AI in risk, audit and compliance

    How AI is changing the work of audit, risk and compliance functions, what regulators and standard setters have said, and the state of the enforcement record as of October 2026.

    AI changes risk, audit and compliance work in two directions at once. It automates the work, replacing sampling with continuous monitoring and manual evidence gathering with automated collection. It also creates a new category of risk to audit, because the models themselves need governing. Organizations handling both sides usually anchor on the NIST AI RMF for risk process, ISO/IEC 42001 for a certifiable management system, and the EU AI Act where it legally applies. As of October 2026, the legal mandate to audit AI in the US and EU is narrowing or deferred, even as the professional infrastructure around AI assurance grows.

    The crucial role of AI in risk, audit and compliance

    AI shows up in a risk or compliance function in two ways, and they need different controls.

    • As a tool. Analytics over full populations, automated evidence collection, policy and regulatory change scanning, transaction monitoring, and drafting of workpapers and reports.
    • As a risk. Models that make or influence decisions, employee use of public AI tools on company data, AI features switched on inside existing SaaS, and AI used by attackers against the organization.

    The second category is where most functions are weakest. Shadow AI is the clearest example. In May 2026, Ontario's Auditor General found provincial staff using unsecured public AI websites, with the approved secure tool accounting for about 6% of generative AI use. An AI acceptable use policy sets expectations; network and SaaS telemetry shows actual use.

    AI and automation: what is actually being transformed

    The real changes in day-to-day work are these:

    • Full-population testing replaces sampling. Where data is structured and accessible, analytics test every transaction or access event. The PCAOB's amendments to AS 1105 and AS 2301 on technology-assisted analysis apply to audits of fiscal years beginning on or after 15 December 2025. They clarify the auditor's responsibilities when doing this and are technology-neutral rather than AI-specific.
    • Continuous control monitoring. Controls are tested as they operate rather than once a year, and exceptions are routed to owners.
    • Automated evidence collection. System connectors pull configurations, logs and approvals directly, which cuts the evidence-request cycle.
    • Drafting and summarization. Generative AI drafts workpapers, issue write-ups and reports, and summarizes regulation and policy.

    The vendor market has reorganized around this. AuditBoard rebranded as Optro in March 2026 with an agentic-AI pitch, and Deloitte launched ControlCatalyst.AI for internal audit, SOX, risk and compliance work on 5 August 2026. Vendor productivity claims remain unverified until measured in an organization's own pilot.

    How much AI audit teams actually use

    Survey figures vary widely, partly because they measure different things:

    These numbers measure different things: individual ad hoc use, licensed departmental tooling and formal programs. Mixing them up hides the gap between claimed and actual adoption, so adoption figures need to be read with the distinction in mind.

    AI's impact on audit practices

    What AI is used for is more consistent than how much it is used. The PCAOB's July 2024 staff update found generative AI in audits focused primarily on administrative and research activities. Later evidence, including CPAB's 2025 inspections below, is consistent with AI mostly drafting the audit rather than performing it.

    Where firms have moved into evidence, regulators are finding problems. In December 2025, Canada's CPAB reported that Canadian firms are embedding advanced technology and starting to use agentic AI. It also found that in its 2025 inspections some auditors relied on automated tools as the primary source of audit evidence while doing limited work to validate the completeness and accuracy of the data going in. This is a familiar data-reliability failure in a new form.

    Preparedness is the weaker number

    Capability lags further behind than adoption:

    • In a February 2026 survey of 373 senior North American internal audit leaders, the IIA and AuditBoard found 85% rated AI-enabled fraud a moderate-to-high risk, but fewer than 40% felt adequately prepared to detect it. Leaders named AI-powered phishing (88%), fabricated invoices or documents (65%) and deepfake impersonation (45%) among their concerns.
    • ISACA's 2026 AI Pulse Poll of more than 3,400 digital trust professionals found 56% did not know how quickly they could halt an AI system during a security incident. Only 12% said their organization had a documented shutdown or override process that is tested regularly.

    The test procedure for shutdown and rollback is in AI audit for IT auditors.

