Blog
Essay · November 12, 2026 · 4 min read

Why almost nobody lets an agent buy for them yet

Airbnb, DoorDash and Booking all said this month that agents are barely ordering anything. I think the reasons are plain, and I think they can be fixed.

I build Calmar, which lets AI agents buy things. So this was an uncomfortable week to read the news.

Brian Chesky said Airbnb is seeing "very, very low numbers" of bookings from Meta's AI agents. Tony Xu said DoorDash has had "maybe several hundred orders in total" from personal AI assistants. Glenn Fogel of Booking said adoption "may be much slower than is anticipated." Semafor put the three together under a headline that said agentic commerce is slow to catch on.

A year ago the pitch was billions of agents filling carts. Several hundred orders at a company the size of DoorDash is a rounding error.

I have spent the week asking myself whether the idea is wrong. I have come out believing that three specific things are broken, and that each one has a fix.

The agents still buy the wrong thing

The same week, Visa published the first results of a benchmark for AI shopping. It gave 22 models from 12 labs 15 ordinary tasks, like choosing a TV or a laptop for a person who had described what they need. Across all of them the models passed 57.2% of the time. The best passed 82.2%.

The number that stayed with me is a different one. In more than half of the failed runs, the model bought something anyway. Every model had three tries at every task, and all three tries succeeded only 34.2% of the time.

Picture a friend who shops for you. Four times out of five they come back with the right laptop. The fifth time they come back with the wrong one and say nothing. You would stop sending that friend to the store. People who refuse to hand an agent their card are behaving sensibly.

Nobody gave the agent a budget

When I ask a person to buy something for me, I tell them how much they can spend. Most agents today either hold a card with no limit or hold no card at all. The first is frightening. The second hands you a link and calls it shopping.

I know the second one well. This week I searched for a flight through our own product and got back one cached fare and a link to another site. A lot of this industry is in that place right now, and on flights so are we. A link is a suggestion. The person still does all the work.

Here is what I think has to exist before anyone relaxes. Every agent has a name, a limit for one order and a limit for one day, and an order over the limit is refused. We added that to Calmar this month. It is dull work, and I believe it matters more than a smarter model.

Stores have no reason to say yes

Semafor made a point near the end of its piece that I think is the real story. Retailers such as Amazon block shopping agents, and slow adoption makes that choice look smart, because an agent cannot be monetized the way a human shopper can.

My reading of why: an agent does not see ads. It does not add a second item on impulse. It skips the page the store spent years tuning. From the store's side, an agent is a customer who costs more to serve and is worth less.

So the store blocks the agent. The agent falls back to links. The person has a worse time and fewer people try again. Every side is acting reasonably, and the result is several hundred orders.

The person who built the agent is stuck in the same way. They pay for every model call and earn nothing when the order goes through. Nobody in that chain is paid to make the purchase go well.

What I would change

Pay for the sale. A store already pays to be found, through ads and marketplaces. It can pay a fee when an agent brings it a finished order, and part of that fee can go to whoever built the agent. Then the store wants agents to come, and the builder wants the order to be right, because a refund takes the money back.

One rule keeps this honest: the fee cannot decide who wins. On Calmar the best deal for the buyer wins, and the fee only settles a tie between two equal offers. If the highest bidder simply won, people would be right to distrust the whole thing.

Let the agent say no. Visa's benchmark expects a model to buy nothing when nothing fits what the person asked for. I would go further and show that to the person as a result. I looked, nothing met what you asked for, here is the closest one and where it falls short. An agent that can refuse is an agent you can leave alone.

Start with the dull purchases. A charger. Printer paper. The same dog food as last month. Nobody wants to browse for those, and a mistake costs twenty dollars. Flights and hotels make the best demos, and they are also where one error ruins a trip. I would earn trust on the charger first.

Where I think this goes

I expect the numbers to stay small through this holiday season. I would be surprised if agents placed more than a sliver of orders before next year, and I would rather say so than pretend.

Buying anything online once felt reckless too. People typed a card number into a website only after there were refunds, a cap on what a thief could cost them, and stores that wanted the order. Those same three things are missing for agents today, and all three can be built.

That is the work we are doing at Calmar. If you build agents or run a store and you see this differently, I would like to hear where I am wrong.

Let your agent buy, and get paid for it

Create a free account
Keep reading