layers that effectively replace or compete with the default router behavior—especially once the ecosystem starts optimizing for different latency/cost/quality tradeoffs. (This has a bit of “early open model ecosystem” déjà vu.)

3. The bigger opportunity is personalized routing: the router reading user intent and context, then choosing models and depth dynamically based on what this specific user actually wants. If router competition opens up—especially via OSS—this kind of personalization will arrive quickly.

The “AI super app” lens

From a UI standpoint, this is also a step closer to the true meaning of an AI super app.

Open ChatGPT today and you can already feel the direction: the model list collapses, “GPT-5” becomes the default, and soon the user likely won’t care what model is behind the curtain. The interface becomes the product. The model becomes infrastructure.

From there, it’s not hard to extend the thought experiment.

Scenario 1: ChatGPT becomes the primary consumer interface

If ChatGPT can leapfrog existing default surfaces (today: iPhone + Siri), it starts to look like the “AI super app”—or even the “AI super device,” the iPhone of the AI era.

One plausible path:

• Even today, the iPhone Action Button can bypass Siri and open a default assistant workflow.

• A future AI device (the Jony Ive direction) could keep consumers always connected to ChatGPT with always-available voice input.

• Output can remain flexible: for the next few years, the iPhone screen will still be the dominant display, but output could also route to home mirrors/TVs, car systems, ambient speakers, etc.

Scenario 2: personalized LLMs + a personal router

A more speculative—but not crazy—scenario is the emergence of personalized LLMs fine-tuned on a person’s data: calendar, contacts, email, health data, family graph, and more.

And it likely won’t be a single model. You could imagine multiple personalized models by “domain of self”: personal, family, work—each evolving as your job and life changes.

Technically, this doesn’t feel like the main barrier, and cost will likely fall into an affordable range over time. The bottleneck is data: collecting it, de-siloing it, and establishing ownership/control—especially for fragmented first-party ecosystems (health data, playlists, purchase history, etc.).

If personalized models exist—and if “my router” can access and orchestrate them—then a truly personalized assistant becomes possible: not just context-injected personalization per prompt, but a persistent personalized intelligence environment.

In that sense, the model router could be the first step toward that future.

(I’ll go deeper on hyper-personalization in the next post.)

Further generalization: routing as cost optimization

Long-term, the router is not just a UI abstraction. It could become the core mechanism for optimizing cost per unit of intelligence.

Extrapolate the system:

• on-device small LLMs (<10B)

• mid-sized models (100B–300B)

• frontier-scale models (300B–1T+)

• small/large reasoning variants

• multimodal specialists

• world models

A router could choose the cheapest adequate path based on task type and difficulty—possibly with the router itself running on-device as a small model.

If that pattern becomes standard (and it likely will, because it will be easy to replicate and commoditize), then we stop thinking “which model should I use?” The user just uses an AI system, the way we stopped worrying about browser compatibility and OS-level constraints as the stack matured. The model becomes like backend architecture: mostly invisible unless you’re building the infrastructure.

And once AI becomes truly ubiquitous at the consumer level, even the terms “LLM” and “AI” will fade from everyday language—like how we no longer say “I’m using electricity” or “I’m doing the internet.”

2) Tools as “generalized agents”

GPT-5’s tool calling has a few notable characteristics:

• Free-form function calling (CFG-based)

• Parallel tool calling

Free-form function calling isn’t conceptually brand new. MCP servers already resemble this pattern: the input is free text, and the tool interprets and executes actions.

The difference is that GPT-5 generalizes it. Instead of a single clean “prompt → API call → response,” the system can:

• chain tools,

• nest calls (including prompting other tools/models),

• analyze results,

• and decide what to call next.

At that point, “tool calling” starts to look less like API invocation and more like calling into agent subsystems. It’s an abstraction for interacting with semi-autonomous components, not just deterministic endpoints.

This is why some people describe GPT-5 as the “stone age” moment for agents: not because it merely uses tools, but because it begins to think with them—building workflows as part of reasoning.

Parallel tool calling reinforces the same direction. If tools are invoked concurrently and results are processed asynchronously, then the system increasingly resembles a world of many interacting agent-like subsystems rather than a single linear API pipeline.

If you push the idea to its logical extreme:

ChatGPT becomes the primary entry point, and behind it sits a generalized “agent-verse”—a multi-agent system that executes work in the background. That architecture also better supports the UI/UX shifts discussed in Part 2 (where interaction is not strictly request-response, but often headless, proactive, and agent-initiated).

A plausible “normal” flow could look like:

• ChatGPT captures user needs from multiple channels (direct requests, signals, other agents).

• Execution happens through ambient agents interacting asynchronously with platform agents (commerce) and brand agents.

• The system returns options and asks for confirmation only where needed.

Consumer behavior: what changes?

My view is that consumer behavior shift in Consumer + AI will be driven less by “GPT-5 as a new model” and more by the shift toward multi-agent systems.

Once interaction moves away from pure request-response and toward headless + intent-driven workflows, consumer behavior will change quickly—if the system delivers real value: convenience, better selection, better price, less time spent.

And historically, when value is obvious, consumers adapt fast—even when the behavioral change is non-trivial (web → mobile is the clearest precedent).

So the significance of GPT-5, in this frame, is not that it immediately creates new consumer behavior. It’s that it adds key ingredients—routing and generalized tool/agent orchestration—that help ChatGPT become the dominant UI surface. Combine that with the coming multi-agent shift, and you start to see how consumer behavior change could accelerate.

AGI?

Does GPT-5 qualify as AGI?

OpenAI may have strategic reasons to want to declare “AGI reached,” but if we define AGI as “average human-level general intelligence,” I think we still need additional capabilities.

For example:

• Temporally synchronized multimodality: not just seeing/hearing independently, but integrating modalities in a time-aligned way—potentially with a 3D world model.

• Continuous incremental learning (and ideally self-learning).