    What the professional bodies and audit regulators have published

    • UK Financial Reporting Council. Published AI in Audit guidance on 26 June 2025 with a thematic review of how the largest firms certify audit technology. It followed with guidance on generative and agentic AI on 30 March 2026, which the FRC describes as the first of its kind from an audit regulator. The FRC sets explainability expectations that depend on context rather than a fixed bar.
    • Canadian Public Accountability Board. Its December 2025 publication on the evolving use of technology in the audit (above), and a 2025 annual report themed on transparency, technology and trust.
    • PCAOB. It has issued no AI-specific standard. The technology-assisted analysis amendments took effect for fiscal years beginning on or after 15 December 2025. Under Chairman Demetrios Logothetis (sworn in 10 February 2026), the Board's updated agendas, published 30 September 2026 after its first public comment period, keep data and technology, including AI, as a research project. No AI standard-setting project has been opened.
    • IAASB. It issued exposure drafts revising ISA 330, ISA 500 and ISA 520 in August 2026, with comments due 15 December 2026. In Canada, the AASB tracks the project for Canadian Auditing Standards.
    • AICPA. As of October 2026, there is no AI attestation standard and no "SOC 2 for AI". SOC 2 still reports against the 2017 Trust Services Criteria (points of focus revised in 2022), which apply to AI systems like any other in-scope system. Vendor references to an AI-specific SOC 2 are not backed by any AICPA standard.
    • IIA. The Global Internal Audit Standards, effective 9 January 2025, make technology competence and resources explicit obligations (Standards 3.1 and 10.3). The IIA's AI Auditing Framework (September 2024) remains guidance. There is no AI Topical Requirement.

    How regulators are using AI in supervision

    Supervisors are adopting AI in their own work while asking firms to govern theirs:

    • Audit regulators. CPAB reports using generative AI tools in some of its own processes. The Office of the Auditor General of Canada's 2026–27 Departmental Plan commits to expanding AI and automation in its audits, building on experiments that showed measurable time savings.
    • Financial supervisors. The US SEC's Division of Examinations lists firms' use and supervision of AI among its FY2026 examination priorities. FINRA's 2026 Annual Regulatory Oversight Report has a dedicated generative AI section covering testing, supervision, vendor diligence and recordkeeping. The UK FCA and Bank of England use AI internally under the FCA's 2025–2030 "smarter regulator" strategy.
    • Implications for supervised firms. Regulators that use analytics can test full populations too. Requests for data rather than samples are likely, and an AI inventory is likely to be among the first documents asked for in an examination.

    When AI produces audit evidence that is wrong

    The clearest documented failure is real, not hypothetical. Deloitte Australia delivered a review of the Targeted Compliance Framework to the Department of Employment and Workplace Relations under a contract worth about A$440,000. The report contained fabricated academic references and a fabricated quotation attributed to a Federal Court judgment. The corrected version disclosed the use of a generative AI model (Azure OpenAI GPT-4o). In October 2025 Deloitte agreed to repay the final instalment, later reported at A$97,587. Fabricated content passed through a Big Four quality process into a government deliverable.

    The research on auditor reliance points the other way from most commentary. Commerford, Dennis, Joe and Ulla (Journal of Accounting Research, 2022) found that auditors receiving contradictory evidence from an AI system, rather than from a human specialist, proposed smaller adjustments to management's complex estimates. That is algorithm aversion, not overreliance. Both failure modes are now documented, and they pull in opposite directions. Practitioners worry about trusting AI too much, while the experimental evidence shows they discount it when it disagrees with management. Review procedures need to catch both.

    Four pressures push compliance teams toward AI tooling, and each brings its own audit question:

    • Regulatory volume and fragmentation. EU, Canadian, US state and sector rules change faster than manual horizon scanning can track. Audit question: who validates the AI's summary of a rule before it drives a control change?
    • Third-party and fourth-party risk. Vendors embed AI in products the organization already uses. The IIA's Third-Party Topical Requirement, effective 15 September 2026, applies when third-party risk is in scope of an assurance engagement. Audit question: does the vendor inventory record which suppliers use AI on your data?
    • Privacy and data protection. AI multiplies personal-data flows into prompts, logs and training sets. See AI and privacy.
    • AI-enabled fraud. Deepfake impersonation and synthetic documents target payment and onboarding controls. See AI identity fraud.

