Consumer agents

Consumer AI has an ARPU problem.

This includes the ChatGPT consumer app: it’s obviously useful, but in many ways thin.

ChatGPT’s consumer app is, for the most part, an answering machine. Many other consumer AI apps are the same: ask a question and AI will answer it.

These “thin” applications create tremendous value for consumers, from parents asking for health advice to kids cheating on homework, but the platforms struggle to capture it. Even their recent moves beyond an AI engine to, for example, personal finance, are still based on a foundation of simple querying (“ask questions to your bank statements”).

But US consumers spend only around 3% of their wallet on digital goods and services. This is made up mostly of streaming, gaming, and other entertainment apps (I wrote about the rise of Reelshort and dramaslop here).

So far, consumer AI has been competing with that 3%; it lives within the realm of the online. This is what drives the ARPU problem for consumer AI; consumers might pay $20/month each for ChatGPT, Claude, as well as perhaps an AI language tutor or AI game.

But very few people will pay $200 per month for a chat experience, no matter how smart the model underneath it is.

The solutions to this problem are being built, and they can be called “consumer agents”.

Consumer agents, like their B2B counterparts, are applications that do work on behalf of the user, and where things progress beyond the user’s direct control.

A consumer travel application tells you which hotel to book. A consumer agent books the hotel, changes it when your flight is delayed, uses your points, coordinates with friends, and gets a refund when it’s needed.

A consumer education application teaches you English via a game. A consumer agent becomes your tutor with realtime avatars and voices, reminds you of words you messed up last time, gives you quizzes to drill on them, and asks you questions personalized to you.

A consumer video application generates captions for your video. A consumer agent edits your videos, posts them, tracks performance, manages brand deals, sends invoices, and helps sell products while testing hooks in parallel and reporting back.

The distinction is not about how complicated the interface is. In fact, the best consumer agents might feel extremely simple. The interface can still be a messaging app, chat box, a camera, a voice note, or a single button.

Even games can be consumer agents. Status AI (AI social media) and Praktika (AI language learning) are both vertical apps, but the user only sees the tip of the experience; the agent is doing huge amounts of work in the background to create a personalized experience in both cases. Neither of them is built around a chat box, but both can be classed as AI agents.

The difference is that consumer agents absorb more of the workflow, more of the transaction, and, importantly more of the spend.

This is how consumer AI escapes the 3% digital services box. It stops competing for subscription budget and starts competing for education spend, travel spend, healthcare spend, work spend, shopping spend, and financial services spend.

This is the solution to the consumer AI ARPU problem.

The obvious flaw in this vision is that in many cases, AI can’t do the vast majority of this. Yet.

That is where human-enabled AI agents come in.

You give the user an interface, often a messaging app (I’ve written about Building on WhatsApp here), but keep a human in the loop, on the thread.

The consumer might feel like they’re talking to a human. Or an AI. Or both. Maybe they don’t even know. It actually doesn’t matter that much.

But in reality, the agent is part model, part workflow software, part offshore operations team, part concierge. The AI handles obvious, repeatable, high-volume tasks, and humans handle the “last mile”: calling the hotel, chasing the insurance company, interpreting unclear instructions, double-checking policies, and escalating when something isn’t working.

This is how the best consumer agents will be built. And gradually, as AI progresses, the AI will take more and more of the workflow, either cutting cost and latency, or allowing the humans in the loop to do even more advanced work.

So human-enabled agents can be a solution to the ARPU problem, but it is not the only one.

Gradually, more and more pure consumer AI products will generate tremendous ARPU by themselves. In our portfolio, several companies are already generating ARPUs more than 10x higher than pre-AI consumer apps, without any human involvement. Simply by doing the work for consumers, or getting as close as possible to it.

So even as the world becomes fixated on enterprise AI, infra, and developer tools, we remain more excited than ever about AI for consumers, particularly those that can create value, and capture it.

Many of those will be consumer agents.