AI Use Cases in Internal Audit
An internal audit team can save twenty minutes drafting an email with artificial intelligence and still leave its operating model essentially unchanged. The larger opportunity appears where AI removes friction from work that consumes audit capacity: understanding large bodies of information, retrieving dispersed knowledge, finding exceptions, preserving traceability, coordinating evidence requests, reviewing quality, and turning fragmented signals into assurance decisions.
Adoption is already moving in that direction. A July 2026 Institute of Internal Auditors article reported that 83% of surveyed internal audit leaders expected their functions to increase AI use over the following year. The most frequently cited areas of current use included planning, reporting, risk assessment, and fieldwork. The important shift is therefore no longer whether AI can help an auditor. It is which parts of the work deserve to be redesigned, and under what controls.
The Global Internal Audit Standards do not prescribe an AI architecture. They do require risk-based planning, appropriate technological resources, sound methodology, supervision, documentation, quality, and effective communication. AI should therefore sit inside that professional system rather than become a shortcut around evidence, skepticism, or accountability.
The tables below combine three kinds of input: uses documented in IIA and Internal Audit Foundation research; professional activities already required or supported by the Standards; and design opportunities derived from current capabilities such as prompting, RAG, document extraction, analytics, and agents. Not every row represents a currently deployed market practice or a formal IIA recommendation. Treat the catalog as a set of candidates to validate against the function's data, risk profile, operating model, and technology environment.
AI should remove friction without removing accountability
A weak AI strategy starts with a tool and then searches for somewhere to use it. A stronger one starts with the work:
Where is there friction, repetition, difficult-to-find information, limited coverage, or avoidable variation in quality?
Only then should the function ask whether AI can change that point without handing over a decision that should remain professional.
The Internal Audit Foundation's study on GenAI across internal audit activities identified benefits across planning, fieldwork, reporting, and follow-up. More recent IIA material also describes quality-assurance applications, analysis of unstructured information, and agentic workflows that request, review, and route evidence. The opportunity is therefore much broader than writing assistance.
Complexity, however, is not maturity. A narrow RAG assistant grounded in approved methodology, with citations and version control, can be more mature than an autonomous agent connected to several systems with poorly defined boundaries. The useful question is not, “What is the most advanced use case?” It is “What creates measurable value at a level of autonomy we know how to govern?”
Six AI patterns auditors should distinguish
Product names will keep changing. Designing around more durable technology patterns makes the use-case portfolio easier to manage:
- Prompting. A model generates, transforms, summarizes, structures, or challenges information supplied in context. It is well suited to ideation, drafting, and assisted analysis when outputs are reviewed.
- RAG —retrieval-augmented generation. The system retrieves content from approved sources before generating an answer. It is useful when methodology, policies, standards, prior reports, or evidence need to be cited and kept current.
- Document AI. Models classify files and extract fields, dates, approvals, clauses, attributes, or entities from large document sets.
- Analytics, ML, and optimization. These find anomalies, patterns, forecasts, segments, or optimal combinations in structured data. Not every valuable AI use case needs an LLM.
- Agents and agentic workflows. The system can perform multiple steps, use tools, move information, or initiate actions. Risk rises materially when the system moves from recommending to acting.
- Hybrid solutions. These combine retrieval, generative models, rules, analytics, and automation. Many of the highest-value production use cases will eventually fall into this category.
The NIST AI RMF, ISO/IEC 42001:2023, and COSO's 2026 GenAI internal-control guidance are useful anchors for thinking about governance, risk, human oversight, monitoring, and controls. They do not replace internal audit methodology, and not every recommendation in them automatically becomes a requirement for an internal audit function.
Use cases across the engagement lifecycle
Planning: from the audit plan to individual engagement scope
Planning operates at two levels. At function level, AI can broaden the capacity to detect signals, maintain the audit universe, and build plan scenarios. At engagement level, it can reduce the work needed to understand an activity, retrieve history, identify criteria, and convert risks into an initial work program.
