AI in internal audit tests every transaction in a population against control rules instead of a sample, ranks the exceptions so auditors start with the likely real issues, and drafts workpapers and summaries. The auditor still chooses the method, judges each exception and signs the conclusions, as the IIA Global Internal Audit Standards (in effect since January 9, 2025) expect. Remolda's AI Readiness Review names the first process to test in one week for $490 CAD plus applicable taxes.
Ready to start? Send a written request with your sector (bank, insurer, credit union, public body, company), the processes you audit most often and the systems that hold the data, or call +1 (613) 867‑7767. We reply within one business day.
What AI changes in an audit, and what the auditor keeps
| Audit task | With AI | The auditor keeps |
|---|---|---|
| Risk assessment and planning | Draft risk notes from prior findings, incidents, KRIs and policy changes | Ranking risks and setting the plan |
| Control testing | Every transaction tested against agreed rules | Choosing the tests and the method |
| Exceptions | Ranked by likelihood and value, with the reason shown | Judging each exception |
| Fieldwork documents | Contracts, policies and minutes compared against the control | Reading the source before relying on it |
| Workpapers and reports | Drafts from the test results and notes | Conclusions, ratings and sign-off |
| Follow-up | Re-running the same tests after remediation | Closing the finding |
AI in this table is a tool of the audit team. Independence, evidence and sign-off stay as they are; AI changes how much of the population the team can look at for the same effort.
Full-population testing and the exception queue
For controls that can be tested from system data, such as payment approvals, segregation of duties in the ERP, vendor master changes, duplicate invoices or access rights, scripts and models can check the whole year rather than a sample of 25 or 60 items. The catch is volume. A full population produces many more exceptions than a sample, many of them explained by legitimate business reasons, and an exception list nobody works through is worse than no test.
Two design choices make it work. First, rank the exceptions: by amount, by how far they sit from normal activity for that vendor or user, and by whether the same pattern appeared in earlier findings, with the reason for each rank visible to the auditor. Second, give the queue an owner and a rule for which exceptions get examined, which get sampled, and which are closed with an explanation from management.
Sampling stays for evidence on paper, inspections and processes whose data is too costly to extract. The auditor documents which method was used and why.
Continuous monitoring and continuous auditing
The two terms are often mixed up, and the difference matters for independence.
- Continuous monitoring belongs to management. The accounts payable manager or the access administrator runs automated checks on their own controls, every day or week, and fixes what fails. It sits in the first or second line.
- Continuous auditing belongs to internal audit. Audit runs its own tests on a recurring schedule and reports to the audit committee. Where management already monitors a control well, audit can test that monitoring and rely on it, rather than duplicating it.
One workable path: audit builds tests during an engagement, hands the stable ones to management as monitoring, and keeps periodic checks that the monitoring still runs. Federal departments will find the Treasury Board's guide on ongoing monitoring of internal controls over financial management useful for the management side. Building the dashboards and alerts is the kind of work our decision support and data pipeline projects cover.
Generative AI in fieldwork and reporting
Language models help with the reading and writing around the tests: summarizing prior reports and incidents for planning, comparing a contract or policy against the control it should support, drafting interview notes, workpapers and the first version of a report. Three rules keep this safe:
- Confidential data stays in an approved tool. Audit files hold personal information, investigations and board material. Use a tool your organization has approved, with data location and retention settled in the contract.
- Every draft cites its source, and the auditor reads the source before relying on it.
- Conclusions and ratings are written by the auditor. A model can draft the wording; it does not decide whether a control is effective.
A written AI use policy for staff, including auditors, is where these rules become day-to-day instructions.
