Agents Need Receipts: The Missing Citation Layer in AI Commerce

A claim without a citation is an assertion. Academic writing worked this out several centuries ago, journalism arrived at it more recently, and software has a version of it in the audit log. Commerce data, at the exact moment it started being consumed by autonomous systems, has almost nothing.

When an AI assistant tells a shopper that something costs £249, that sentence carries no source, no time, and no way for anyone downstream — the shopper, a second assistant, a regulator, a journalist — to establish what was actually checked. The number is repeated onward as if it were a fact and there is no thread back to whatever made it one.

Why this gets worse rather than better

Claims used to end at a human. Someone read a comparison site, formed a view, and either bought or did not. If the price was wrong they discovered it at the checkout and corrected silently.

In an agentic chain a claim keeps moving. One service produces a price, an assistant relays it, another agent acts on the relay, and a purchase happens several hops from the original reading. Every hop adds confidence and removes context, because natural language is very good at carrying the number and very bad at carrying the caveat.

Each retelling of an unsourced price makes it sound more certain and leaves it no more true.

What a commerce evidence receipt is

The idea is deliberately modest. When a service answers a question about a price, it also emits a small structured object describing that answer: what was asked, what was found, when the underlying observation was made, how the product was identified, and under which published rules the answer was produced.

Critically, a receipt adds no new information. It does not make an answer more true. It serialises what was already returned, so that the answer can be referred to later by something other than a paraphrase.

The properties that make one useful

• Stable identity. Two identical answers should produce the same receipt id, and a materially different answer should produce a different one. That is what makes an id a citation rather than a serial number.
• Bound to the content. The id should be derived from the answer itself, so an edited claim cannot keep the original's reference. A receipt that floats free of what it describes is decoration.
• A named standard and version. "Verified" means nothing without the rules it was verified against, and those rules must be citable and versioned so a reader can see which edition applied.
• The observation time, distinct from the answer time. These are different facts and conflating them is how a month-old reading gets presented as a fresh one.
• Honest about signing. If a receipt is cryptographically signed, say so and publish what checks it. If it is not, say that too — an unsigned receipt still binds an id to a payload, which is useful, and claiming more than that is worse than claiming less.

The last point deserves emphasis, because it is where this idea most easily turns into theatre. A signature proves that a particular party issued a particular set of fields. It says nothing whatever about whether those fields are correct. A signed receipt over a wrong price is a well-attested wrong price.

What it does not solve

Being clear about the limits is most of what makes the idea credible.

1. It does not make a price accurate. Evidence quality is a separate problem, addressed by what you observe and how, not by how you cite it.
2. It does not bind the retailer. The merchant sets the price at the checkout, and a receipt from a third party is not an offer.
3. It does not create trust in the issuer. It makes the issuer checkable, which is a different and more useful thing.

What it does is convert an unfalsifiable claim into a falsifiable one. "It cost £249" cannot be shown wrong. "This service, under version X of these published rules, reported £249 at this retailer from an observation made at this time" can be — and something that can be shown wrong is the only kind of claim worth trusting.

Who this is for

For assistants, a receipt is the difference between relaying a number and citing a check. An assistant that can say where its price came from is more useful and considerably more defensible than one that cannot.

For retailers, it is a way to see what is being said about your prices, and to challenge it with specifics rather than with a complaint. A dispute about an undated number goes nowhere; a dispute about a dated, sourced, referenced claim resolves.

For anyone building commerce data, it is a discipline more than a feature. You cannot issue a receipt for something you did not record, so the act of trying to emit one immediately shows you which of your fields are evidence and which are habit.