Show three things, linked: the certificate, the evidence behind it, and the record of what the agent did while it was valid. An auditor will pick actions and ask which certificate allowed each, what testing supported it, who accepted that testing, and whether the limits were enforced. If the records answer in minutes, you are ready.
What will the auditor actually ask?
Auditors rarely start with the certificate. They start with an action, such as a refund the agent issued or a letter it sent, and work backwards. The questions follow a pattern:
- Was the agent permitted to do this, at this depth, on this date?
- What evidence supported that permission?
- Who decided the evidence was good enough?
- Was the permission enforced, or only written down?
- What happened when the agent failed or changed?
Certification only satisfies an auditor if each answer is in a record, not in someone's memory.
What are the steps, and who acts at each?
| Step | Who acts | What evidence it leaves |
|---|---|---|
| 1. Keep a register of every agent, capability and certificate | Agent owner | A list with capability, depth, issue and lapse dates |
| 2. Retain the test run behind each certificate | Testing team | Case set, version of brief, sources and model, scores per gate |
| 3. Record the floors set before the run | Process owner with risk | Floors with the date they were set |
| 4. Record acceptance by an independent reviewer | Named reviewer | Signed acceptance, conditions, any exceptions |
| 5. File every agent action with its verdict | The platform, automatically | Action, capability, depth, deciding limit, certificate reference |
| 6. Record failures, lapses and restorations | Agent owner | Failed run reports, drop in depth, rerun results, restoration date |
| 7. Monitor for drift and keep the results | Operations | Monitoring outputs and actions taken |
| 8. Answer the sample | Agent owner, with the auditor | The chain from action to certificate to evidence to reviewer |
Step 5 is the one that makes the rest usable. If each filed action carries a reference to the certificate in force, the auditor's sample can be traced in one step.
What does a good answer to one sampled action look like?
Take an action from the record: the agent issued a fee refund of 25 on a given day. The trace runs:
- The filed verdict shows "allowed", the capability "fee refunds within policy", depth 3, and the certificate reference.
- The certificate shows that capability at depth 3, issued six weeks before the action and not yet lapsed.
- The test run shows the case set, including past refund requests and requests that should have been refused, with scores per gate above the floors.
- The acceptance shows who reviewed it, when, and the condition that refunds above a set amount are held.
- The refund was below that amount, so the condition did not apply. A nearby filed action, a larger refund, shows "held" and the name of the person who approved it.
That chain answers every question in the list above.
Which mistakes make auditors uneasy?
- A model's or vendor's certificate offered as the agent's. The auditor will ask what testing was done on your cases. The distinction is explained in whether a certified model makes a certified agent.
- Floors set after the results. Dates matter; record them.
- No record of failures. An unbroken pass history looks less credible than one with a failure, a fix and a rerun.
- Certificates that did not lapse. If a certificate was relied on after its date, or after a change, the auditor will find it.
- Limits written in the agent's instructions, not enforced. Auditors ask for evidence of enforcement, such as held and refused actions.
How does monitoring fit in?
A certificate is evidence about the past. Auditors increasingly ask how you know behaviour has not moved since. Monitoring outputs, and the actions taken when monitoring showed a change, answer that; see how to detect drift in a certified AI agent.
What about regulatory documentation?
Some auditors will be checking against specific rules. In the EU, the AI Act sets documentation, record-keeping and human oversight duties for high-risk systems, and a deployer's auditor may ask for that pack alongside certification evidence; see EU AI Act documentation an auditor will ask for about your AI agents. Agents can also help assemble responses to auditor requests, covered in how to answer auditor queries with an AI agent.
Where Pryme fits
In Pryme, every verdict is filed in an append-only record with the capability, the depth and the bound that decided, so a sampled action leads straight to the Attest certificate in force and the run behind it. Lapses and failed runs are in the same record.
Questions people ask
How far back will an auditor look?
As far as the audit period and your retention rules allow. Keep certificates, test runs and the action record for at least as long as you keep the records of the work itself.
Can the auditor rerun the tests?
They may ask to. Keep the case set and the agent configuration so a rerun is possible, or explain why it is not.
What if the agent's certificate has lapsed since the action?
That is fine, if it was valid at the time. The question is always whether the certificate was in force when the action happened. The broader questions of whether an AI agent can be certified and what it means to certify an agent give context.