By that bar, the path to AGI still looks long and steep 🙂

Call for Startups

The purpose of sharing this thinking is straightforward. As an early-stage investor focused on Consumer + AI, I hope this series helps existing startups better leverage AI-driven shifts—and helps new founders reduce trial-and-error as they search for meaningful opportunities.

In that sense, this is Two Cents’ version of a Call for Startups.

If you are an early-stage founder or startup in Consumer + AI and believe you are onto something, my inbox is always open. Feel free to reach out via DM or email:

hur at hanriverpartners dot com

It’s becoming clear that the market’s “readiness” for Consumer AI has crossed a tipping point.

What we need now is to get far more concrete about how AI-driven market change will unfold—the direction, the mechanisms, and the implications for industry structure, competitive dynamics, and the economics between participants.

For founders, the job is to identify those opportunities a little earlier and move first. For investors, the job is to recognize those early moves quickly and support them aggressively.

This series—my “Flights of Thought”—is an attempt to share how I’m thinking through what will happen, what it will unlock, and what kinds of ideas are likely to matter.

Agents are emerging

My base case is that AI-delivered services will rapidly shift toward an agent-based architecture—especially multi-agent systems.

The problem is that “agent” is one of the most overloaded words in AI right now. It can mean wildly different things depending on who’s using it, and in what context.

So this post does two things:

1. Break down agents across a few useful lenses, so we can reason about where each type will be used.

2. Outline the kinds of shifts agents will drive as they become the substrate for AI apps, services, and systems.

We’re still at an early stage. There are more questions than answers. But the direction is becoming clearer.

How to segment AI agents

1) Segment by structure and role

In [Two Cents #71], I proposed five broad types based on how an agent operates and what role it plays.

Human Agents

Agents that replace a human performing a task in the same way a human would, without changing the underlying workflow.

The canonical examples are customer support calls, outbound sales calls, and other human-to-human interactions replicated by an AI.

Auto Agents

You give the agent a goal, and it autonomously executes multiple steps across tools and services to deliver an outcome.

Examples: booking travel, handling shopping, or completing an entire coding task end-to-end (in the spirit of Cursor/Devin).

Workflow Agents

An extension of auto agents. Instead of completing a single task for a user, these agents replace part—or all—of an organization’s workflow over time.

This likely becomes multi-agent / graph-agent by default, with a wide spectrum of human-in-the-loop designs.

One example of the “shape” of this: Stanford experiments where a team of agents operated like a research group over an extended period and produced meaningful scientific outputs (including protein design work relevant to COVID therapeutics).

Ambient Agents

Always-on, cloud-resident agents that proactively work in the background on behalf of a specific user.

They can access personal data sources (email, calendar, health, financial transactions), make decisions autonomously or with human confirmation, and trigger actions.

Virtual Human Agents

Agents whose primary role is to be a human-like counterpart: companions, “AI waifus,” and character-based agents.

Examples: CarynAI, Character.AI personas, and “Samantha” from Her.

This category can expand further—NPCs and characters in virtual worlds and games, or even autonomous “organizations” built to pursue an objective (a hedge fund agent collective, an IP manager, etc.).

(These labels are simply a working taxonomy. New terms will emerge and stabilize—as “ambient agent” already has.)

2) Segment by position inside an AI service ecosystem

In a true multi-agent environment, agents will exist as a population: created, destroyed, and persisted depending on who initiates requests and what roles need to be served.

This isn’t a clean MECE taxonomy—more a practical set of examples.

Personalization agents

• My secretary/concierge agent: handles requests on my behalf, confirms decisions and outcomes with me, and orchestrates execution. (Do we strictly need this to be an “agent”? Possibly not—but the role will exist.)

• My data layer / personalization router: manages access to my personal data sources and responds to personalization requests. This could look like an agent, an MCP-style server, or a dedicated data layer. In some architectures, the data layer itself could be implemented as a multi-agent system.

Open questions here include where this lives and how it operates:

• On-device capture and inference vs. cloud-based ambient agents

• Role split between device and cloud

• Privacy, delegation, and trust boundaries

Service agents

• Platform agents (commerce, travel, marketplaces, financial services): receive tasks, execute them, and return results. In many cases, this might be better represented as an MCP server rather than a full “agent,” depending on interaction patterns.

  If the interaction is synchronous request-response, an MCP server may be sufficient.

  If the interaction is asynchronous—e.g., bidding workflows—agent-based participation makes more sense.

These agents can take multiple shapes:

1. Execute tasks within a platform and return results (the platform interface becomes agent-facing rather than human-facing).

2. Request bids or results from other services/agents.

3. Decompose tasks into sub-tasks, orchestrate multiple agents, negotiate across them, and assemble a final output—effectively acting as an intermediary inside a multi-agent graph.

Brand agents

Agents that represent a brand or service externally—serving other agents.

They might:

• Respond to AI search queries with structured outputs (the GEO world), or

• Submit bids/proposals in response to a user’s project request (leaf nodes inside a multi-agent market).

As the ecosystem becomes more agent-native, we should expect many more agent roles to emerge.

3) Segment by automation maturity

You can also classify agents by capability level, though I’m not sure this is always the most useful lens:

• Glorified personalized help: personalized Q&A

• Reactive execution: performs requested tasks (most “agents” today)

• Proactive recommendations: suggests actions beyond explicit requests

• Proactive action: recommends and completes tasks

• Autonomous workflows: fully end-to-end execution, including collaboration/negotiation with other agents and multi-agent OKR-style work

Other possible segmentation

We could also segment by interaction behavior: initiating agents, serving agents (MCP servers), navigating agents, ambient agents, etc. But it’s not yet clear whether this adds real explanatory power.

What is clear: the division of labor between agents—and the UI/UX for how agents interact with users and services—will remain a major design problem for years.

How agents change the world

1) UI/UX and consumer interaction patterns

The first thing we need to internalize is that once autonomous multi-agent systems (including ambient agents) become real, the interaction model we’ve lived with for decades can change fundamentally.

