Internal auditors need to know what each recertification should leave behind, so they can test it without taking part in it. Every 90 days, each agent's certificate lapses unless renewed on the organisation's own cases. Auditors stay ready by refreshing three skills each cycle: reading gate results, tracing verdicts in the record, and checking that the people approving held work were current.
What does the 90-day cycle mean for internal audit?
For operations teams, recertification changes the work. For internal audit it creates a fresh population to test. Each cycle should produce a certificate per agent, gate results, a change note, briefing records for the people affected and, if limits changed, a record of who changed them. If any of those is missing, the control has a gap.
Auditors may also use an agent themselves. An audit agent can sample records at depth 1, draft workpapers and test scripts at depth 2, and draft findings at depth 2. It does not rate findings, agree management actions or sign the report. Those stay with the auditor, and the audit agent is itself recertified every 90 days, which auditors need to track like anyone else.
For example, an auditor testing a claims agent picks 25 settlements from the week after the effective date. For each, they check the certificate in force, the depth it allowed, the organisation's limit at the time and, where the payment was held, the approver and that approver's certificate status on that day. One settlement released above the new limit is a finding, whatever its amount.
The skills that decay fastest are technical. Auditors who last looked at an agent two cycles ago may not know that a capability moved from depth 2 to depth 3, which changes the test from "was every draft reviewed" to "were exceptions handled and limits respected".
Who owns what?
The head of internal audit owns the audit plan and auditor readiness. The builder owns recertification. Management owns the controls being tested. Auditors must not sign change notes or approve depth changes, because they would then be auditing their own decisions.
Steps for each cycle
- Agent, depth 1: list certificates due across the organisation in the coming quarter, with capabilities and current depth.
- Person (audit manager): choose which cycles to test, weighting to agents at depth 3 or 4 and to any capability that failed in the last two cycles.
- Person (auditors): refresh on what changed. Read the change notes for selected agents and any update to the gates or depth model.
- Agent, depth 1: sample the record for the selected agents: verdicts before and after the effective date, held actions and their approvers.
- Person (auditor): test. Was recertification complete before lapse? Did the agent act above its new depth after the effective date? Were approvers' certificates current at the time of approval?
- Agent, depth 2: draft workpapers and findings.
- Person (auditor): edit, rate and agree findings with management.
- Person (head of internal audit): report to the audit committee.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Population | Lists certificates and depths due | 1 | Choosing what to test |
| Sampling | Pulls verdicts, holds and approvers | 1 | Setting the sample |
| Testing | Matches approvals to certificate status | 1 | Judging exceptions |
| Workpapers | Drafts | 2 | Reviewing |
| Findings | Drafts | 2 | Rating and agreeing actions |
| Reporting | Nothing | none | The head of internal audit |
What should auditors relearn each cycle?
Gate results. Seven gates must each clear on their own: grounded, refuses, depth, escalates, scope, adversarial and stable. Auditors should be able to say which gate a capability passed narrowly and whether management responded.
The minimum chain. Every action meets five bounds at once: role scope, certificate, sources, the organisation's own limits and the platform guardrails. The lowest wins, and the verdict names which bound decided. That named bound is the auditor's most useful field.
Refused, held and withheld. A test that counts refusals as failures has misread the system. Refusals and withheld actions are controls working. Held actions are where human judgement must show in the record.
Approver currency. Professional certificates for people last 12 or 24 months and can lapse or be revoked. Auditors should check status at source on the date of each sampled approval, not today. Pryme Intelligence Academy certificates, for example, can be checked by number or QR code through VerifyDoc.ai.
Controls auditors test, and the evidence each leaves
- Recertification before lapse. Evidence: certificate dates against verdict dates.
- Change note issued and briefed. Evidence: dated note and attendance.
- Limits changed by authorised people. Evidence: limit history with names.
- Approvals by certified people. Evidence: approver, time and certificate status for each held action.
- Auditor readiness. Evidence: the practicum log template for internal auditors for auditors who approve audit-agent outputs.
What to measure
For the audit function: the share of agents at depth 3 or 4 tested at least once a year, repeat findings across cycles, and the time management takes to close findings. For auditor readiness: whether each auditor on an agent engagement has read the latest change note before fieldwork begins.
Where does it go wrong?
Auditors test against an old depth. The workpaper says "every draft was reviewed" when the agent now executes at depth 3 and reviews happen by exception. The test passes and proves nothing.
Auditors become part of the control. They are asked to sign off a change note "for comfort". Decline politely; the role is independent.
Auditors approve audit-agent outputs without understanding them. Whether they need a certificate is a choice the function makes; see do internal auditors need a certificate to approve AI agent actions and are internal auditors accountable for an AI agent's mistakes. Risk managers own a related cycle; compare keeping risk managers ready and the guide to AI literacy and certification for people who work with agents.
Questions people ask
How often should internal audit test recertification?
Not every cycle for every agent. Test each agent at depth 3 or above at least yearly, and any agent that failed a run soon after.
Can auditors rely on the builder's gate results?
They can use them as evidence, but should re-perform a sample: pick a few replay cases and check the scoring.
What if the record has a gap?
Report it as a finding. An append-only record with a gap means something else failed, and management must explain what.