Internal auditors in Canadian departments and agencies need training to audit AI agents that pre-screen grant applications, check benefit files and draft correspondence. They must test verdicts, approvals and certificates from the record, and check the agent against federal policy on automated decisions. Audit conclusions stay with them. Public-sector privacy law and transparency duties shape oversight.
What does internal audit focus on in Canadian government?
Departmental internal audit serves the deputy head and a departmental audit committee that includes external members. The risk-based plan covers grants and contributions, where funding agreements must be monitored against eligibility and results; benefits and services delivered to the public; procurement and contracting; IT security; and financial controls. Auditors also follow up management action plans and report on whether they were carried out.
The function is expected to be independent and to follow professional internal audit standards. Its reports are often published, so findings must be clear to a non-specialist reader, and conclusions must rest on evidence that would survive an access to information request.
What changes when agents arrive?
Programmes are beginning to use agents for document-heavy steps, each with its own depth.
- Grant application screening (depth 1). The agent checks an application against published eligibility criteria and lists missing items. A programme officer decides.
- Benefit file completeness checks (depth 1). It reviews a file and flags gaps before an officer assesses entitlement.
- Correspondence drafting in both official languages (depth 2). It drafts replies to applicants. An officer checks both versions and sends.
- Contract milestone tracking (depth 3). Within the department's limits, it sends reminders to suppliers on reporting dates and logs replies.
Audit can also use an agent: at depth 1 to test every grant screening verdict in the append-only record, and at depth 2 to draft workpapers that an auditor reviews and signs.
What must auditors learn, and what stays with them?
Auditors should learn the system's vocabulary first. Our guide on what internal auditors should learn before working with AI agents lists it: depth, the minimum chain of five bounds, the four verdicts and the record.
The audit test that matters most in government is whether a person decided where the policy requires one. Federal policy on automated decision-making scales the level of human involvement to the impact of the decision. An auditor should be able to take a sample of eligibility outcomes and show, from the record, that an officer made each one and that the agent's depth was limited to screening. Our list of what internal auditors should never let an AI agent do is a useful starting checklist.
Certification and provenance follow. Agent certificates lapse after 90 days and a run that clears no depth drops the agent to depth 1 the same day, so auditors test certificate dates against verdict dates. On provenance, an eligibility fact taken from the programme's own system of record is sealed; one stated in the application is declared. Auditors should test that declared facts were verified before any decision relied on them.
Auditors who want hands-on knowledge of review can take operator-level training and keep a practicum log template for internal auditors as evidence.
What stays with the auditor: planning, conclusions, recommendations, and reporting to the deputy head and audit committee.
Which Canadian rules shape oversight and literacy?
Canada has no federal AI act in force. For federal institutions, government policy on automated decision-making requires an assessment of a system's impact before it is used for administrative decisions, with requirements for notice, explanation, human involvement and peer review that rise with the impact level. Auditors should test that the assessment exists, matches how the agent is used, and is kept current as depth changes.
Privacy in government is governed by public-sector privacy law, federal or provincial, rather than PIPEDA, which covers the private sector. That law limits collection and use of personal information and gives individuals access rights. Public bodies also carry transparency and fairness duties that apply to every administrative decision.
OSFI and FCAC supervise financial institutions and do not reach a department's programmes. ISO/IEC 42001 and the NIST AI Risk Management Framework are useful references for structuring an AI audit.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
A 90-day training plan for an audit unit
Month one. All auditors complete basic literacy. The chief audit executive adds agent deployments to the audit universe and asks programmes for their impact assessments.
Month two. Run an advisory engagement on one deployment, such as grant screening. Test that officers made every eligibility decision, that the agent stayed at depth 1, that certificates were current, and that bilingual correspondence matched.
Month three. Write the standing audit programme for agents, link it to the impact assessment requirements, and set up banking corrections for audit's own agent.
Where the Academy fits
Pryme Intelligence Academy offers a free readiness assessment for teams of up to 25. An audit unit takes the Foundations test set in the public sector, and the report shows scores by concept and the gaps to close. Auditors leading agent engagements can then take the Operator track, whose certificates anyone can check at source through VerifyDoc.ai.
Questions people ask
Should internal audit be consulted before an agent goes live?
An advisory engagement before go-live is useful and does not compromise independence if audit does not design the controls. It often catches a missing impact assessment.
How do auditors test bilingual correspondence?
By sampling replies from the record and comparing both versions with each other and with the source. Differences in substance are findings.
How should audit report agent findings to the departmental audit committee?
In plain terms: which agent, which decision, which bound held or failed, and what it means for the programme. External members may not know the vocabulary, so a short glossary helps. Our piece on why internal auditors are the real control and the guide to AI agent skills by role give useful framing.