Chatbot Is Not the Interface

The claim

Brian Chesky's position is that the text chatbot is the wrong interface for travel and e-commerce, and that the future, while agentic, will be rich and visual rather than text-forward. "The future are not apps. The future are agents, but I don't think they're going to be text-forward. They're going to be really rich user interface."¹

The failure modes of a chatbot for commerce

Chesky agrees with the broad direction that apps give way to agents, but breaks sharply on what the agent should look like, enumerating specific ways today's chatbot fails commerce. It is text-based, when travel and commerce are inherently visual and photos are treated as an afterthought. It offers no direct manipulation: a user must type every prompt, which is fine for conversation but rules out adding filters, clicking around, or tweaking a result, the things that make shopping efficient. Comparison is hard, since much of e-commerce and travel is comparison shopping, and with thousands of options an AI has to infer exactly what a user wants in order to surface one or two results, when users usually want to see more choices rather than get lost in a purely text back-and-forth. And it is single-player, when most trips and purchases are planned by two people; on Airbnb specifically, 85 percent of guests send a message to someone else during the process.

When an interviewer described an idealized flow, an agent given a location, candidate hotels, party size, and dates that pulls listings, pricing, and photos, runs a trade-off analysis against the wider internet, and then lets two people share and review the chat together, Chesky's reply names the crux directly: "Yes, that's the future. And that's not a chatbot, what you described." The capability can technically run on top of a chatbot, but the right surface, in his view, is much more immersive: rich maps, a sense of the neighborhood, a system that talks back, photo comparison across options.

He grounds the claim historically: ChatGPT launched third-party apps and then shut them down, a retreat he attributes to needing a richer SDK than a chat thread can offer. The App Store worked precisely because every app had a unique, rich interface, not because everything ran through one shared conversational surface. He argues the unlock is now arriving on the model side, with image and video generation improving quickly enough to make going beyond text possible, and predicts real-time AI-generated interfaces within one to two years. His summary of where things stand: modern AI today is text, photos, and buttons, and the next paradigm is generative, visual, and immersive.

Why it matters

This is a specific, falsifiable bet against the dominant framing that everything in consumer AI collapses into a chat box. If Chesky is right, value in consumer AI accrues to whoever builds the rich agentic surface rather than to the raw chat model, which reopens a lane that pure AI labs, shipping text chatbots with no product or design instinct, have left open. It also rehabilitates design and front-end craft as a frontier skill rather than a casualty of AI, since a generative visual interface still needs designers who can build. And it puts a specific bound on the broader claim that agents replace apps: agents win, in this account, but the winning agent looks like an immersive, directly manipulable, multiplayer visual tool, closer to a regenerated app than to a message thread.

Independent corroboration from fintech

Eric Glyman reached a similar conclusion earlier and in production, from a different industry. Describing 2023 on the MAD Podcast, he recalled: "there was this weird, funny phase where people wanted to chat with everything and just put a chatbot everywhere, and we never met a single person who said, I just wish I could chat with my bank account. It felt really off to us that that was going to be the way it was going to go."² Ramp's alternative question was whether there are data-oriented products where AI is needed to produce a materially better result, which led to AI embedded directly in the existing operational surface, price, expense, and accounting intelligence, rather than a chat layer bolted on top. Notably, Ramp's own agentic features are not chat-first either: a screen agent acts on the product's own interface, clicking tabs and filling inputs, rather than replacing it with a thread, consistent with Chesky's prediction and coming from a company that had already rejected the chat box on its own terms.

Tensions and open questions

The chatbot-is-not-the-interface claim agrees with the view that apps give way to agents but disagrees about the endpoint: some argue most apps dissolve into conversational agents, while Chesky argues the winning agent must be a rich visual surface, closer to a regenerated app than a chat thread. It also remains a prediction rather than a shipped product; Chesky himself concedes he is not sure Airbnb will be the company to prove it out, and no reference implementation yet exists. There is also counter-evidence within his own account: chatbots currently send referral traffic to travel companies that converts at a higher rate than Google search, meaning today they are additive to travel commerce rather than disintermediating it, and the rich-visual disruption he describes is still hypothetical.