Modern UX is essentially:

A GUI presents options → a human selects an action → the system responds → repeat as request-response pairs until the user gets what they want.

The core ingredients are:

1. presenting selectable options,

2. actions as request-response pairs,

3. sequences of human-driven steps.

In a multi-agent world—especially with ambient agents—each of these breaks.

The system can become headless, intent-driven, and agent-initiated, with a new balance between autonomy and user engagement. UI becomes less about “navigation” and more about confirmation, exception handling, and control boundaries.

Today, most consumers still can’t directly experience a real multi-agent environment. But early experiments are showing up (e.g., products like Eigent—still developer-leaning at this stage).

Aside: it’s notable how many recent agent-focused projects come from China (Eigent, Skywork, Fellou, Manus). My guess is that China’s hyper-competitive consumer internet environment forces faster experimentation on new interaction patterns.

This new interaction model will take time to reach mass-market polish—but I wouldn’t bet on it taking more than ~12 months before we see commercial products that feel meaningfully different from today’s UI.

2) Business model and market structure shifts

When agents act as delegates—executing work on behalf of humans—the interaction and power dynamics between stakeholders (users, services, agents) will shift dramatically.

This is the engine behind the “fundamental shifts” discussed in [Two Cents #75]. Agents—especially autonomous multi-agent systems—pressure the existing division of labor across platforms, distribution, and monetization.

Commerce

Agent commerce will reshape:

• competitive dynamics among existing commerce players, and

• the emergence of new purchase funnels that bypass incumbents entirely.

We’re already seeing divergent strategies from incumbents:

• Some have no clear agent policy yet.

• Some block third-party shopping agents entirely.

• Others allow discovery but block checkout.

The specific position depends on each player’s moat, market posture, and incentive structure. This will broaden into many more disputes and equilibrium shifts until the market stabilizes under new “rules.”

Search and other consumer categories

Commerce shifts cascade directly into search (because purchase intent is a large fraction of high-value search). And the impact won’t stop there—marketplaces, fintech, and essentially every consumer category will feel the downstream effects.

When market structure rewires, startup opportunity expands. That’s why the phrase “this is the last big opportunity before AGI” resonates—it captures the idea that platform transitions create rare windows where rules reset.

A2A infrastructure: what needs to exist

I covered parts of this in [Two Cents #71], but it’s worth repeating. A multi-agent, A2A world requires new rails and operating norms. Much of it is still underwater, but momentum is building.

Areas likely to become very active:

Payment rails

M2M payment infrastructure for agent transactions.

Stablecoin-based rails are accelerating, and we’re seeing competition across incumbents and startups.

Identity rails

Authentication for agent transactions, proof-of-personhood for delegated actions, and delegation / access management.

Incumbents like Okta may move early, though startups will emerge.

Personalization data layer

Likely the core asset layer in consumer AI.

Security

Permissions, execution scope, and human-in-the-loop controls.

Privacy

Agent access to sensitive personal data (health, finance), and potentially on-chain or self-sovereign privacy architectures.

Tools

Automation and orchestration frameworks, observability, permissioning, and HITL systems.

Additional issues that will matter

Agent identity

• Temporary vs persistent agents (task-based vs daemon-like ambient agents)

• Who the agent represents (user vs service)

• Machine-readable task specs

• Delegation boundaries (e.g., purchase authority)

Agent code of conduct

We already see standards forming for LLM crawling (e.g., llms.txt as an analogue to robots.txt).

As agents gain delegated authority to decide and act, we will need behavioral and interaction norms—early examples like Agent Interaction Guidelines (AIG) point in that direction.

If agents can even “hire” humans to complete tasks, we’ll need conventions for that too.

More broadly, this will be a technical and social co-evolution problem. The system will converge—through trial and error—on norms that are not perfectly optimal, but acceptable enough for most participants.

In a sense, it’s a micro-version of how human societies evolved governance structures over time.

Key traits and implications of the agent transition

Because we are still early, the downstream effects—consumer behavior shifts, stakeholder dynamics, new economic equilibria—are not yet fully visible.

Most “agents” in market today still look like workflow automation with a thin layer of reasoning—call it “Zapier plus.” Even impressive products often feel like “Zapier plus” combined with deeper research modes.

Ambient agents are even earlier. Many features that look “ambient” today are effectively daemon-style automations that draft actions (e.g., email replies) and request confirmation.

The deeper shifts—new UI/UX primitives and new business structures—will emerge gradually as ecosystems become more agent-native. This may take 5–10 years to fully play out. That sounds long until you remember: Uber took ~11 years from founding to IPO.

What will drive the structural change?

Machine-to-machine interaction (M2M)

Agents interacting with services or other agents flips a foundational assumption: the “decision-maker” is no longer a human. That changes discoverability, flows, checkout, and incentives. Entire systems will need to be redesigned around this assumption.

Autonomy

Agents can make and execute decisions—sometimes material ones (format of requests, purchase choices, checkout)—which forces new primitives: authentication, delegation, privacy, liability boundaries, and new revenue-sharing models.

There are likely more drivers, but these two alone are enough to force rewrites across the stack.

Over the next 12–18 months, I expect the agent category—especially multi-agent A2A—will be one of the most dynamic areas in AI. We’ll see rapid experimentation, real failures, and iterative construction of the infrastructure and norms needed for a sustainable equilibrium.

Some of this will sound like science fiction. But it’s increasingly “SF” in the other sense: San Francisco, happening now. (Yes—pun intended 🙂)

Call for Startups

The purpose of sharing this thinking is straightforward. As an early-stage investor focused on Consumer + AI, I hope this series helps existing startups better leverage AI-driven shifts—and helps new founders reduce trial-and-error as they search for meaningful opportunities.

In that sense, this is Two Cents’ version of a Call for Startups.

