Auditors examining AI agents ask a predictable set of questions about people: who touches which agent, what they must know, how you know they know it, and whether only qualified people approved held actions. A training evidence pack answers all four in one bundle. The template below has ten sections, each with the evidence it should contain.
What is the pack for?
It saves the audit from turning into a fortnight of emails. Without a pack, an auditor asks for the training policy, then the list of approvers, then their certificates, then the access logs, then a sample of approvals, and each request arrives in a different format from a different person. With a pack, the auditor starts from an index and tests.
It also forces you to find your own gaps first. Assembling section 9, approvals traced to the approver's credential on the date, is where most organisations discover that someone approved held payments for a fortnight on a lapsed certificate. The argument for treating these records with the same care as financial records is made in training records are evidence; treat them that way.
When should you build it?
Before any audit that covers AI agents: an internal audit, an external attestation such as SOC 2, a certification audit against ISO/IEC 42001, or a supervisor's review. Build it once, then refresh it each quarter so it is never more than three months stale. For preparation steps, see how to prepare training evidence for an AI audit. If the audit is a SOC 2 examination, SOC 2 and evidence of training for teams running AI agents covers what that report looks at.
How do you assemble it?
- Put every document in the pack with a version and date. An undated policy is not evidence.
- Use the same person identifiers across sections, so the auditor can follow one person from the role map to their approvals.
- Include failures and remediation. A pack that shows only passes looks curated.
- Point to the system of record for anything too large to include, and say how the auditor can query it.
The evidence pack template
| No. | Section | What it contains | Evidence format |
|---|---|---|---|
| 1 | Index and owner | List of sections, owner, date assembled | One page |
| 2 | Scope | Agents in scope, their capabilities and certified depth | Table from the workspace |
| 3 | Policy | AI literacy policy and AI agent policy, with versions | Signed documents |
| 4 | Role requirements | Each role, its contact with agents, the tier or credential it requires | Role map |
| 5 | People | Everyone in scope: role, required credential, held credential, number, expiry | Credential register extract |
| 6 | Checks at source | Date and result of the last check of each credential with the issuer | Register columns plus screenshots or exports for a sample |
| 7 | Assessment evidence | Exam results as counts, practicum logs, sign-offs | Exports and signed logs |
| 8 | Rights control | How approval rights are granted, suspended on lapse and restored | Procedure plus access log extract |
| 9 | Approvals traced | A sample of held actions approved in the period, each traced to the approver's credential status on that date | Sample table |
| 10 | Exceptions and remediation | Lapses, late renewals, unqualified approvals found, and what was done | Exception log |
A worked example
Take a sample payments company preparing for a SOC 2 examination that includes its refunds and payment-repair agents. Values are illustrative.
Section 2 lists two agents: payment repair at depth 3, refund release at depth 4 held for a named approver up to 2,000. Section 4 shows three roles touching them: 12 reviewers who release held refunds, 3 team leads who approve above 2,000, and 2 builders who maintain the agents' briefs. Section 5 shows all 12 reviewers and 3 leads holding a professional operator certificate, and both builders holding a builder-level certificate.
Section 9 samples 60 approvals across the period. Fifty-nine trace cleanly. One refund of 1,600 was approved by a reviewer whose certificate had lapsed four days earlier. Section 10 records it: found during pack assembly, refund re-reviewed and upheld by a team lead, rights now suspended automatically on lapse through the procedure in section 8.
The auditor can now test the claim in section 8 directly, by checking the access log for the next lapse after the change. The pack has turned a finding into evidence that the control was fixed. The difference between that and a certificate on file is the subject of what evidence of competence is; agents have their own equivalent in the certification evidence pack.
How does this pack relate to the agent's own evidence?
It is the people half. Agents carry their own certificates, which lapse after 90 days, and the evidence for those comes from certification runs. An auditor will usually want both and will test where they meet: a held action approved by a qualified person, on an agent certified for that depth, on that date. Pryme Intelligence Academy certificates can be checked at source by number or QR code, which makes section 6 a lookup rather than a request. More in the guide to AI literacy and certification for people who work with agents.
Questions people ask
How large should the approval sample in section 9 be?
Large enough to cover every approver at least once and every month of the period. For a team of around 15 approvers over a quarter, 50 to 60 items is typical.
Should failed exam attempts go in the pack?
Yes, in section 7. They show the assessment is real, and an auditor may ask how many people failed and what happened next.
Who should sign the pack off?
The executive who answers for the AI deployment, after the register owner and the head of the reviewing team have confirmed their sections.