Department-level planning
| Use case | What it does | Value signal | Key guardrail |
|---|---|---|---|
| Risk & horizon scanner (RAG + agent) | Continuously scans approved external and internal sources for emerging risks, regulatory change, incidents, and strategic signals. | Faster identification of relevant change | Curated sources, freshness rules, auditor validation |
| Risk universe triage engine (RAG + analytics) | Groups and prioritizes risk signals before the formal risk assessment, highlighting themes that warrant auditor attention. | Time saved in risk synthesis; coverage of risk signals | AI proposes, CAE and auditors decide priorities |
| Audit universe completeness mapper (Hybrid) | Compares organizational structures, processes, systems, entities, projects, and prior plans to identify possible gaps in the audit universe. | Potential omissions identified and resolved | Treat matches as hypotheses; reconcile to authoritative master data |
| Annual audit plan scenario optimizer (Analytics + LLM) | Creates alternative plan scenarios using risk priority, mandatory coverage, capacity, skills, timing, and stakeholder requests. | Scenario cycle time; capacity utilization | No autonomous plan approval; assumptions must be visible |
| Stakeholder signal synthesizer (Prompting + document AI) | Synthesizes interviews, surveys, board requests, management concerns, and risk workshops into structured themes for planning. | Time saved; number of themes traced to sources | Retain source attribution and minority views |
| Assurance coverage mapper (RAG + analytics) | Links risks to internal audit, compliance, risk, external audit, and other assurance activity to identify duplication and gaps. | Coverage gaps and duplication identified | Reliance decisions remain professional judgments |
Engagement-level planning
| Use case | What it does | Value signal | Key guardrail |
|---|---|---|---|
| Engagement scoping assistant (RAG) | Uses objectives, process information, prior issues, current risks, and stakeholder requests to draft scope options and exclusions. | Planning cycle time; fewer late scope changes | Scope is approved by the engagement lead, not the model |
| Prior-audit and issue history retriever (RAG) | Finds relevant prior reports, workpapers, findings, action plans, and recurring themes for the activity under review. | Relevant history found; search time saved | Permissions, document versioning, and citation accuracy |
| Risk-control hypothesis generator (Prompting + RAG) | Generates initial risk, control, fraud, and failure-mode hypotheses to challenge during walkthroughs and planning interviews. | Breadth of hypotheses considered | Never treat generated risks or controls as established facts |
| Audit criteria navigator (RAG) | Retrieves relevant internal policies, regulations, standards, contracts, and control frameworks and maps them to potential objectives. | Time to identify criteria; citation coverage | Use authoritative and current sources only |
| Historical work program suggester (RAG) | Finds comparable prior work programs and proposes reusable procedures, highlighting where current risks require changes. | Reuse rate; planning time saved | Avoid copying obsolete procedures; require current-risk review |
| Evidence request builder (Prompting + RAG) | Drafts an initial PBC/evidence request list aligned to objectives, risks, controls, and planned procedures. | Fewer missing requests; fewer PBC cycles | Engagement lead approves necessity and proportionality |
The professional boundary matters. AI can propose a priority, scope option, or procedure; it should not become the authority that decides what risk deserves assurance coverage. Standard 9.4 places responsibility for the internal audit plan with the chief audit executive, while Standards 13.2 through 13.6 frame engagement risk assessment, objectives, scope, criteria, resources, and the work program.
Fieldwork: moving from document review to evidence at scale
The biggest fieldwork opportunity is not to “let AI perform the audit.” It is to expand the amount and variety of information an auditor can consider without losing traceability. The current CIA Part 2 syllabus already recognizes artificial intelligence, machine learning, automation, continuous monitoring, and analytics as technology options that can support findings and conclusions; the Global Internal Audit Standards continue to require relevant, reliable, and sufficient information and appropriate engagement documentation.
