Airbnb CEO Says AI Agents Like Muse and Instinct Barely Bring In Bookings, And That Tells Us a Lot About the Future of AI

Every few months, a new wave of AI agents arrives with the same promise: soon, software will do your shopping, your planning, and your booking for you. You will simply describe what you want, and an agent will go out into the world, compare options, and complete the transaction.

That story just met a hard reality check. Airbnb's CEO says AI agents like Muse and Instinct are barely bringing in bookings, and he is blunt about what that means for his business: "I don't see them as a threat."

It is a short statement with a long tail. It does not mean AI agents are useless. It means something narrower and more useful: there is a wide gap between agents that can find things and agents that can close them. Understanding that gap is the key to knowing where AI is actually heading next.

The Claim, Plainly Stated

Strip away the hype and the message from Airbnb's corner office is simple. Agents like Muse and Instinct are out there. People are using them. But when you look at the number that matters, completed bookings, the agents show up as a rounding error rather than a revenue line.

That is worth pausing on. Airbnb is not a company that struggles with distribution. It sits on a massive base of demand, a familiar brand, and a product people already know how to use. If AI agents were truly becoming the front door to travel purchasing, this is exactly the kind of business where you would expect to see it first. Instead, the CEO's read is that the agents are noise, not threat.

The confident phrasing is itself the story. A CEO does not usually dismiss a technology this openly unless the internal numbers are clear enough to make the dismissal safe.

Why "Barely Any Bookings" Is Not the Same as "Not Working"

Here is where a lot of commentary gets this wrong. Low booking volume does not prove that agents are broken. It proves that agents are stuck at a specific stage of the funnel.

Think about how a trip actually gets booked. It usually happens in stages:

• Inspiration, someone starts thinking about a destination.
• Research, they compare places, prices, and dates.
• Decision, they pick one option and commit.
• Transaction, they pay, confirm, and handle changes if things go wrong.

AI agents are genuinely good at the first two stages. They can summarize, compare, and narrow options quickly. They are far weaker at the last two. The decision stage involves taste, trust, and the fear of getting it wrong. The transaction stage involves payments, identity, cancellation rules, and someone to call when the plan falls apart.

So when a CEO says agents are "barely bringing in bookings," the most likely reading is that agents are influencing the journey without owning the checkout. They are a research assistant, not a travel agent with a credit card.

Discovery Is Cheap. Conversion Is Expensive.

This is the single most important lesson the AI industry keeps relearning. Building an agent that can browse and recommend is now almost trivial. Building an agent that can reliably complete a purchase on someone else's behalf is a much harder problem, and most of the difficulty is not in the AI at all.

To close a booking, an agent needs more than language skills. It needs:

• Permission. A payment method it is authorized to use, with limits and safeguards.
• Accountability. A clear answer to "who pays if the agent gets it wrong?"
• Integration. Real access to inventory, pricing, and availability, not a scraped web page.
• Reversibility. The ability to cancel, change, or fix a mistake without a human fighting through a support queue.

Every one of those is an organizational and legal problem before it is a technical one. That is why agent-driven commerce is arriving slowly even as agent-driven research arrives fast. The demo is easy. The transaction is the moat.

Why "Not a Threat" Is a Strategic Signal, Not Just a Boast

There is a second layer to the CEO's comment. When an incumbent says a new technology is not a threat, that is rarely just an observation. It is a positioning statement, and it usually comes with an implicit plan.

Three things tend to follow from a statement like this:

1. Build it yourself, on your own terms

If outside agents are not converting, the natural move is to build assistant-style experiences inside your own product, where you control the data, the checkout, and the customer relationship. You get the benefits of the technology without handing the front door to someone else.

2. Keep the direct relationship

Platforms have spent a decade and enormous sums building direct traffic. Letting third-party agents sit between the brand and the buyer would weaken pricing power and reduce the value of that relationship. Being lukewarm on external agents is a way of protecting the channel that already works.