    Auditing AI itself

    Auditing an AI system is a different discipline from using AI to audit. For definitions, engagement types and the evidence auditors ask for, see what is an AI audit. For the control-by-control work program, see AI audit for IT auditors.

    The assurance market is maturing, but accreditation still trails demand.

    • Standards. ISO/IEC 42006:2025, published July 2025, sets requirements for bodies certifying against ISO/IEC 42001. Accreditation bodies including Canada's SCC, ANAB and UKAS now apply it. In the EU, EN 18286:2026, the AI Act quality management system standard, was published in mid-2026 and, as of October 2026, had not yet been cited in the Official Journal.
    • Firms. PwC launched Assurance for AI in June 2025, including in Canada. KPMG US obtained ISO/IEC 42001 certification for its own AI management system in November 2025, and KPMG Canada followed in 2026. Deloitte expanded its AI controls and assurance services in August 2026.
    • Market size. The UK's Trusted Third-Party AI Assurance Roadmap (DSIT, 3 September 2025) counted over 524 firms in the UK AI assurance market, generating about £1.01 billion in gross value added in 2024. It also noted that efforts to certify AI assurance processes are "few and far between".
    • Legal weight of voluntary frameworks. Texas's TRAIGA, in force 1 January 2026, gives a defence to organizations that substantially comply with the NIST AI RMF Generative AI Profile (AI 600-1) or another recognized framework. A voluntary framework now carries evidentiary weight.

    The regulatory mandate is contracting, not expanding

    As of October 2026, the direction in the US and EU is toward fewer or later audit mandates, with one important exception.

    • Colorado repealed its AI Act before it took effect. xAI sued on 9 April 2026 (xAI v. Weiser, D. Colo.) and the DOJ intervened on 24 April. Magistrate Judge Chung's order of 27 April 2026 barred enforcement of SB 24-205 pending the preliminary-injunction ruling. SB 26-189, signed 14 May 2026, replaced SB 24-205 with a narrower automated decision-making framework operative 1 January 2027. It drops the duty of care against algorithmic discrimination, the risk management program obligation and the impact assessment requirement.
    • The SEC withdrew its predictive data analytics proposal on 12 June 2025, leaving no SEC rule specific to AI use by advisers and broker-dealers.
    • The US federal approach is preemption, not mandates. Executive Order 14365, "Ensuring a National Policy Framework for Artificial Intelligence" (11 December 2025), led to a DOJ AI Litigation Task Force (formalized 9 January 2026) and to the White House's National Policy Framework for AI (20 March 2026), which asks Congress to preempt state AI laws. The framework is a legislative roadmap, not a binding rule, and as of October 2026 Congress had not enacted preemption.
    • The EU deferred. The Digital Omnibus on AI, in force 27 July 2026, moved Annex III high-risk obligations to 2 December 2027 and Annex I to 2 August 2028.
    • The exception is Canada's regulated sectors. OSFI's Guideline E-23, final on 11 September 2025 and effective 1 May 2027, explicitly brings AI and machine-learning models into enterprise model risk management for federally regulated financial institutions. The Treasury Board's amended Directive on Automated Decision-Making required existing federal systems to comply by 24 June 2026.

    One detail is frequently misreported. Under Article 43 of the EU AI Act, most Annex III high-risk systems are self-assessed under internal control, not audited by a notified body. A notified body is needed mainly for biometric systems under Annex III point 1 where harmonized standards or common specifications are not fully applied, and for AI in products whose sectoral law already requires third-party assessment. Penalties are also often overstated. Breaching high-risk obligations carries fines of up to €15 million or 3% of worldwide turnover. The €35 million or 7% tier is reserved for prohibited practices.

    Where mandates exist, compliance is close to nil

    New York City Local Law 144 is the only mature algorithmic audit mandate with measurable outcomes, and the outcomes are poor. Research presented at FAccT 2024 ("Null Compliance") had student investigators check 391 employers. Only 18 (about 5%) had posted a bias audit report and 13 (about 3%) had posted the required transparency notice. The authors argue that because the law lets employers decide whether their tool is in scope, a missing audit cannot even be proven to be non-compliance.

