Rehearse before the auditor arrives. Prepare AI agents for an audit by fixing the scope and period, taking a snapshot of every agent's brief, capabilities and depths as they stood, checking the decision record has no gaps, and reconstructing a handful of decisions yourself. Then list the known issues before the auditor finds them. Each step leaves a document.
What does audit preparation cover, and who owns it?
Audits of agents come in several kinds: internal audit, an external auditor testing controls that agents operate, a certification audit against a management system standard, or a supervisor's review. All of them ask the same core questions. What was each agent allowed to do during the period? Who decided that? Did it stay within those limits? Can you show it?
The agent operations lead owns preparation. Process owners answer for their agents. Risk and compliance confirm the control set being audited. The platform owner makes sure records can be produced. If internal audit is new to agents, brief internal audit on how AI agents are governed before fieldwork rather than during it.
What are the steps, and who acts at each one?
- Person, agent operations lead, with the auditor: agrees scope, period and criteria. Which agents, which processes, which dates and against which policy or standard.
- Agent, depth 1: builds the population. Every agent active in the period, with its status changes, drawn from the inventory and identity records.
- Agent, depth 2: assembles snapshots. For each in-scope agent, the brief, role scope, capability and depth matrix, sources and limits as they stood at each point in the period, with change history between them.
- Agent, depth 1: checks record completeness. It looks for days with no records from an active agent, verdicts without a named bound, held actions without an approver, and actions under a lapsed certificate.
- Person, process owners: explain or fix each gap, in writing.
- Person, agent operations lead: reconstructs sample decisions. Pick five or ten actions, including a refusal and a held action, and trace each from request to filed verdict using only the record. If you cannot, neither can the auditor.
- Agent, depth 2: drafts the evidence index: policies, approvals, certificates, change log, incident log and access reviews, each with its location.
- Person, agent operations lead: writes the known-issues list, with status and remediation dates.
- Person, platform owner: sets up read-only access for auditors under their own identities, time-limited to the fieldwork.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Scope | Nothing | none | Agreeing scope and criteria |
| Population | Lists agents active in period | 1 | Nothing |
| Snapshots | Assembles limits as they stood | 2 | Nothing |
| Completeness | Finds gaps in the record | 1 | Nothing |
| Explain gaps | Nothing | none | Owners' written explanations |
| Rehearsal | Nothing | none | Tracing sample decisions |
| Evidence index | Drafts the index | 2 | Checking it |
| Known issues | Nothing | none | Disclosing problems |
| Access | Nothing | none | Granting auditor access |
What evidence does each step leave?
- Scope: an agreed memo with period and criteria.
- Population: the list of in-scope agents with status dates.
- Snapshots: versioned documents per agent for the period.
- Completeness: a gap report.
- Explain gaps: signed explanations and any fix made.
- Rehearsal: the traced decisions, with record identifiers.
- Evidence index: a document mapping each control to its evidence.
- Known issues: a dated list with owners.
- Access: the access approval and its expiry.
What will the auditor test?
Expect tests of design and operation. Design: does the policy say what agents may do, are limits set per capability, are there named owners and approvers? Operation: for a sample of actions, was each within the agent's limits at that moment; were held actions approved by someone on the roster; did refusals happen when they should; were certificates current? The approach is set out in how do you audit what an AI agent did?
Every one of those tests depends on the record. If you are unsure what makes a record usable for this, read what is an audit trail? before preparing anything else.
Which controls make preparation quick?
Point-in-time snapshots. Auditors test what applied on the date of an action, not today's settings. Versioned limits with dates make this possible.
An append-only record. If records could have been changed, the auditor must test the change controls too. A record that cannot be altered removes that work.
Named bounds on every verdict. A verdict that says which bound decided lets the auditor confirm the limit worked without re-running the agent.
One evidence index. Auditors lose time asking where things are. Give them the map on day one.
What should you measure?
Time to produce a requested sample, gaps found during preparation compared with gaps found by the auditor, and findings repeated from the previous audit. A finding the auditor raises that your own gap report missed tells you where preparation is blind.
Where does it usually go wrong?
Current settings are shown as historical. The auditor asks what an agent could do in March, and the team shows today's configuration. Keep the history.
Records exist but cannot be read. Logs are complete but technical, and nobody outside engineering can follow them. Rehearsal exposes this early.
Gaps are explained after the auditor finds them. An explanation offered in advance is a known issue; the same explanation offered later reads as a control failure.
Vendor agents are out of scope by accident. If a vendor's agent acted on your data, it belongs in the population.
Where Pryme fits
In Pryme, every verdict is filed in an append-only record with the bound that decided, and each agent's certificate history shows which depths applied when. That makes snapshots and sample tracing a matter of retrieval. The control model being audited is described in agent governance is a bound, not a model.
Questions people ask
How far ahead should preparation start?
Several weeks before fieldwork, so gaps can be explained or fixed. The rehearsal is the step that most often reveals work, so do it first if time is short.
Should we use an agent to answer auditors' questions?
An agent can retrieve records and draft summaries at depth 1 or 2. Answers to the auditor should come from named people, who stand behind them.
What if a sampled decision cannot be reconstructed?
Disclose it, explain why, and show what has changed so it cannot recur. A disclosed gap with a fix is a smaller finding than an answer that does not hold up.