If you are an early-stage founder or startup in Consumer + AI and believe you are onto something, my inbox is always open. Feel free to reach out via DM or email:

hur at hanriverpartners dot com

It’s becoming clear that the market’s “readiness” for Consumer AI has crossed a tipping point.

What we need now is to get far more concrete about how AI-driven market change will unfold—the direction, the mechanisms, and the implications for industry structure, competitive dynamics, and the economics between participants.

For founders, the job is to identify those opportunities a little earlier and move first. For investors, the job is to recognize those early moves quickly and support them aggressively.

This series—my “Flights of Thought”—is an attempt to share how I’m thinking through what will happen, what it will unlock, and what kinds of ideas are likely to matter.

Let’s start with the most fundamental question: where consumers will “play”, and what the primary UI surfaces and UX patterns will look like in that world.

The Primary UI Surface for Consumers

A few questions worth asking—thinking far forward:

Where will consumers primarily interact with AI? On the smartphone? Back to desktop? Or in ambient environments—cars, offices, living rooms, bedrooms, subways, sidewalks—spaces saturated with microphones, speakers, cameras, and sensors? Or through ambient devices—pins, AR glasses, wristbands?

And what will the dominant modality be in those environments? The current pattern—staring at a phone screen, uploading photos, typing? Voice conversations on the phone? If so, is the “counterparty” Siri? Or an ambient device’s Alexa-like interface? Or direct voice conversations with an AI app while looking at a screen? In an ambient setting, do we talk to a local device that routes requests, or does the cloud agent triage everything by default?

Before we reach that future, what becomes the dominant interface within today’s constraints? Does Siri (as a voice agent) become the primary router, with everything else running behind it? Or do ChatGPT/Claude become the new super-apps—and effectively the new app stores? Or do users continue to discover services by hunting for individual AI apps and URLs? Or does the browser reassert itself?

With those scenarios in mind, let’s return to the present and look at what’s already happening.

Battle for Attention

Several competing moves are unfolding in parallel.

1) LLM chatbots racing to become the “AI Super App” and the AI app store

ChatGPT and Claude are trying to own the consumer’s main UI surface.

They’re making the chatbot the default interface—and then expanding what’s reachable inside it: prompt-based apps (GPTs, artifacts) and workflow agents (Computer Use, Operators). Functionally, both are “AI apps.” The platform that aggregates them naturally becomes the app store layer as well.

If an LLM chatbot becomes the default super-app + app store, it captures the two most strategic assets at the UI layer—similar to Chrome on desktop or WeChat in China. Control the default surface, and you control consumer access to the broader Consumer + AI universe.

Even leaked strategy memos point in the same direction: the goal is to become the “super assistant” that functions as the primary interface to the internet.

At the same time, for the near term, the smartphone still acts as each person’s “server/center”—where activity accumulates and identity, permissions, and data live. And Apple/Google are unlikely to allow full OS-level third-party access to that center. Which means LLM chatbots will not have default access to the system agent (e.g., Siri) in any straightforward way.

Aside: this is why I suspect the OpenAI–Jony Ive device effort is about bypassing that constraint—potentially creating a hardware-level shortcut to become the default input layer (e.g., a microphone tethered to ChatGPT), or a path to “intercept” agent routing without OS-level permission.

2) The competing race: AI browsers

A second group is trying to win the UI surface through the browser itself.

We’re already seeing entrants like Perplexity’s Comet, The Browser Company’s Dia, and AI-browser-like products from Genspark and Fellou. OpenAI is also rumored to be working on an AI browser.

Aside: many “AI browsers” today feel closer to a repackaging of bookmarks and extensions than a truly independent browser stack. They may look like browsers in the current phase, but the harder question is whether these pseudo-browsers can actually hold the consumer surface once the real competition shifts into ecosystem capture—i.e., the AI app store battle. My read is that some players (e.g., Genspark) still look more like “agent platforms” than full super-app contenders, and it’s not obvious a startup can pursue both strategies at full intensity without stretching resources.

AI browsers can also build internal ecosystems—skills galleries, agent directories, app catalogs. Dia has already started; others are moving in a similar direction.

At the core, this is the same war as the chatbot war: who owns the consumer’s access path to the explosion of “AI apps” arriving as agents and prompt workflows.

This is still largely desktop-led today. Mobile versions will come. But browsers on mobile are constrained by OS-level limitations, which raises the question of whether an AI browser can realistically become the mobile super-app.

WeChat is the historical exception, but it required extraordinary, context-specific concessions from Apple early on due to China market dynamics. Given how uncomfortable that later made Apple—and how the power balance shifted—it’s hard to imagine that kind of exception happening again.

3) The mobile gatekeepers: Apple and Android

Apple hasn’t moved aggressively yet, but Google has. If you watch Google’s AI Mode demos, you can see deeper integration into the smartphone experience: beyond Lens, into live camera input, and into “what’s happening on the screen” across both desktop and mobile.

This matters because it makes two things possible for the gatekeeper:

1. Taking action on everything happening on the screen (search is the first obvious action; Apple will likely pursue a different mechanism).

2. Capturing that entire activity stream as data.

And that leads directly to the long game: attention control plus a proprietary personalization data layer—potentially fused with first-party data like Apple Health. In the long run, personalization data is one of the most durable moats in Consumer AI, and the OS gatekeeper is structurally best positioned to capture it.

The big variable is the one we started with: does the smartphone remain the primary consumer interface “until it isn’t”? If the surface shifts to new ambient devices (OpenAI–Ive, Meta AR glasses, or new startups), the outcome of this long game could change.

Multi-agent environments

Multi-agent systems are spreading fast in prototypes and experimentation—but consumers still have limited direct touchpoints.

Most multi-agent workflows will reach consumers through one of three routes: standalone AI apps, aggregators/super-apps, chatbot app stores, or AI browsers (desktop and mobile).

