When your AI agent books a hotel, who pays it?

38 of the 84 agents in this directory are free to start. 1 tells you what it earns from the things it recommends.

That gap is small now because these products mostly answer questions. It gets larger every time one of them books something.

What was actually described

On 28 September 2026 Instinct's founder gave his first long interview and set out how a free product is meant to pay for itself. The shape is a percentage of what the agent spends, charged to the other side. He compares it to platforms that already work this way. He puts Shopify at around two and a half to three percent, Amazon at "upwards of 10%", and Apple's in-app purchases at thirty. Of boutique hotels he says some are "offering to pay up to upwards of 30% for every transaction" delivered to them. He is careful to add that he is not claiming Instinct will be at thirty.

He also says what the product is doing at that scale: over a billion dollars a year in transaction volume, half of it travel, on an invite-only user base.

Two things are worth separating. The take rate is described in detail. Staying free is not promised: asked directly, he says he will not commit to free for a lifetime and calls it a personal goal.

The distinction he draws

The interview raises the alignment question head on, and his answer is the interesting part. He names the advertising model as the thing he does not want to build. When the host puts the old line to him, that if you are not paying you are the product, he agrees. A world where a system smarter than the user can push them toward a purchase they do not want is, he says, dangerous.

The distinction is real. An advertisement interrupts you and competes for attention you were spending on something else. A take rate is charged on a thing you already asked for.

What it does not settle is the picking. In his own travel example the user says they need to be in New York tonight, and everything after that is chosen by the agent: the airline, the seat, the hotel, the car. If hotels pay different percentages, and the published ranges say they do, then the choice has a price attached to it that the person reading the answer cannot see. He addresses this by saying Instinct should act on what the user wants and should follow higher level objectives rather than completing tasks blindly. That is an intention, and this site records intentions as claims rather than as controls.

The first one arrived the same day

Hours after that interview, Instinct announced a partnership with Shopify: its agent now searches the merchant network for live prices, sizes and availability, and you check out with Shop Pay using details already saved to your Shop account. The company says 35 percent of its users already use it for shopping.

The announcement lists what the user gets. Discovery across brands, price-drop alerts, delivery updates, return-window reminders, automatic discounts. It does not say what Instinct is paid, and there is no reason it would: no rule currently requires it and no competitor does it either.

The gap this piece is about was demonstrated within a day of being described. Shopify is also the number he named as the low end of the range, around two and a half to three percent. It now sits inside Muse as well, which has carried it since launch day.

Somebody outside worked to the same answer that week. A researcher-investor at Leonis Capital modelled the unit economics from the other end and published the chart. An actively used personal agent, on her estimate, costs more each month to run than people have historically paid for software. That is a poor subscription.

Bring the serving cost down and the operator gets a choice: keep charging and enjoy the margin, or price at zero and earn a small percentage of the purchases the agent routes. Her line for the end state is that the valuable asset is the purchasing decisions flowing through it.

Her figures are estimates of other companies' costs and none of them is a field here. What is worth noting is that two people reasoning from opposite directions, one running the company and one modelling it from outside, landed on the same business.

The rules were written for a page you can look at

Paid placement is not a new problem and it is not unregulated. Under the United States endorsement rules a material connection has to be disclosed wherever it would change how much weight a reader gives a recommendation and the reader would not expect it. A commission or an affiliate share is named as exactly that kind of connection.

The catch is the format those rules imagine. A labelled advertisement sits next to the organic result, so the reader sees both and knows which is which. Guidance on the newer interfaces adds a condition that is hard to meet in a text thread. A disclosure has to travel with the claim. It cannot sit in a footer, or in something you agreed to during onboarding three weeks earlier.

An agent texting you one hotel has no next to. There is no second column, no "sponsored" chip above a list, and in the voice case there is no screen at all. Regulators have not published guidance for agents that transact, and a bill drafted to cover autonomous agents is still a draft. So for now the question of whether the hotel in the reply paid to be there is answered by whoever built the agent, or not answered.

What the 84 listings publish today

1 of them says it earns from what it suggests. It is Zapia, whose policy describes affiliate links, keeps records of commissions, and gives you the right to opt out of affiliate conversion tracking and to ask what is shared. Whatever else that is, it is the arrangement written down, with a switch.

Everything else in this market is quieter. 25 listings publish no price anywhere. 38 are free to start and most of those say nothing at all about where the money comes from instead. That is not evidence that any of them takes a commission. It is the reason you cannot tell.

What to watch

Whether a second listing publishes how it earns from what it recommends. One is a company being unusually forthcoming; two is the beginning of a norm, and the norm is the thing that decides whether you ever get told.

And whether the first agent that books on your behalf puts the disclosure in the message. Not in the terms, not in onboarding, in the reply that names the hotel.