| Use case | What it does | Value signal | Key guardrail |
|---|---|---|---|
| Evidence PDF assistant (RAG) | Queries long reports, policies, contracts, SOC reports, and other evidence while preserving page-level source references. | Review time; citation accuracy | Do not infer beyond retrieved evidence; verify cited passages |
| PBC intake analyzer (Agent + document AI) | Checks incoming evidence against requests, identifies missing or unusable items, and prepares follow-up requests. | PBC cycle time; missing-item rate | Agent may request clarification, not accept evidence as sufficient |
| Document classifier and extractor (Document AI) | Classifies large evidence sets and extracts defined fields, dates, approvals, clauses, or control attributes. | Documents processed; extraction accuracy | Quality sample and exception review required |
| Evidence-to-criteria traceability checker (RAG + rules) | Links evidence items to audit criteria and planned procedures, flagging unsupported assertions or missing links. | Unsupported claims detected | Auditor decides sufficiency and relevance |
| Contract and policy comparator (RAG + document AI) | Compares agreements, policies, procedures, or versions to identify deviations, missing clauses, or inconsistent requirements. | Exceptions found; review time | Legal interpretation remains with qualified professionals |
| Interview and walkthrough synthesizer (Prompting) | Turns approved notes or transcripts into process steps, controls, open questions, contradictions, and follow-up items. | Documentation time; unresolved questions surfaced | Preserve speaker context and verify material statements |
| Transaction anomaly investigator (Analytics + LLM) | Combines analytics with natural-language explanations to prioritize unusual transactions and suggest investigation paths. | Exception yield; investigation time | Anomaly is not evidence of wrongdoing; investigate independently |
| Risk-based sample selector (Analytics / ML) | Prioritizes samples using risk factors, anomalies, value, frequency, control history, or defined sampling logic. | Coverage of high-risk population; reproducibility | Document sampling rationale and avoid hidden bias |
| Control test assistant (Hybrid) | Executes bounded checks such as matching approvals, dates, thresholds, required fields, or document attributes and summarizes exceptions. | Tests completed; exception precision | Predefined test logic, QA sample, and human conclusion |
| Cross-evidence contradiction finder (RAG + LLM) | Searches across interviews, policies, tickets, reports, and records for inconsistent statements, dates, ownership, or control descriptions. | Contradictions surfaced and resolved | Treat as leads; inspect original sources before concluding |
| Workpaper completeness checker (RAG + rules) | Checks whether workpapers contain required objectives, procedures, evidence references, reviewer notes, conclusions, and sign-offs. | Review notes avoided; completeness rate | Methodology rules must be current; reviewer remains accountable |
| Evidence collection workflow agent (Agent) | Orchestrates approved evidence requests, reminders, intake, naming, metadata, and routing to the right audit workstream. | Administrative hours saved; request turnaround | No uncontrolled access, deletion, or acceptance decisions |
Agentic use creates one additional boundary: an action is more consequential than a suggestion. An agent may organize requests or execute bounded tests, but allowing it to accept evidence as sufficient, alter workpapers without traceability, or close testing without review materially changes the control environment.
Reporting: the opportunity is not faster writing but more consistent reasoning
Drafting is one of the most visible GenAI applications, but a strong reporting use case should improve more than speed. AI can reinforce finding logic, check traceability, compare ratings, challenge action plans, and bring quality review closer to the moment the content is created.
| Use case | What it does | Value signal | Key guardrail |
|---|---|---|---|
| Finding draft optimizer (Prompting + RAG) | Turns validated facts into a structured first draft covering criteria, condition, cause, risk/consequence, and action. | Drafting time; edit distance | Only validated evidence may enter the prompt/context |
| Finding logic checker (RAG + rules) | Checks whether each finding is internally coherent and whether the conclusion is actually supported by the stated evidence. | Unsupported logic detected before review | Flag issues, do not rewrite facts to fit the conclusion |
| Root-cause hypothesis generator (Prompting) | Suggests plausible causal hypotheses and follow-up questions based on evidence patterns and control design. | Root-cause discussions improved | Hypotheses require corroboration with management and evidence |
| Evidence support checker (RAG) | Verifies that factual statements, numbers, dates, and conclusions in the draft can be traced to approved workpapers or sources. | Unsupported statements removed | Retrieval must respect workpaper permissions and versioning |
| Risk-rating consistency reviewer (RAG + analytics) | Compares proposed ratings with defined criteria and similar prior findings to surface possible inconsistency. | Rating consistency; reviewer overrides analyzed | AI does not assign the final rating |
| Management response analyzer (Prompting + RAG) | Assesses whether management responses address the stated risk, root cause, owner, timing, and evidence expectations. | Rework cycles; weak actions challenged | Auditor negotiates and approves agreed actions |
| Executive summary generator (Prompting + RAG) | Synthesizes validated engagement results into concise messages focused on significance, themes, exposure, and required decisions. | Drafting time; stakeholder comprehension | No new facts or stronger claims than underlying findings |
| Stakeholder-tailored communication assistant (Prompting) | Reframes the same approved conclusions for process owners, senior management, the board, or regulators without changing substance. | Fewer revisions; message clarity | Maintain one source of truth and approved terminology |
| Portfolio theme synthesizer (RAG + analytics) | Aggregates multiple engagements to identify recurring control failures, cross-business themes, systemic causes, or emerging risk signals. | Themes escalated; cross-audit insight | Avoid overgeneralizing from incomparable engagements |
| Report quality reviewer (RAG + agent) | Reviews drafts for accuracy, objectivity, clarity, concision, constructiveness, completeness, tone, and methodology requirements. | Review notes; cycle time; readability | Quality review supports, not replaces, supervisory review |
| Audit committee briefing builder (RAG + prompting) | Creates a board-level narrative from approved reports, themes, overdue actions, risk acceptance, and performance data. | Preparation time; completeness of key messages | CAE approves materiality and escalation judgments |
One rule should remain intact: AI cannot create evidence that does not exist. If a draft contains a statement the workpapers do not support, the answer is not to ask the model to phrase it more convincingly. The statement should be removed, additional evidence obtained, or the conclusion changed.