In that world, multi-agent apps are typically not the primary surface—they are the content that the surfaces fight over. That said, there is always a possibility that a breakout emerges and becomes the surface itself—WeChat-style—especially if it avoids being packaged as a traditional mobile app that requires App Store distribution.

Ambient Agents and “ambient intelligence”

Today’s agent ecosystem is still mostly workflow agents: agents that execute a defined task, sometimes using computer-use patterns or connectors to external apps and data sources.

Cloud-resident ambient agents—always-on, continuously context-aware—likely need more time. But the direction is clear. As A2A systems evolve into multi-agent and autonomous systems, ambient agents become a plausible dominant delivery model.

If I force myself to imagine a five-year horizon: consumer-facing workflow agents (interacting through phone or ambient devices) and cloud-resident ambient agents may end up at roughly similar scale—within a factor of ~2.

If the shift to ambient agents becomes that large, then the “consumer interface” must change even more dramatically than the desktop-to-mobile transition did.

UI/UX in an Agent World

Assuming the world moves quickly toward Agent-to-Agent (A2A) interaction, what does UI/UX look like?

A working taxonomy of agents

Agents likely segment into a handful of types:

• Human Agents

• Auto Agents

• Workflow Agents

• Ambient Agents

• Virtual Human Agents

This isn’t a final taxonomy—new terms will emerge and settle (as “ambient agent” already has). Workflow and ambient agents generally assume a multi-agent environment.

You could also segment by interaction role—initiating agents, serving agents (MCP servers), navigators, autonomous agents—but it’s not yet clear that added granularity is practically useful.

UI/UX for “AI apps” and agents

I’ll use “AI app” broadly: anything that executes a function the user wants—mobile/web apps, agents, prompt workflows, etc.

What forms might AI apps take, and how do users reach them?

AI app as a visible surface

1. Standalone app or URL: a mobile/web app discovered through app stores or the browser.

2. Inside an AI super-app: accessed via something like /agent-name or @agent-name inside the chatbot.

3. Through an AI app store directory: GPTs, artifact directories, skills galleries, agent directories embedded in chatbots or AI browsers.

These routes likely cover most auto/workflow agents.

Headless execution

1. Daemon-like cloud agent: initiated server-side without a direct UI surface.

2. Initiated through a consumer-facing router agent: Siri/Alexa-like master agent, on-device or cloud-based.

This is the natural pattern for ambient agents.

But headless agents still require a human control surface: a dashboard, feed, confirmations, exception handling—possibly via messenger, a dedicated app, a master agent, or integration into existing super-apps.

Character-like interfaces

Some agents may live as “characters”—profile-pic-like identities, game-avatar metaphors, or persistent voice-mode embodiments. Virtual human agents will likely appear this way, embedded inside other apps and AI experiences.

And beyond these, we should expect many new interaction patterns to emerge.

Voice UX and ambient devices

Voice UX (especially headless + conversational) and ambient device UI (pins, AR glasses) will not behave like mobile/web apps. They may become the next primary consumer surface.

This is why big tech players without existing consumer surface dominance—Meta and OpenAI in particular—are betting hard here. A new tech wave can create new behavioral shifts, and those shifts can create new UI surfaces up for grabs.

A few implications for voice and ambient UX:

• The user’s ability to make explicit “choices” via GUI becomes limited or disappears. Think fewer options than an IVR menu.

• Voice is temporal—interaction takes time—so instead of presenting many choices, the system will likely optimize for capturing intent and executing quickly.

• The classic UX model—guiding users through fragmented GUI actions so they can “find and choose well”—loses power.

Key words here: personalization, defaults, intent.

To make this work, we’ll need new UI/UX patterns—likely agent-native—and a deep personalization data foundation that can reconstruct user workflows end-to-end.

This will not resolve quickly. It’s a multi-year design and iteration problem.

The personalization data layer / agents

A central layer will emerge that either aggregates personal data or routes access to it.

Data sources include: calendars, email, messaging, social graphs, photo libraries, files, and behavioral logs (purchases, services, travel bookings, course consumption, etc.).

Forms could include reactive agents, ambient agents, MCP servers, or explicit data layers.

Key words: personalization (depth and type), privacy, trust, data ownership/sovereignty.

On-device vs. cloud-based

Personalization will likely be hybrid.

On-device

• Parts of “my agent” live locally.

• On-device capture becomes critical: workflows, screen state changes, user choices, messaging data.

• An on-device personalization model is also plausible.

• Hybrid pipelines: capture locally, process/store selectively, personalize across layers.

Cloud-based

• Backend processing and inference over captured data.

• Handling cloud-resident personal data (email, messaging, photos, files) and proactive actions (drafting emails, requesting confirmations for purchases/tasks).

• Natural home for ambient agents.

Examples and projects worth noting:

• mem0, Context: tool-layer approaches that try to unify personal data.

• Apple ReALM: an internal Apple effort to use on-screen context and user reactions to improve Siri through personalization data.

• Scribe-style workflow recording: if workflow recording expands into default capture, then recording + vLLMs could become a powerful platform for understanding and personalizing individual workflows—though it lacks the OS-level privilege Apple has.

Aside: if I let my imagination run: this “personalization data layer/agent” category feels like one of the few places a true “second-generation model”—Google-scale—could emerge. Today we still lack clear answers on data capture (OS-level first-party collection is ideal; third parties face constraints) and on how to translate raw data into durable personalization. But Web 1.0 also looked unclear before the two enabling breakthroughs—PageRank and keyword ads—made Google inevitable.

Key takeaways

• Agents are spreading into B2C faster than most people in Korea currently feel on the ground. Today, consumer AI experiences still split between native web/apps and “browser-first” distribution (often as Chrome extensions). As AI browsers and chatbot-based agent ecosystems scale, we should expect a rapid migration toward those surfaces.

• Native app/web vs. browser/chat-based AI apps have fundamentally different discoverability mechanics, which implies fundamentally different GTM and marketing strategies. Early GPT stores remind me of the early App Store era—when trivial apps could explode because distribution was wide open. That said, as with the early App Store, not all of this is sustainable.

