> ## Content Index > Fetch the complete content index at: https://fintekcafe.com/llms.txt > Use this file to discover other available public pages before exploring further. # AI in Internal Audit: How Auditors Use It, and How to Audit the AI the Business Already Runs - URL: https://fintekcafe.com/ai-internal-audit-explained/ - Published: 2026-11-07T09:00:00.000Z - Updated: 2026-11-07T09:00:00.000Z - Description: AI in internal audit: full-population testing, continuous monitoring, where evidence breaks down, and an assurance checklist for the AI your business runs. - Author: FinTekCafe - Tags: finance, audit, ai, governance, seo **AI in internal audit means two different jobs: using AI to test controls and transactions faster and more completely than sampling allows, and giving the board assurance over the AI systems and agents the rest of the business has already deployed.** Most audit functions have started on the first. Gartner found in a survey of 119 chief audit executives, published in January 2026, that 83 percent were piloting or using AI. The second job is further behind. Only 21 percent of Fortune 100 companies disclosed audit committee oversight of AI in 2025, according to EY's review of proxy filings, and only 22.6 percent of boards in a 2025 NACD survey had audited where AI is used in their company. The position taken here: the bigger risk, and the bigger value, sits in the second job. An audit function that uses AI to test expenses but cannot tell the audit committee which agents approve refunds, screen candidates or onboard suppliers is optimizing the wrong half. ![Audit procedures](https://fintekcafe.com/content/images/2026/11/inline-01-audit-procedures.png) ## Key Takeaways - Internal audit sits between the board and every function, so AI changes both its tools and its scope. Treat them as two programs with two owners. - AI makes full-population testing practical. US public-company auditing standards now explicitly address technology-assisted analysis of entire populations, effective for fiscal years beginning on or after December 15, 2025. - The evidence standard does not relax because a model did the work. A government report from Deloitte Australia that contained fabricated references, and led to a partial refund in October 2025, is the cautionary case. - Auditing AI in the business starts with an inventory, then tests ownership, data lineage, change control after model updates, and logging of agent actions. - The main frameworks are the IIA's AI Auditing Framework, the NIST AI Risk Management Framework, ISO/IEC 42001 and, since February 2026, COSO's guidance on internal control over generative AI. None is a mandatory AI audit standard. ![Oversight gap](https://fintekcafe.com/content/images/2026/11/inline-02-oversight-gap.png) ## Two Jobs at Once Internal audit gives independent assurance to the board and audit committee that risks are managed and controls work. Every function falls inside its scope: HR, procurement, customer service, IT, finance, sales. That position is why AI hits it twice. On the tool side, audit has always been constrained by sampling. A team testing expense claims looks at 60 out of 60,000 and extrapolates. AI and analytics make it possible to look at all 60,000, flag the anomalies and spend human time on those. On the scope side, the business is deploying AI faster than governance is keeping up. ISACA's 2026 AI Pulse Poll of more than 3,400 professionals found 90 percent saying employees use AI at work, but only 38 percent reporting a formal, comprehensive AI policy. Every one of those deployments is a set of controls that someone will eventually need to give assurance on. The adoption figures for audit itself are high, with caveats. Gartner's January 2026 survey found 83 percent of audit functions piloting or using AI. A Gartner webinar poll of 161 chief audit executives in May 2026 found 93 percent reporting some AI use but 60 percent without a formal AI strategy. Webinar polls are not probability samples, but the direction is consistent: broad experimentation, thin structure. ![Assurance checklist](https://fintekcafe.com/content/images/2026/11/inline-03-assurance-checklist.png) ## AI for the Audit Itself ### Full-population testing **Full-population testing means running a test over every transaction or record in scope rather than a statistical sample.** The analytics are not new; what AI adds is the ability to read unstructured evidence, such as invoices, contracts and approval emails, at the same scale as structured data. Standard setters have moved. The PCAOB, which sets auditing standards for US public-company external audits, adopted amendments in June 2024 addressing technology-assisted analysis, noting it can be used to perform substantive procedures on all items in a population. They apply to audits of fiscal years beginning on or after December 15, 2025\. Internal auditors are not bound by PCAOB standards, but their external auditors are, and the expectation travels. ### Continuous controls monitoring Instead of testing a control once a year, continuous monitoring checks it every day: segregation of duties in the ERP, approval limits on payments, privileged access reviews. Exceptions go to the control owner and to audit. This is the audit counterpart of the finance function's own shift, described in [how AI enables the continuous close](https://fintekcafe.com/continuous-close-ai-finance-function/). ### Anomaly detection in expenses, payments and access logs Fraud is the classic