Follow-up: moving from chasing dates to testing whether risk changed
Follow-up is often full of administrative friction: reminders, evidence intake, status updates, and escalation preparation. That makes it attractive for automation, provided that confirmation of implementation and assessment of residual risk remain audit judgments.
| Use case | What it does | Value signal | Key guardrail |
|---|---|---|---|
| Remediation evidence validator (RAG + document AI) | Compares submitted closure evidence with the agreed action and expected proof, identifying missing or inconsistent support. | Closure review time; reopen rate | AI recommends; auditor confirms implementation |
| Action-plan sufficiency challenger (Prompting + RAG) | Evaluates whether a proposed action addresses the root cause and risk rather than only the symptom. | Weak actions revised before acceptance | Use cost-benefit and business context; human agreement required |
| Closure recommendation assistant (RAG) | Summarizes action status, evidence, remaining gaps, and residual risk to support close, extend, or escalate decisions. | Consistent closure decisions; time saved | No autonomous closure of findings |
| Overdue action prioritizer (Analytics + LLM) | Ranks overdue actions using risk rating, age, exposure, recurrence, dependency, and management responsiveness. | High-risk overdue actions addressed sooner | Transparent weighting and manual override |
| Follow-up request agent (Agent) | Sends approved reminders, requests evidence, classifies responses, and routes unresolved items to the appropriate auditor. | Administrative hours; response turnaround | Escalation thresholds and communication templates approved |
| Recurring-issue detector (RAG + analytics) | Finds repeated themes, control failures, root causes, business units, or owners across closed and open findings. | Recurrence identified earlier | Normalize taxonomy and verify comparability |
| Continuous indicator monitor (Analytics / ML) | Monitors selected control or risk indicators and produces alerts for conditions that warrant follow-up or new assurance work. | Alert precision; time to detect deterioration | Defined thresholds, ownership, false-positive review |
| Residual-risk reassessment assistant (Prompting + RAG) | Uses remediation evidence and current context to structure a reassessment of remaining exposure after management action. | Consistency of reassessment; time saved | Risk acceptance and residual-risk conclusions remain human |
| Escalation pack builder (RAG + prompting) | Assembles chronology, prior commitments, evidence, open risk, communications, and decision points for escalation to senior management or the board. | Preparation time; completeness | CAE determines whether and how to escalate |
| Management status classifier (Document AI) | Classifies free-text updates into standardized statuses such as on track, delayed, evidence incomplete, disputed, or ready for review. | Manual triage hours; classification accuracy | Do not infer implementation from management wording alone |
| Reopened-risk detector (RAG + analytics) | Looks for new incidents, complaints, control failures, or business changes that may invalidate a previously closed action. | Previously closed issues reconsidered when warranted | Use defined trigger criteria and audit judgment |
Standard 15.2 requires a process for confirming implementation of recommendations or action plans. Technology can make that process more continuous and risk focused, but “evidence received” should never automatically mean “action implemented.”