• Beyond incumbents like Naver/Kakao, consumer AI players with meaningful DAU/MAU (e.g., WRTN, Liner) may want to explore the emerging “AI app store” opportunity and the personalization layer (agents/data layers). But there are still many open questions: how the market structure settles, how Korea vs. global dynamics evolve, and what becomes truly durable.

• Two emerging “truths” seem directionally right so far:

  1. AI apps don’t naturally have strong direct network effects.

  2. Velocity/momentum can be a moat—until it isn’t.

  Momentum is often time arbitrage, not inherent defensibility. Some leaders will convert momentum into data moats or indirect network effects, but momentum alone is not durable. Consumer apps currently winning on velocity (including Cluely, WRTN, Liner) need a deliberate strategy to translate that lead into sustainable moats.

• Voice UX remains a large opportunity—with real difficulty. Whether Korean-language “locality” becomes a meaningful moat (as it sometimes did in Web 1.0) is still unclear. But there’s also a reverse argument: starting outside the U.S. can force multi-language excellence by default. ElevenLabs starting in Poland is a good reminder that “local constraints” can become global product strength.

• In every tech shift, UI surfaces and UX patterns changed massively—and that created enormous opportunity (and many casualties). This cycle will be no different. For incumbents, it’s a threat. For startups, it’s a generational opening—especially in the personalization data layer/agent category.

Call for Startups

The purpose of sharing this thinking is straightforward. As an early-stage investor focused on Consumer + AI, I hope this series helps existing startups better leverage AI-driven shifts—and helps new founders reduce trial-and-error as they search for meaningful opportunities.

In that sense, this is Two Cents’ version of a Call for Startups.

If you are an early-stage founder or startup in Consumer + AI and believe you are onto something, my inbox is always open. Feel free to reach out via DM or email:

hur at hanriverpartners dot com

It’s becoming clear that the market’s “readiness” for Consumer AI has crossed a tipping point.

What we need now is to get far more concrete about how AI-driven market change will unfold—the direction, the mechanisms, and the implications for industry structure, competitive dynamics, and the economics between participants.

For founders, the job is to identify those opportunities a little earlier and move first. For investors, the job is to recognize those early moves quickly and support them aggressively.

This series—my “Flights of Thought”—is an attempt to share how I’m thinking through what will happen, what it will unlock, and what kinds of ideas are likely to matter.

Fundamental Shifts

The first thing to pay attention to is that AI is driving structural change in the internet and mobile-era playbooks we’ve all internalized.

This isn’t just about new startups launching new products. It’s about the counterparties startups interact with—platforms, distribution channels, partners, competitors—shifting underneath them. When that happens, the economics of competition and transaction change as well. The relationship map changes, and with it, the revenue model and unit economics a new company can realistically build on.

So before debating specific “startup opportunities,” we need to understand the foundational shift in market structure that’s now underway.

Search and SEO

It’s already obvious that Google’s search-driven SEO structure is changing at the root.

Yes—search via chatbots like ChatGPT is still tiny relative to Google (on the order of a fraction of a percent). But even just adding AI Overviews has materially altered organic traffic flows to blue links—by something like 30–60% in many cases. If conversational search (e.g., AI Mode) becomes the default interaction pattern, it’s not hard to imagine the first real cracks forming in Google’s historically unassailable position.

In commerce—where visibility is oxygen—we’re already seeing early movement toward AI SEO, and the shift from “SEO” to what many now call “GEO” is increasingly inevitable.

More importantly, the Agent-to-Agent service architecture that’s now starting to spread will reshape discovery, selection, and transaction across most online commerce. I’ll go deeper on Agent-driven markets in later posts; here I’ll focus on the shape of the shift.

The biggest change is simple: the “actor” executing a transaction moves from humans to agents, turning many interactions into machine-to-machine (M2M) flows. Once that happens, every business model built on the assumption that “a human will browse, compare, and click” is forced to change.

Google Search’s original logic is: show a ranked list of relevant links, assume humans will evaluate them, and compete to be the link humans click. SEO is the arms race to win that ranking and that click.

Two things have now changed:

1. The decision-making layer for “which result matters” is shifting from human evaluation to AI mediation.

2. The place where discovery happens is shifting away from Google toward other surfaces—LLM chat, AI-native search products, and agent interfaces.

That is the core of the SEO → GEO transition.

At a deeper level:

• Information retrieval is changing from “collect fragments via links and synthesize yourself” to “receive a synthesized answer (and, when needed, deeper research and analysis) as the default interaction.”

• Even more significantly, the highest-value part of search—the path from purchase intent to purchase decision—is increasingly likely to migrate away from search as we’ve known it. Instead of users actively searching for the best product, AI and agents will increasingly find and propose the best option on the user’s behalf.

That’s why this isn’t just an SEO story. It’s a search business model story—and potentially a major re-platforming of the entire purchase funnel.

Commerce and Marketplaces

The same pattern is playing out in commerce. With “agent commerce,” the entity converting intent into action shifts from human to agent, and the logic of discovery and selection—and the stakeholder map—changes fundamentally.

There are two structural shifts to watch.

1) What happens when incumbents absorb agents.

The early example is Amazon’s “Buy for Me”-style agent shopping direction: consumers can use Amazon as the default destination, but access third-party inventory and complete the full purchase cycle. In that world, Amazon becomes an even stronger front door to commerce—while also creating a strategic dilemma: if third-party inventory is reachable through Amazon, then Amazon’s moat in “exclusive destination value” can weaken at the margin.

As agent commerce becomes real, it will impact everything tied to the purchase cycle: product pages, keyword ads, brand marketing, and retail search optimization. And this matters because retail ads are not a side business—Amazon’s shopping ad engine is enormous and contributes disproportionately to profit. This structural shift is a real threat to incumbents, but it will also create large whitespace for startups (the exact shapes are still emerging).