use. The ACFE's 2026 Report to the Nations, covering 2,402 cases in 143 countries, estimated that organizations lose 5 percent of revenue to fraud each year, put the median loss per case at 104,000 US dollars, and found 43 percent of frauds detected by tips. Anomaly detection gives audit a second detection channel. It also has a new adversary: generative AI makes convincing fake receipts trivial to produce, as covered in [how AI-generated receipts beat expense controls](https://fintekcafe.com/ai-fake-receipts-expense-fraud-explained/). An IIA and AuditBoard survey of 373 audit leaders, published in February 2026, found 27 percent rating AI-enabled fraud a high risk and fewer than 40 percent believing their function was adequately prepared. ### Drafting workpapers and reports The most common current use is writing. A Gartner webinar poll of 743 audit professionals, released in August 2026, found 60 percent using AI to draft audit issues, ratings or reports, but only 30 percent using it for audit testing. That ratio matters: writing is where AI saves time, testing is where it changes coverage. The large audit firms are building at scale. KPMG added AI agents for expense vouching and searching for unrecorded liabilities to its Clara platform in April 2025\. Deloitte added generative and agentic capabilities to its Omnia platform in July 2025\. EY announced agentic AI in its Canvas platform in April 2026, with a target of agent support across all end-to-end audit activities by 2028. ### Audit procedures transformed by AI | Procedure | Old method | AI method | Evidence caveat | | --------------------------------- | --------------------------------------------- | -------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------- | | Expense claim testing | Statistical sample, manual receipt review | Every claim scored; receipt images checked for duplication and manipulation | AI-generated receipts can pass image checks; corroborate with card and merchant data | | Payment and vendor master testing | Sample of payments; annual vendor file review | All payments matched to purchase orders and contracts; duplicate and shell-vendor patterns flagged | Model flags are leads, not findings; each exception needs human-verified source evidence | | Access and segregation of duties | Annual review of role assignments | Daily monitoring of entitlements and conflicting access | Logs must be complete; a monitoring gap is an evidence gap | | Journal entry testing | Risk-based filters on a sample | All entries scored for unusual timing, users, amounts and descriptions | Scoring logic must be documented well enough to re-perform | | Contract compliance | Sample of contracts read by auditors | Clause extraction across all contracts; obligations compared to actual billing | Extraction error rates must be measured and disclosed in the workpaper | | Workpaper and report drafting | Auditor writes from notes | Model drafts from test results; auditor edits | Every factual statement and reference checked against source | ## Where AI in the Audit Fails ### Explainability to the audit committee An audit finding has to stand up to the person whose area was audited, and then to the audit committee. "The model scored this transaction as anomalous" is not a finding. The audit team must be able to explain why, in terms of the evidence, and re-perform the test without the model if challenged. ### Evidence standards The cautionary case is not from internal audit, but it sets the standard. In October 2025 Deloitte agreed to partially refund an AUD 440,000 contract with Australia's Department of Employment and Workplace Relations after a report it produced was found to contain fabricated references and a made-up quote from a court judgment. Deloitte disclosed it had used a generative AI model and said the recommendations were unchanged. The reputational damage came from the citations, not the conclusions. Regulators have responded with documentation requirements rather than prohibitions. The UK Financial Reporting Council published its first guidance on AI in audit in June 2025, focused on how firms document and certify their tools. The PCAOB staff published observations on generative AI in July 2024, noting firms mostly used it for administrative and research tasks and emphasized human review. ### False positives that swamp the team Full-population testing finds more exceptions, most of them benign. An audit team that runs an anomaly model over a million payments and gets ten thousand flags has created a workload problem, not an insight. Tune thresholds on historical data before the live run, and report the false-positive rate alongside the findings. ## Auditing the AI the Business Already Runs This is the half of the job most audit plans underweight. The questions are familiar control questions, applied to a new kind of system. ### Inventory You cannot audit what you cannot find. The starting point is an inventory of AI systems and agents: what each does, which data it touches, which decisions it makes or influences, and who owns it. It should include AI features switched on inside existing SaaS tools, which are the easiest to miss. The NACD finding that fewer than a quarter of boards had audited where AI is used suggests most organizations do not yet have one. ### Ownership Every model and agent needs a named business owner accountable for its outcomes, not just a technical owner who keeps it running. An agent that approves refunds belongs to the head of customer service, not