Most training records are kept like a guest list: who turned up, roughly when, in a spreadsheet anyone can edit. Once people approve what AI agents prepare, those records become evidence that the approver was competent on the day. Evidence must be attributable, time-stamped, unaltered and retrievable for a given date. Keep training records to that standard.
Why have training records become evidence?
In a governed deployment, every verdict is filed in an append-only record. When an agent's action is held and a person approves it, the record shows who approved, when, and on what basis. That record is designed to be reconstructed: anyone examining a decision months later can see exactly what happened.
The obvious next question is whether the person who approved was qualified to. That question can only be answered from training records. If they are thin, editable or vague about dates, the verdict record leads to a dead end. You know who approved; you cannot show they were competent.
That is the shift. Training records used to support HR processes. They now support the control environment, and they will be read by people whose job is to test evidence: internal audit, external auditors, supervisors, and sometimes lawyers.
What makes a training record good evidence?
The standards are the same as for any other evidence in a control environment.
| Property | What it means for a training record | What usually goes wrong |
|---|---|---|
| Attributable | It identifies the person and the issuer unambiguously | Records keyed by a name that several people share |
| Time-stamped | It shows when competence was established and when it expires | A "completed" flag with no dates |
| Unaltered | Changes are added, not overwritten | Someone edits an old row to fix a typo, and the history is lost |
| Specific | It names what the person was assessed on: role, level, product version | "AI training: done" |
| Retrievable as at a date | You can show status on the day of any decision | Only current status is visible |
| Checkable | An outsider can confirm it without trusting your copy | A PDF on a shared drive |
The guide to what records a company should keep of AI agent training covers the content in more detail. The point here is the standard: if a record would not survive the questions an auditor asks of a payment authorisation, it will not survive the questions they ask of a training record either.
What is the difference between attendance and evidence of competence?
Attendance shows that someone was present. Evidence of competence shows that they can do the work. For approvers, that means an assessment of their judgement, ideally on observed real decisions, with a result and a standard it was measured against.
Many organisations hold plenty of attendance records and almost no evidence of competence. When the question comes, they produce a completion report and hope it is enough. It rarely is, because the question was never "did they attend?" but "were they qualified to approve this?"
How does this connect to audit frameworks?
Auditors already treat training as part of controls in many frameworks. A SOC 2 report, for example, often tests whether staff with access to systems received appropriate training. When teams run agents, the scope naturally extends to the people who approve agent actions; SOC 2 and evidence of training for teams running agents works through what a service auditor is likely to sample.
The practical work of assembling records for an audit is covered in preparing training evidence for an AI audit, and a training evidence pack template gives a structure to fill in.
What is the trade-off?
Treating records as evidence makes them less flexible. Fixing a mistake means adding a correction rather than overwriting, which is clumsier. Every change leaves a trace, and some people find that uncomfortable.
There is a data protection dimension too. Records that are attributable, time-stamped and retained for audit are personal data, held longer and more carefully than a casual spreadsheet. They need a lawful basis, a retention period and access controls.
And it costs effort. Moving from a shared spreadsheet to a proper system, with history and date-based retrieval, takes time and sometimes money.
These costs are real and worth paying. The alternative is records that collapse under the first serious question, at exactly the moment the business needs them. Data protection is an argument for doing records properly, not for keeping them loosely: minimal, well-controlled, purpose-limited records are both better evidence and better practice.
What should you do differently?
- Stop editing history. Add corrections as new entries; never overwrite past records.
- Record dates of issue and expiry for every certificate that supports approval rights.
- Record what was assessed, including role, level and the product version where relevant.
- Make status retrievable for any date, so you can answer "was this person current on the day they approved?"
- Prefer records that outsiders can check at source, such as certificates looked up by number or code with the issuer.
- Link training records to approval rights, so a lapse has a defined, recorded effect.
- Set a retention period that matches how far back an auditor or supervisor might reasonably look.
The guide to AI literacy and certification for people who work with agents shows how certification and records fit together.
Questions people ask
Who should own training records for agent approvers?
Usually the function that grants approval rights, with risk or compliance holding a copy. HR may administer them, but ownership should sit where the control does.
Is a certificate from an external issuer enough on its own?
It is strong evidence of competence if it can be checked at source and shows dates. You still need your own record linking it to the approval rights you granted and when.
How long should training records be kept?
Long enough to cover any decision a supervisor or auditor might examine, which in regulated firms often means several years. Set the period deliberately and record why.