2) The shopping “surface” itself is moving.

The next shift is commerce moving into non-commerce platforms—LLM chat products and AI-native apps—where the entire loop from discovery to checkout is completed inside the AI interface.

In the classic flow, shopping is distributed across multiple layers: search (Google/Amazon/comparison sites), product pages (commerce platforms), browsing and comparison, checkout (wallets/PSPs/platform checkout).

In the AI-native flow, that whole loop can collapse into one interface: search → discovery → decision → checkout, happening inside ChatGPT/Gemini/Perplexity or a new AI app layer.

Greg Brockman’s early ChatGPT-4 demo—planning a meal, ordering groceries via Instacart, reviewing a cart, then completing checkout—was a clear preview: B2C platforms risk being pushed into the role of backend execution layers behind the AI front-end. Commerce is now where that shift begins to feel tangible.

Zooming out: this doesn’t stop at commerce. It extends to marketplaces and any platform that intermediates discovery, connection, and transaction. Generalized further, it implies that many B2C platform playbooks built over the last 30 years—from search to commerce to marketplaces—could see their operating models fundamentally rewritten.

Compared with that, the web-to-mobile transition looks relatively incremental: same democratized distribution structure, new form factor, richer modalities (always connected, location, camera). The AI transition is more profound—it replaces the division of labor among market participants with a different process and a different structure.

Software

Software and SaaS won’t be exempt.

Coding agents and “vibe coding” are already changing what software creation even means. And once the production function changes, the SaaS market structure has to change with it.

The emergence of AI-native startups that hit scale with radically different headcount, timelines, and go-to-market dynamics is a signal that the old SaaS playbook—ARR milestones, T2D3, conventional LTV/CAC heuristics, net dollar retention as the north star—won’t map cleanly onto what’s coming.

We’re watching a new playbook form in real time: sell work, not tools, price outcomes, rethink what it means to “build,” “serve,” and “use” software.

At an even deeper level, it’s fair to ask whether the term “SaaS” will still be the right abstraction a decade from now—or whether the concept of “software” itself shifts.

(Aside: one way to think about “software” is that it has historically been humans doing the work to translate intent into the limited language computers can understand—machine code to assembly to higher-level languages to scripts. We’re now reaching a stage where computers can increasingly understand work described in natural language and execute it. If the worker becomes far smarter—and has rich context about you, your data, your preferences—it starts to resemble a “great operator” or even an old-school butler: you don’t need to specify everything precisely; you express intent, and the system figures out execution.)

Net-net: across categories, the playbooks we’ve relied on for the last 20 years are being rewritten.

The specific form differs by sector, but the common thread is: the old playbooks are not valid in an AI-native economy. We have to write new ones.

Key Takeaways

Over the last 10 years (mobile and SaaS), and arguably the last 20+ years (web), we treated many industry playbooks as “default truths.” Most of them are no longer safe assumptions.

Trying to grow by layering incremental AI features onto existing products—while staying inside the same old distribution, monetization, and organizational frameworks—will increasingly fail.

One thing that has frustrated me over the past 2–3 years is how many early consumer startups were still chasing problem statements that looked structurally similar to what we saw over the prior decade. My hope is that the “AI opportunity set” pushes more founders to look for opportunities in fundamentally new frames.

In this environment, what matters is less “what happens if we tweak the old system?” and more:

• What structural shifts are happening (or will happen)?

• What are the first-order and second-order effects?

• What new solutions become necessary—or newly possible—because of those shifts?

That kind of first-principles thinking is not optional. It’s required.

Across every category, we should question the default assumptions, analyze the underlying structure, and move from “finding” incremental value to building new value from the foundations up.

Aside 1: Incumbents Won’t Simply Be Disrupted

One important difference versus prior platform shifts is that incumbents may not just be targets of disruption—they may remain meaningful actors.

We already saw this in enterprise AI: Microsoft, Adobe, Oracle, Databricks, and others have adapted faster than many expected.

Why?

At the macro level, incumbents have learned from previous cycles, and their tech literacy is materially higher than it was in earlier transitions.

More concretely, in enterprise AI, incumbents often control the customer’s data pipelines and workflows. That creates a real data and distribution moat versus new entrants.

That said, the consumer market is different. Consumer opportunity is more tightly tied to behavioral shifts enabled by tech shifts, and consumer lock-in to legacy data infrastructure is weaker. In newly created consumer markets, incumbents are often less equipped to adapt—and new entrants are more likely to define the category.

Aside 2: AI Data Center Build-Out Feels Familiar

The current “$1T AI data center” conversation looks structurally similar to the telco infrastructure cycle from around 2000—fiber build-out, data centers, overhang, then consolidation.

The scale is larger—maybe 10–100x—partly explained by inflation and partly by compute intensity. But the pattern rhymes.

It wouldn’t surprise me if, 20 years out, we see echoes like: fiber glut → power glut, and telco consolidation → AI data center consolidation.

That doesn’t mean the opportunities and risks in this build-out are irrelevant. It just means: this infrastructure story is important—but it is not “the entire world.” It’s a major layer, not the whole stack.

Call for Startups

The purpose of sharing this thinking is straightforward. As an early-stage investor focused on Consumer + AI, I hope this series helps existing startups better leverage AI-driven shifts—and helps new founders reduce trial-and-error as they search for meaningful opportunities.

In that sense, this is Two Cents’ version of a Call for Startups.

If you are an early-stage founder or startup in Consumer + AI and believe you are onto something, my inbox is always open. Feel free to reach out via DM or email:

hur at hanriverpartners dot com

It increasingly feels like the market’s readiness for Consumer AI has crossed a tipping point.

We are no longer just seeing isolated, novel services popping up in scattered corners of the market. Instead, we are entering a phase where fundamental discussions about reshaping entrenched industries and business models—SEO, marketplaces, commerce, and beyond—are happening in earnest.