    The regulator's own performance was examined too. The New York State Comptroller's audit of 2 December 2025, covering July 2023 to June 2025, found the Department of Consumer and Worker Protection had received only two complaints in that period. Three-quarters of the auditors' test calls to 311 about these tools were misrouted. DCWP's own review of 32 companies found one instance of non-compliance, while the Comptroller's reviewers found at least 17 potential issues at the same companies. DCWP has committed to proactive enforcement.

    Enforcement so far

    Enforcement has come through existing law rather than AI statutes:

    • Employment discrimination. EEOC v. iTutorGroup settled for US$365,000 in August 2023 over software configured to auto-reject older applicants.
    • Securities. The SEC brought its first AI-washing actions in March 2024, fining Delphia US$225,000 and Global Predictions US$175,000 for overstated AI claims.
    • Consumer protection. The FTC banned Rite Aid from using facial recognition for five years in December 2023, and finalized a US$1 million order against accessiBe in April 2025 over unsupported claims for its AI accessibility tool.
    • Litigation. In Mobley v. Workday, a nationwide age-discrimination collective was conditionally certified in May 2025. In late May 2026, Magistrate Judge Laurel Beeler held that bias-testing data curated by counsel and used to give legal advice was protected by attorney-client privilege, while ordering production of EEO-1 and OFCCP material. That creates a direct tension with the transparency purpose of bias auditing, which is best resolved before an AI audit starts.

    No penalties under the EU AI Act had been publicly reported by October 2026, and no audit firm has been publicly sanctioned by an audit regulator for AI misuse in an engagement. See AI lawsuits and where the courts have landed.

    Priorities for risk, audit and compliance leaders

    • Separate the two agendas. Track "AI as a tool" (productivity, quality and validation of AI-generated work) separately from "AI as a risk" (inventory, governance and assurance of AI systems).
    • Validate AI-generated evidence. Apply the CPAB lesson. Test the completeness and accuracy of data feeding automated tools, and check citations in anything AI drafted.
    • Plan to the deferred dates, not away from them. EU high-risk obligations still arrive in 2027 and 2028, and OSFI E-23 in May 2027.
    • Test the off-switch. Most organizations cannot say how fast they could stop an AI system.
    • Do not wait for a mandate. The legal requirement to audit AI is narrowing in places, but boards, customers and insurers are not waiting for it. See board oversight of AI.

    Sources and references

    Frequently asked questions

    How is AI changing internal audit?

    Mostly by speeding up drafting, research and summarization, and by enabling full-population analytics and continuous control monitoring where data allows. Use in core evidence-gathering is growing but uneven. Regulators such as CPAB have found auditors relying on automated tools without validating the underlying data, so the change brings new review obligations with it.

    Is there a SOC 2 for AI?

    No. As of October 2026 the AICPA has not issued AI-specific Trust Services Criteria or an AI attestation standard. SOC 2 reports cover AI systems under the existing 2017 Trust Services Criteria. Organizations wanting AI-specific assurance use ISO/IEC 42001 certification, attestation engagements against stated criteria, or both.

    Does the EU AI Act require an independent audit of AI systems?

    Not for most systems. Most Annex III high-risk AI is self-assessed by the provider under internal control. A notified body is required mainly for certain biometric systems and for AI in products already subject to third-party assessment. After the Digital Omnibus, these obligations apply from 2 December 2027 and 2 August 2028 respectively.

    What have audit regulators said about using AI in audits?

    The UK FRC issued AI in Audit guidance in June 2025 and generative and agentic AI guidance in March 2026. Canada's CPAB reported inspection findings on automated tools in December 2025. The PCAOB found generative AI used mainly for administrative and research tasks, and keeps data and technology on its research agenda without an AI-specific standard.

    Can AI be used for regulatory compliance monitoring?

    Yes, for horizon scanning, transaction and communications monitoring, control testing and evidence collection. The compliance function remains accountable for the result. Validate AI summaries of regulation before they drive control changes, keep human review for high-impact decisions, and include the compliance AI tools themselves in the AI inventory and audit plan.

    What is AI-washing?

    AI-washing is overstating or misrepresenting the use of AI in products, services or investment processes. Regulators treat it as ordinary misrepresentation. The SEC fined two advisers in March 2024 and the FTC has acted against unsupported AI marketing claims. Compliance teams should require evidence for every external AI claim.

    Share on LinkedIn