When Your AI Buys the Wrong Thing: Who Is Actually Responsible?

An assistant books the wrong size. An agent buys at a price that had already changed. An automated purchase goes through for an item that turns out to be a different model from the one the shopper described. None of these are hypothetical for much longer, and the question they raise is one no protocol specification answers: whose problem is it?

What follows is a description of the landscape as it stands, not legal advice. If you are building agentic checkout as a business, this is a conversation to have with a solicitor rather than a blog.

The framework was built for a human clicking Buy

UK consumer protection rests on a set of assumptions that agentic commerce quietly breaks. The Consumer Rights Act gives a buyer remedies when goods do not match their description. The Consumer Contracts Regulations give distance-selling cancellation rights and require that key information be given to the consumer before they are bound. The Digital Markets, Competition and Consumers Act tightened the rules on misleading actions and omissions, including in how prices are presented.

Every one of those protections assumes a consumer who was shown something and acted on it. When an agent is shown the information and the consumer is shown only the agent's summary, the chain of who-knew-what becomes genuinely unclear — and the interesting cases are the ones where each party behaved reasonably.

Four parties, four different exposures

The retailer

A merchant's traditional defence is that the price and description on their page were correct, and they were. If an agent read a stale copy from elsewhere and the shopper never saw the page, the merchant may still face a consumer who believes they were quoted a different price. Agentic protocols address part of this by having the merchant confirm price and availability at the point of transaction — which is exactly why that confirmation step exists and why it should not be treated as a formality.

The assistant

The party that stated the price to the human is in the most novel position. If it presented an unchecked number as a fact, it made a representation. The defensible posture is to be able to show what was known at the moment of the answer: the source, the timestamp, and whether the claim was verified or merely relayed. An assistant that cannot reconstruct that has no account of its own behaviour.

The data provider

Anyone supplying prices into the chain — feed aggregators, comparison services, verification tools — is now upstream of automated purchasing decisions rather than of human ones. The practical implication is that provenance and freshness stop being nice-to-have metadata and start being the record of what you actually asserted.

The shopper

Statutory rights against the retailer do not disappear because software placed the order. The harder question is agency: what exactly did the consumer authorise? "Buy me running shoes under £80" is a much broader instruction than "buy this specific pair", and the gap between those two is where most disputes will live.

The design decisions that reduce disputes

Most of this is avoidable, and the mitigations are unglamorous.

• Confirm at the merchant before charging. Price and availability should be re-read from the retailer at transaction time, never taken from the discovery step. Both major agentic protocols support this; the failure mode is treating it as optional.
• Keep the human in the loop above a threshold. A confirmation step on anything expensive costs a few seconds and removes the whole class of "I never agreed to that".
• Record what was shown. The assistant's answer, its sources, and the timestamps — kept long enough to resolve a dispute, and no longer than needed.
• Make the scope of authority explicit. An agent should know, and be able to state, the difference between browsing on someone's behalf and spending their money.
• Preserve cancellation rights end to end. Distance-selling rights are the shopper's, and a flow that makes them hard to exercise is a problem regardless of who wrote the software.

The cheapest dispute-resolution mechanism in agentic commerce is a re-read at the merchant before the card is charged.

Where the regulators are looking

Nothing here is settled, but the direction of travel in UK and EU consumer regulation over the last few years has been consistent: more scrutiny of how prices are presented, more emphasis on the information a consumer had before committing, and less patience with interfaces designed to obscure either.

It is not difficult to see how that applies to an assistant that states a price without qualifying it. A confidently-worded number with no source and no date is, at minimum, an odd fit for a regime built around informed consent.

The practical advice, for now

For shoppers: for anything expensive, let the agent find the product and then complete the purchase yourself, or at least look at the retailer's page before you confirm. Keep the assistant's answer if you may need it later.

For builders: assume that any price you state may one day need to be defended with evidence of what you knew and when you knew it. That is a data-model decision, and it is much cheaper to make before the first dispute than after.