3. Wait for the standards to settle

Agentic payments, identity verification, and liability rules are all still forming. There is little upside in rebuilding your tech stack around standards that could change twice before they stabilize. Watching and testing is cheaper than committing.

None of that is a rejection of AI. It is a bet on sequencing, the belief that the agent era arrives on the incumbent's timetable, not the startup's.

The Trust Gap Is the Real Bottleneck

Consider what a traveler is actually risking. A bad restaurant recommendation costs nothing. A bad booking costs money, time, and possibly a ruined trip. People are remarkably tolerant of AI that suggests and remarkably intolerant of AI that commits.

This shows up in behavior everywhere, not just travel. People happily use AI to draft, plan, and compare. They hesitate when the AI wants to press the buy button. The hesitation is rational. Until an agent can be trusted with the consequences of a mistake, the human stays in the loop at the final step, and the final step is where the money is.

That is why "barely any bookings" is such a precise phrase. It does not describe a technology that failed. It describes a technology that has not yet earned the right to finish the job.

What This Means for the Future of AI

If the Airbnb CEO's read holds across the industry, the next phase of AI will look less like a world of autonomous agents running errands and more like a world of assisted decisions with narrow, well-guarded moments of autonomy.

Expect the progress to show up in a specific order:

• First: research and comparison. Already working, and getting better fast.
• Second: structured, low-risk transactions. Repeat purchases, subscriptions, and rebookings where the choice is already made and the risk is small.
• Third: high-consideration, high-cost purchases. Travel, housing, insurance, big-ticket retail. This is where agents will arrive last, and where trust infrastructure will decide the winners.

The companies that win will not be the ones with the most impressive demos. They will be the ones that solve the unglamorous parts: clean inventory data, clear liability rules, fast refunds, and a human handoff that actually works.

Practical Playbook: What Businesses Should Do Now

The right response to this story is neither panic nor complacency. It is measurement. Here is a practical set of moves for teams watching the agent wave.

Measure agent influence, not just agent revenue

An agent that never completes a booking can still shape which brand a traveler considers. Track agent-referred traffic separately from agent-completed transactions, and watch the whole funnel. Judging agents only by conversions today will make you underrate them tomorrow.

Make your inventory machine-readable

Agents need structured data, availability, pricing rules, policies, and constraints, not a marketing page optimized for human eyes. Clean, consistent, well-documented data is the price of admission for any agent to recommend you accurately.

Protect the direct channel while you experiment

Run agent pilots on the edges of your business, not the core. Give agents a test lane, measure outcomes, and keep your owned channels strong. The companies that lose in a platform shift are usually the ones that handed away distribution before they had proof it would pay off.

Design for the "last mile of trust"

Assume that for the near future, a human will want to confirm anything expensive. Build flows where an agent can prepare a decision and a human can approve it in one tap. That hybrid model is likely to be the dominant pattern for years, not a temporary workaround.

Watch the plumbing, not the demos

Agentic commerce will break through when payments, identity, and dispute resolution become boring and standardized. Track those developments more closely than you track product launches. The demos will always look impressive; the plumbing is what actually unlocks volume.

The Bigger Picture

It is tempting to read this story as a verdict: AI agents are overhyped, the bubble is deflating, move along. That reading misses the point.

The more accurate takeaway is that the AI industry has entered the phase where talk gets replaced by bookkeeping. It is no longer enough for an agent to seem smart. It has to complete a transaction, absorb a mistake, and survive a refund request. Those are boring tests, and they are exactly the tests that separate durable products from viral demos.

A CEO saying "I don't see them as a threat" is not the end of the agent story. It is a marker of where the story currently stands: agents have won the conversation and lost the checkout. Whoever closes that gap, with trust, integration, and accountability rather than better copywriting, will own the next chapter.

For now, the most honest summary of the agent economy is this. The research layer is here. The transaction layer is coming. And the distance between those two things is measured in trust, not in tokens.