Who Pays When the Agent Is Wrong?

Meta recently released “Muse”, and they are going headfirst into agentic commerce. The pitch is that Muse will make real-world transactions on your behalf, through a partnership with Stripe, and “Connectors” (including initial partner connectors with Instacart and Expedia). Amazon, for its part, has blocked Muse agents from purchasing on its store, citing violations of its Terms of Service.

My position is fairly simple: we are not yet at the point where agents should be making unsupervised purchases, and particularly not large ones like airfare, hotels, and travel. Pushing ahead anyways, the people who are likely to end up holding the bag for failures are not going to be Meta (or other agentic developers): it will be the end consumers.

This Is Not a Technical Problem#

It’s tempting to frame this as a question of whether the agent is “good enough” yet. Will it pick the right flight? Will it book the right dates? That framing misses the point. A better agent will still make mistakes, and “who pays for the mistake” isn’t a question a better model answers. The limitation is legal and social, not technical.

When Muse is acting as an agent of a person, especially without the explicit knowledge or agreement of the other party, Muse is legally acting as that person, and that person is the one responsible for what the agent does. Not Muse, and not Meta. That’s the whole point of acting as someone’s agent.¹

A large purchase is a contract. If Muse books a flight, the person it booked it for has entered into that contract. There is no “but it was actually my computer!” escape clause. So what happens when it gets it wrong? Who is responsible when it books the wrong dates, or the wrong city, or a non-refundable fare you never would have picked yourself? Do refund policies change if the purchase was “agentic”? Does the airline care that it wasn’t really you who clicked the button?

I haven’t seen good answers to these questions yet. And “we’ll figure it out later” is a bad place to be when the thing in question is someone’s money.

Where This Does Work#

None of this means agents should never touch a purchase. It means the purchase needs something in place before the contract forms.

Supervised purchases are a fit when a person looks at the thing and says “yes, that one” before any money moves. At that point the agent did the searching, and the person made the decision, and the responsibility lines up with who actually decided.

Unsupervised purchases are a fit when the downside is small enough that the person would genuinely shrug it off. Re-ordering the same groceries is a very different risk than booking a week in another country.

What doesn’t work is the unsupervised, large, hard-to-reverse purchase, where the agent decides, and the person pays for the decision.

Amazon Is Only One Example#

Amazon is the easiest example to look at, because it pushed back. Amazon’s ToS claims are generally assumed to be about “data ownership” and “advertising margins”. Whatever the motive, there is a more immediate cost here: in this case, Amazon is the one that has to eat the loss when Muse makes a mistake.

Amazon’s return policy doesn’t exclude agentic purchases (yet). So every Muse purchase that turns into a return costs Amazon money, and Amazon has no financial recourse against the agent that caused it.

That’s only the initial state, though. I wouldn’t be too surprised to see Amazon stop accepting returns on agentic orders, especially from third-party agents that are already in violation of its ToS. If that happens, the loss doesn’t entirely disappear; it moves to the consumer. Either way, it never lands on Meta, or on the developer of whatever agent is at play.

It gets worse from there. Say a person “abuses” Amazon returns, and their account is suspended. If that return abuse was really Muse making bad purchases, does Meta take the hit for it? No. The account that gets suspended belongs to the end consumer. The agent acts, and the consumer carries the consequences. Muse/Meta just continue on, and (likely) claim that they are not party to whatever disagreement they caused.

Shopify is another company Muse is purchasing from, with the difference being that Shopify is fully bought in to the agentic commerce story, partnering with Muse directly. There is a strict difference between Shopify and Amazon, though. Amazon is the actual merchant². Shopify is, essentially, a software platform and marketplace. When a Muse order on a Shopify store goes wrong, the escalated returns and customer support issues don’t land on Shopify; they land on the merchant.

It’s the same pattern we’ve previously seen. The companies going “all in” on agentic commerce are the ones that can externalize the risk to their users.³

The Partners Are the Bigger Risk#

Amazon opted out. The larger financial risk to end users sits with the companies that opted in: Instacart as an initial partner, and even more so Expedia.⁴ With an integrated partner, the purchase doesn’t get blocked. It goes through.

A bad Instacart order is annoying. A bad Expedia order is a flight, or a hotel, or a whole trip, which is exactly the kind of large purchase this all started with. Will Expedia treat a Muse booking any differently from one you made yourself? If it doesn’t, the booking is yours, along with whatever it costs to undo.

We’ve Seen This Playbook Before#

Framing social and societal problems as technical problems is nothing new to Silicon Valley. Rideshare, Airbnb, DoorDash, and plenty of others have all run the same play: become embedded enough that you can’t be stopped, and externalize all of the risk.

We’re seeing that again with “Agents”. The difference this time is that it isn’t a scrappy startup trying to get big before the rules catch up. It’s being pushed by a ~$2T mega-corp.⁵

The questions above (who is liable, who absorbs the refund, whose account gets suspended) are going to get answered eventually. The concern is that they get answered after the fact, and one consumer at a time.

If there is any upside here, it is that these questions might be answered when the downside of agentic missteps is still monetary. We haven’t yet gotten to the point where agents are making practically-irreversible decisions on behalf of naive consumers. Hopefully the open questions are resolved long before we get to that point; unfortunately I do not believe that will be the case.

The agent makes the decision, and the consumer holds the liability. Until that changes, agents shouldn’t be making any large purchases, or binding their users to contracts or agreements.