Alongside new categories of consumer services (e.g., Daybreak, Rosebud, Pika Social), we are also seeing rapid emergence of the infrastructure that supports them (memory layers such as mem0 and Context, payments), as well as new methodologies (multi-agent systems, ambient agents). Some categories—most notably AI companion services—are already showing explosive revenue growth, hinting at what may become the first true Consumer AI killer app.

This is not yet a universally observable, dominant market shift. But across nearly every sector, we are clearly in a phase where countless early shoots are emerging simultaneously.

After nearly three years of post-ChatGPT-3.5 turbulence, parts of the AI market are beginning to settle. In LLM infrastructure, key players are emerging. In the enterprise market, coding has proven itself as the first undeniable killer app. Certain verticals such as legal are also starting to show clearer market structure. In other words, visibility is finally improving across segments of the AI landscape.

And while first movers do not always become ultimate winners, some Consumer AI categories—again, AI companions being the most notable—are beginning to demonstrate the potential to become the consumer-side equivalent of coding in enterprise AI.

Thoughts on the AI Tech Wave

As part of the broader AI technology wave, how might Consumer AI meaningfully take shape and scale? One useful mental model is to revisit how Web 1.0 unfolded in the late 1990s.

First, building on today’s installed user base and LLM infrastructure, we should expect a flood of experimental services across many domains, mushrooming rapidly as teams test market demand.

From a timing perspective, the U.S. market is likely to see a surge of such experimental products in H2 2025 through H1 2026, all competing aggressively for early market leadership. Korea will likely follow with a 6–18 month lag, initially through benchmarking and copy-driven execution.

This lag differs from the mobile transition era, when Korea tracked the U.S. much more closely. Mobile was largely a form-factor and UX shift, not a deep restructuring of industrial or business models—whereas AI clearly is.

Most of these experiments will fail. A small fraction—perhaps 1–10%—will emerge as early market winners. Among them, some will compound into enduring leaders (Amazon-like), while others will dominate temporarily before handing the baton to second-generation models (Yahoo, AltaVista-like).

Separately—and importantly—in the Korean market, beyond this “following the global market” dynamic, we expect the emergence of entirely new, native consumer experiences, particularly centered on play. As seen in iloveschool, SayClub, and Lineage, these tend to emerge at the intersection of content, gaming, social, and entertainment.

In this domain, innovation often proceeds independently of U.S. trends—and in some cases flows back into global markets. The virtual item–driven freemium model pioneered by SayClub and PMang later became the dominant global monetization model for digital content over the next two decades.

At Han River Partners, we refer to this domain as “Spectrum of Play”, and treat it as a distinct investment area within the broader consumer sector.

Naturally, the infrastructure and tooling required to support these shifts will emerge in parallel. In areas such as agents, personalization, multimodality, and UX, we should expect fundamentally new approaches—not incremental extensions of existing paradigms.

As with keyword advertising in Web 1.0, new business models and infrastructure layers will enable second-generation winners to emerge—Google, Uber, and TikTok–like outcomes in the Consumer AI era.

This inflection point is precisely where Han River Partners is most focused as an early-stage investor in Consumer + AI.

AI vs. Web and Mobile: What’s Different?

To frame this more clearly, it helps to contrast GenAI with prior platform shifts.

Web 1.0 introduced technologies (web protocols, HTTP) that made previously impossible things possible, fundamentally restructuring information distribution and enabling new business models.

Mobile, by contrast, was largely an incremental expansion—new form factors and UX layered onto existing capabilities, with added attributes such as location awareness, always-on connectivity, and ubiquitous cameras, which in turn shaped new but still incremental user behaviors.

AI is different. It makes existing tasks orders of magnitude faster and more efficient, while simultaneously enabling things that were theoretically possible but practically infeasible—video generation, music creation, and beyond. Doing familiar things dramatically better at scale effectively unlocks vast new categories that previously could not exist.

Conceptually, AI functions as an amplifier of intellectual capability. Just as the Industrial Revolution automated physical labor and amplified material production for over two centuries, AI will play a similar role for cognitive labor.

Understanding where AI aligns with—and diverges from—web and mobile is critical when evaluating how it reshapes opportunities and market dynamics.

Keywords to Explore in Consumer + AI

The work ahead is to think concretely about how AI-driven change will unfold, what it means for industry structure, competitive dynamics, and economic models—and to identify opportunities slightly ahead of the curve.

For founders, this means starting earlier with sharper hypotheses. For investors, it means recognizing and supporting these efforts sooner.

I plan to share my evolving “Flights of Thought” on these dynamics—across markets, technologies, and business ideas—through a series of essays. (The title, fittingly, was suggested by ChatGPT.)

The current plan is to publish one theme per week, covering market structure, observed signals, or emerging business ideas, and to extend these discussions into offline forums such as small AI salons. The list below may evolve—items may be added, removed, or reordered.

Themes to explore:

• Fundamental shifts in existing playbooks

• UX in Consumer AI

• What exactly is an Agent?

• Hyper-personalization

• AI companions & social: the first Consumer AI killer app?

• Entertainment: reborn or disrupted? (“Spectrum of Play”)

• Creativity unlocked at consumer scale

• Consumer AI by segment — Commerce, Education, Social, Healthcare, Fintech …

• Required infrastructure layers

• Consumer needs and primal instincts: AI as companion vs. solution

• Where tech shifts intersect with behavioral shifts

Call for Startups

The purpose of sharing this thinking is straightforward. As an early-stage investor focused on Consumer + AI, I hope this series helps existing startups better leverage AI-driven shifts—and helps new founders reduce trial-and-error as they search for meaningful opportunities.

In that sense, this is Two Cents’ version of a Call for Startups.

If you are an early-stage founder or startup in Consumer + AI and believe you are onto something, my inbox is always open. Feel free to reach out via DM or email:

hur at hanriverpartners dot com