the AI platform team. ### Data lineage Which data trained or grounds the system, where it came from, whether the organization had the right to use it, and whether it contains personal data. Lineage failures become privacy and discrimination findings later. ### Change control after a model update Traditional change control assumes the organization decides when software changes. With AI, the vendor may update the underlying model without notice, and behaviour can shift with no code change on your side. Audit should test whether the business re-validates AI systems after model updates, and whether it would notice if a vendor changed one. ### Logging of agent actions An agent that takes actions, such as issuing refunds, changing records or sending emails, must leave a log complete enough to reconstruct what it did and why. Without it, there is no audit trail. The engineering for this is covered in [agent observability in production](https://fintekcafe.com/agent-observability-production-guide/). The IBM 2026 Cost of a Data Breach study found 92 percent of organizations that suffered an AI-related breach lacked proper AI access controls, which is usually a logging and permissions failure underneath. ### Examples across functions - **HR screening.** In Mobley v. Workday, a federal court in California granted preliminary certification of a nationwide age-discrimination collective in May 2025 over AI-driven applicant screening; the case is ongoing. In New York City, the State Comptroller's December 2025 audit of Local Law 144, which requires bias audits of automated hiring tools, found enforcement ineffective: the city agency identified one issue across 32 companies, where auditors found at least 17 potential violations. - **Customer service agents.** In Moffatt v. Air Canada, decided in February 2024, a British Columbia tribunal held the airline liable for its chatbot's incorrect advice on bereavement fares, rejecting the argument that the chatbot was a separate entity. The damages were small; the principle is that the company owns what its agent says. - **Procurement.** An agent that onboards suppliers or approves purchase orders inherits every control the human process had: vendor due diligence, approval limits, segregation of duties. Audit should test that the agent's permissions match the human role it replaced, not the broader permissions of the service account it runs on. - **Finance.** In banking, model risk management guidance changed in April 2026, when the Federal Reserve, OCC and FDIC issued SR 26-2, superseding the long-standing SR 11-7\. SR 26-2 explicitly leaves generative and agentic AI out of its scope, so language-model agents in a bank need governance of their own; conventional forecasting and credit models stay within model risk scope, and audit should confirm they are in the model inventory. ### The AI-assurance checklist | Control area | What to test | Evidence | | ---------------------- | ------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- | | Inventory | Completeness of the AI and agent register, including AI features in SaaS tools | Register reconciled to procurement records, SSO logs and network traffic to AI providers | | Ownership | Each system has a named business owner with documented accountability | RACI or policy record; owner sign-off on use case and risk rating | | Risk classification | Each system rated by impact on people, money and compliance | Risk assessment with criteria; mapping to EU AI Act or other applicable rules | | Data lineage | Sources, rights and personal data for training and grounding data | Data inventory, contracts and privacy assessments | | Access and permissions | Agent credentials scoped to the task; no shared or standing admin rights | Entitlement listings; comparison to the human role's access | | Change control | Re-validation after model, prompt or vendor updates | Change tickets, test results and vendor release notices | | Performance monitoring | Accuracy, drift and error rates tracked against thresholds | Monitoring dashboards and incident records | | Human oversight | Approval steps for high-impact actions work as designed | Sample of approvals and overrides; timestamps | | Logging | Agent actions reconstructable end to end | Log extracts for a sample of transactions | | Incident handling | AI failures reported, investigated and fixed | Incident register entries and post-incident reviews | | Value | Benefits claimed for the system are measured | Business case against actual results | The last row is not a traditional audit control, but audit committees increasingly ask it. The method for testing it is in the [CFO's guide to measuring AI return](https://fintekcafe.com/ai-roi-measurement-cfo/). ## The Frameworks Auditors Are Using | Framework | Issuer | Current version | Status | | ----------------------------------- | ------------------------------ | ----------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------- | | AI Auditing Framework | Institute of Internal Auditors | Updated September 2024; three domains: governance, management, internal audit | Voluntary guidance | | Global Internal Audit Standards | Institute of Internal Auditors | Issued January 2024, effective January 9, 2025 | Mandatory for IIA members; no AI-specific topical requirement issued | | AI Risk Management Framework | NIST | Version 1.0, January 2023; Generative AI Profile, July 2024 | Voluntary; under revision following the July 2025 US AI Action Plan | | ISO/IEC 42001 | ISO and IEC | Published December 2023; ISO/IEC 42006 for certification bodies, July 2025 | Certifiable; 271 accredited certificates in 40 countries at the end of 2025, per the ISO Survey | | Internal Control Over Generative AI | COSO | Released February 2026 | Voluntary; maps generative AI to the COSO internal control components | | EU AI Act | European Union | Digital Omnibus published July 2026 | Most high-risk obligations now apply from December 2, 2027; those for AI embedded in regulated products from August 2, 2028 | Two points for executives. First, there is no single mandatory AI audit standard; the frameworks overlap, and most organizations pick one as a backbone and map the others to it. The IIA framework suits audit planning; ISO/IEC 42001 suits organizations that want external certification; NIST suits US organizations with federal exposure. Second, the regulatory calendar keeps moving. Colorado repealed and replaced its AI Act in May 2026 with a narrower, disclosure-based law effective January 1, 2027\. Build the audit program on controls that make sense regardless of which rule lands, an approach the [AI governance policy guide](https://fintekcafe.com/ai-governance-policy-guide/) sets out from the policy side. ## What the Audit Committee Should Ask For Audit committees know AI matters and doubt they are overseeing it well. The Deloitte and Center for Audit Quality audit committee survey, released in September 2026 with close to 250 respondents, found 75 percent ranking AI governance a top-three priority. Among respondents whose audit committee holds primary responsibility for AI governance, 82 percent described that oversight as emerging or limited. Five requests close most of that gap: 1. **The inventory,** with risk ratings, reported at least twice a year, and a count of systems added since the last report. 2. **The top five AI risks by impact,** each with a named owner and the control that addresses it. 3. **Internal audit's coverage plan,** showing which AI systems were audited this year and which were not. 4. **Incidents and near misses** involving AI, including vendor model changes that altered behaviour. 5. **Audit's own use of AI,** with the controls over it, so the committee knows how findings were produced. Skills are the constraint. ISACA launched its Advanced in AI Audit certification in May 2025, open to holders of existing audit credentials, which gives audit leaders one benchmark for building capability. ## The Position, and Its Tradeoff **The recommended stance is assurance first, efficiency second.** Put the first serious investment into auditing the AI the business runs: inventory, ownership, change control, logging. Use AI in the audit itself where it extends coverage, such as full-population testing and continuous monitoring, and treat drafting as a convenience, not a strategy. The tradeoff is visible savings. Using AI to draft reports and automate testing produces a quick, reportable efficiency gain for the audit budget. Auditing agents and models produces findings, which create work for everyone else and rarely show up as savings. An audit leader under budget pressure will be tempted to lead with the efficiency story. The case for resisting is that the audit committee's exposure is not the cost of audit; it is an unmonitored agent making decisions about customers, candidates or money that nobody can reconstruct. ## Frequently Asked Questions ### How is AI used in internal audit? Audit teams use AI to test entire populations of transactions instead of samples, to monitor controls continuously, to detect anomalies in expenses, payments and access logs, and to draft workpapers and reports. Gartner polling in 2026 found drafting the most common use and audit testing far less common. ### How do you audit an AI system? Start with an inventory of AI systems and agents, then test ownership, risk classification, data lineage, access permissions, change control after model updates, performance monitoring, human oversight and logging. The evidence is the same kind auditors use elsewhere: registers, tickets, logs, approvals and test results. ### What framework should internal audit use for AI? The IIA's AI Auditing Framework, updated in September 2024, is designed for audit planning. NIST's AI Risk Management Framework and ISO/IEC 42001 are widely used for governance, and COSO published guidance on internal control over generative AI in February 2026\. None is a mandatory AI audit standard, so most functions choose one backbone and map the rest. ### Can AI replace audit sampling? For structured data, often yes: technology-assisted analysis can test every item in a population, and US public-company auditing standards now address it explicitly. The findings still need human-verified evidence, documented logic that can be re-performed, and thresholds tuned so that false positives do not overwhelm the team. ## The Bottom Line Internal audit's value has always been independence and evidence. AI tests both. It gives auditors the power to see every transaction, which is a genuine leap in coverage, and it fills the business with systems that make decisions faster than any annual audit cycle can follow. The executive decision is where to point the audit function first. Point it at the AI the business already runs, with an inventory the audit committee can read and controls that hold when a vendor changes the model. Then let AI make the audit itself broader and faster, with every finding still traceable to evidence a human has checked.