Standards and rules that apply in Canada
| Standard or rule | Who it covers | What it means for AI in audit |
|---|---|---|
| IIA Global Internal Audit Standards | Internal audit functions that conform to the IIA | Principle 10: technology plan, regular evaluation, training, reporting limits to the board |
| Treasury Board Directive on Internal Audit | Federal departments | IIA framework applies, plus mandatory federal procedures |
| OSFI Guideline E-23 | Federally regulated financial institutions, from May 1, 2027 | AI/ML and vendor models: inventory, rating, independent review, monitoring |
| AMF AI guideline | Quebec-authorized financial institutions, from May 1, 2027 | Register of all AI systems with a risk rating |
| COSO guidance on generative AI | Any organization using the COSO framework | Controls over GenAI mapped to the five COSO components |
IIA Standards. Issued January 9, 2024 and in effect since January 9, 2025. They do not single out AI. Principle 10 (Manage Resources) covers technology: the chief audit executive should strive to provide technology that supports the audit process, evaluate it regularly, train auditors when a new tool is introduced and report to the board and senior management when technology limits the function's effectiveness.
Treasury Board Directive on Internal Audit. In effect since June 15, 2023, it replaced the 2017 directive. The chief audit executive applies the IIA's International Professional Practices Framework, unless it conflicts with the Treasury Board policy or directive, together with mandatory federal procedures.
OSFI Guideline E-23. In effect from May 1, 2027. It does not name internal audit, but its model definition includes AI and machine learning methods and it covers vendor models, so an anomaly-scoring model used by audit can be a model in the institution's inventory. E-23 asks for review independent from model development and allows the use of internal reviewers or objective third parties. In their September 2024 risk report, OSFI and FCAC reported that about 50% of financial institutions used AI in 2023, up from about 30% in 2019, and 70% expected to by 2026, so the institution's own AI models belong on the audit plan.
AMF guideline (Quebec). Issued April 7, 2026 and in effect from May 1, 2027 for authorized insurers, financial services cooperatives, trust companies and deposit institutions: a register of all AI systems, a risk rating for each and a senior manager accountable for AI.
COSO. The 2013 Internal Control – Integrated Framework has five components: control environment, risk assessment, control activities, information and communication, and monitoring activities. COSO's publication Achieving Effective Internal Control Over Generative AI (2026) applies those components to generative AI, groups uses into eight capability types such as ingestion, transformation, posting and judgment, and includes control mapping and testing procedures auditors can use.
Our AI compliance review maps these to your own AI systems, which is also the starting point for an audit of AI governance. For banks and credit unions, see AI for banks in Canada; for public bodies, AI for federal departments.
Where to start
Pick one process with high volume, data already in a system and a clear control owner:
- Accounts payable: duplicate invoices, payments over approval limits, vendor master changes followed by a payment, invoices split under a threshold.
- Employee expenses: policy limits, weekend or duplicate claims, missing receipts.
- User access: segregation-of-duties conflicts, access kept after a role change or departure, privileged accounts without review.
Agree the tests with the process owner, run them on a full year, and measure how many exceptions turn out to be real. That number decides whether the test moves to a monthly schedule and who owns it. In our experience, extracting and cleaning the data takes longer than writing the tests; it depends on the ERP and on access approvals.
How Remolda works with audit teams
- AI Readiness Review ($490, one week): a working session with the chief audit executive and the team, a readiness score across data, process, people, leadership, infrastructure and culture, and three to five candidate tests or tools.
- AI Opportunity Audit ($2,900, two weeks): interviews with audit, finance, risk and IT, ranked use cases for audit and for the controls it relies on, and a 12-month roadmap.
- AI Pilot Sprint ($9,800, six weeks): one set of tests or one assistant built in your environment, with the auditor reviewing every exception and the result measured at week six.
All prices are in CAD plus applicable taxes and listed on the pricing page. Audit teams that want to use the tools themselves can add department training.
Know which control you would test on a full population first? Send a written request with the process, the system and the period. We reply within one business day.
Sources
- The IIA: Global Internal Audit Standards (issued January 9, 2024, effective January 9, 2025)
- The IIA: How do the Standards address the use of technology in auditing?
- Treasury Board of Canada Secretariat: Directive on Internal Audit
- OSFI: Guideline E-23, Model Risk Management (2027)
- OSFI and FCAC: Risk Report, AI Uses and Risks at Federally Regulated Financial Institutions (September 2024)
- AMF: Guideline for the Use of Artificial Intelligence
- COSO: Achieving Effective Internal Control Over Generative AI
- Internal Auditor: COSO issues GenAI guidance (March 2026)

