Your Auditors Will Ask What Your AI Agents Did. Here Is the Answer.

The question never arrives during the work. It arrives months later, in a quiet room, with a date attached: on this Tuesday, your agent touched the customer data export. What exactly did it run, what did it see, and who told it to?

By then the chat is gone, or trimmed, or living in a tool your team half migrated away from. What survives is worse than nothing: a confident summary the agent wrote about its own work. Asking an agent to report on itself and calling the answer evidence is the one move no auditor accepts from a human employee. It should not pass for a machine either.

This is not a hypothetical backlog item anymore. A global bank is currently rolling AI agents out across its whole workforce, into work like code review, defensive security, and client advice. Every regulated firm watching that rollout is doing the same private math: thousands of agents doing real work means thousands of runs somebody will eventually have to account for. The accounting cannot be reconstructed after the fact. It has to be captured while the work runs.

What an answer actually looks like

An answer is a record made at execution time, by the system that ran the call, that a stranger can recompute without asking you for anything. On Zambo, every tool call through https://zambo.dev/mcp leaves an AER-1 receipt: the caller, the tool, the saved result, the timestamp, and the call's position in its job's sequence, sealed with a SHA-256 commitment over the canonical result bytes. The receipt lives at a public link. Anyone can open it, recompute the digest, and compare. Match means the saved result is the result that was saved. One changed character means the check fails loudly.

Here is a two-step sample from this morning, the kind of quiet dependency an audit eventually reaches: a review that relied on two reference values the agent recorded as it worked. Step one saved the BTC reference value, $74,664.62 at the moment it was recorded:

https://zambo.dev/run/f3010bb2-7f88-4a1b-9294-006cf1912484

Step two saved the ETH reference value, $2,111.86:

https://zambo.dev/run/d14d50c0-1ded-4f0b-8872-53fb19d95b2c

Both steps were then sealed into one workflow record, the chain an auditor actually receives:

https://zambo.dev/workflow/49c5adee-9045-4ffa-be2a-cf70b97e6f0f

Open any of the three. The saved value sits on each receipt page under its digest, and the workflow page shows the sequence with its root commitment. Prices move. The saved values do not, and the digest is how anyone proves it. No login, no export request, no trusting my description of what you are looking at. That is the entire standard: the record answers the question by itself.

Why the usual substitutes fail

Chat transcripts fail because they are editable, partial, and written for conversation, not evidence. Screenshots fail because they prove a screen looked a certain way once. Agent self-reports fail for the reason above: the witness is the worker. Internal logs come closest, and they fail on the two points auditors press hardest. The log lives on your infrastructure, under your control, so its integrity is your word. And a log line records what the system says happened, with no commitment a third party can recompute to confirm the entry was not edited in the months since.

The receipt inverts both points. The commitment is computed over the result bytes at the moment the call returns, and the record is public, so editing it later breaks the math in front of anyone who checks. As of this morning the public stream holds 125,177 of these receipts, and the format is not one vendor's promise: AER-1 is an open Internet-Draft (draft-zambo-aer1, closed at revision -26) with a public conformance corpus of 143 vectors and 146 checks that five independent implementations pass. If Zambo disappeared tomorrow, the records your auditors hold would still recompute.

The honest boundary

A receipt proves the saved result was not changed after it was saved. It does not prove the result was right, and it does not make anyone compliant with anything. What it gives a compliance team is the thing they actually cannot manufacture later: a fixed, checkable account of what ran, in what order, producing what. Whether that account satisfies a given rule is the auditor's call, made on evidence instead of on your summary of evidence.

The firms getting this right are not buying a dashboard. They are making one decision early: work runs through an endpoint that receipts every call, so the record exists before anyone knows which runs will be questioned. Twenty free calls plus one wrap-up land in every account daily, pickup is always free, and a $0.99 Day Pass makes a full day of runs unlimited, so the habit costs nothing to start and pennies to keep. The expensive version of this decision is making it after the first audit letter arrives, when the runs being asked about are already unrecoverable.

Your AI said it did the work. When the auditor asks, here is the receipt.