Every AI agent should leave a receipt.

What is an AI agent receipt?

A receipt is the record an agent leaves after a run, written for the person who has to stand behind the result. It names the approved context, the rules in force, what changed, who decided what, and what is still open. If it takes longer to read than the work takes to check, it is a log, not a receipt.

Logs matter, and most agent tools keep them. But a system can record every click and still leave you unable to answer the question that comes up on Monday: was this allowed, and is it still the brand we approved?

Don't always-on agents already keep an activity log?

Always-on agents keep the first half of a receipt. OpenAI says dots start with built-in rules for when to act alone and when to ask, that Custom Rules let you allow, require approval for, or block specific actions, and that an Activity View lets you follow the work, including background work. That covers what the agent did. It cannot cover whether the work matches your brand, because the agent only knows the version of your positioning someone gave it.

OpenAI is plain about the limit on its own launch page: “Dots can still make mistakes, so always review consequential work.” A receipt is what turns that review from an afternoon of retracing steps into a few minutes of checking five lines.

Six days later, OpenAI gave qualifying ChatGPT advertisers Negative Phrases, for “narrow placement requirements specific to a brand’s own policies.” Different product, same lesson. The platform supplies the control. Your company has to supply the rules, in writing, before the system runs.

What should a brand agent's receipt show?

Five lines, each one answerable by a person who was not watching. Longer lists exist for security teams. For a marketing team, these five decide whether the work can ship.

The list is ours, and it lines up with what researchers and security groups are asking for. The Shared AI Findings Exchange proposal, a proposed initiative for sharing AI incidents and near misses, asks members to preserve evidence including tool calls, the permissions available during the run, human approval and intervention events, and the files created or modified. The paper Agent Safety Should Be a Runtime Contract argues that an agent should show “hard evidence” such as file diffs and log captures before a task counts as done. One is a proposal and the other a position paper. Neither is an adopted standard.

What does a receipt look like in practice?

This article has one. Before our blog publisher releases a post, a reviewer who did not write it has to return a go, and that go is sealed to the exact article and the exact desktop and phone screenshots it reviewed. Change a word after the review and the publisher refuses to release it.

We built it that way after a mistake we kept. On September 17, one of our articles went live with “Cover pending” where its cover should have been, and a reviewer had passed it. We changed the review and the publisher that day. Now the publisher refuses any article without its approved cover, and the rule sits in the file every later run reads.

A receipt will not stop every mistake. It makes each one traceable to the check that missed it, so the fix lands on the check and not only on the post.

Where should the receipt point?

To a source your company owns and approves. A receipt that says the agent used “the brand” proves nothing. One that names the approved file, its version, and who last changed it can be checked by anyone on the team.

That is the job of a customer-owned Brand Brain: your approved positioning, voice, claims, and boundaries in plain, versioned files your company keeps, changed only when a person approves the change. Agents read it before they work. Their receipts point back to it after.

If you are writing the rules an agent follows, read how to test a brand rule without blocking the work. If you are still deciding where AI fits in your brand plan, start with what an AI brand strategy is for.

What can I do this week?

Pick one agent or automation that already touches customer-facing work. Ask it, or the person who runs it, for the five lines from its last run. Every line nobody can answer is the next thing to fix, before you give that agent more access.

If you are about to switch on an always-on agent, write two lists first: what it may never say about your company, and which claims it may make only with proof attached. Those lists are what its receipt gets checked against.