Use cases that strengthen the function, not only the engagement
Quality: IQA, EQA, supervision, and continuous improvement
Quality should not enter only at final report review. The Global Internal Audit Standards distribute quality responsibilities across internal quality assessment, performance measurement, and engagement supervision and improvement; Standard 8.4 also establishes the external quality assessment requirement. AI can help turn these activities into a more continuous, evidence-based system.
| Use case | What it does | Value signal | Key guardrail |
|---|---|---|---|
| QAIP knowledge assistant (RAG) | Answers methodology and quality questions from approved GIAS, QA manuals, internal methodology, and local procedures. | Search time; answer citation rate | Version-controlled sources and citation verification |
| IQA conformance mapper (RAG + rules) | Maps internal evidence to Standards requirements for periodic self-assessment and identifies missing support or ambiguous conclusions. | Assessment preparation time; evidence gaps | Qualified assessor makes conformance judgment |
| Ongoing quality monitor (Analytics + RAG) | Reviews completed engagements for recurring methodology deviations, late reviews, missing sign-offs, unsupported findings, or documentation gaps. | Defects detected; trend reduction | Use transparent rules and validate false positives |
| EQA readiness evidence index (RAG) | Creates a searchable evidence map for external quality assessment, linking standards, policies, workpapers, board communications, metrics, and prior IQA results. | EQA preparation hours; missing evidence identified | Do not self-certify conformance or assessor independence |
| EQA document pack organizer (Agent + RAG) | Collects, labels, indexes, and routes approved documents requested by external assessors while respecting access restrictions. | Admin hours; request turnaround | Access control and human approval before release |
| Engagement supervision assistant (RAG + rules) | Flags open review notes, inconsistent conclusions, incomplete procedures, overdue workpapers, and unresolved scope changes during the engagement. | Review cycle time; defects caught earlier | Supervisor remains responsible for coaching and approval |
| Methodology compliance checker (Rules + RAG) | Checks whether required templates, approvals, planning steps, testing documentation, and communications are present for the engagement type. | Methodology exceptions; rework avoided | Rules must reflect current methodology and approved exceptions |
| Cross-engagement consistency reviewer (RAG + analytics) | Compares similar engagements for inconsistencies in risk ratings, evidence expectations, issue wording, scope depth, or treatment of recurring themes. | Consistency issues identified | Context may justify differences; do not force uniformity |
| Performance metric analyst (Analytics + LLM) | Analyzes timeliness, coverage, cycle time, productivity, stakeholder feedback, plan delivery, quality defects, and value indicators. | Better KPI insight; time saved in monthly reporting | Avoid optimizing a single metric at the expense of audit quality |
| Stakeholder feedback synthesizer (Prompting + analytics) | Summarizes survey comments and interviews to identify recurring strengths, concerns, expectations, and opportunities for improvement. | Response themes; actionability of feedback | Protect anonymity and avoid sentiment overinterpretation |
| Report quality benchmarker (RAG) | Compares approved reports against internal quality criteria and a curated set of strong examples to support coaching and continuous improvement. | Quality scores; coaching topics | Do not reward style over evidence or substance |
| Nonconformance remediation tracker (Agent + RAG) | Tracks quality assessment findings, root causes, owners, action plans, evidence, due dates, and closure status. | On-time remediation; recurrence reduction | CAE and board retain required oversight |
A system can flag a possible conformance issue; the conformance judgment belongs to a competent assessor. That distinction is particularly important for IQA and EQA, where independence, evidence, and interpretation of requirements cannot be reduced to an automated score.
Operations: knowledge, capacity, and the machinery of the audit function
Standard 10.3 on technological resources makes a management question explicit: is the function using technology in ways that improve effectiveness and efficiency? That includes engagement delivery, but it also includes the operational infrastructure that allows the function to work consistently.
| Use case | What it does | Value signal | Key guardrail |
|---|---|---|---|
| Audit manual knowledge assistant (RAG) | Lets auditors query approved methodology, templates, definitions, examples, and procedures with citations. | Search time; adoption of standard methodology | Source permissions, versioning, and citations |
| AI prompt library (Prompting) | Provides reviewed prompt patterns for common audit tasks with examples, constraints, and required verification steps. | Reuse rate; time saved; output quality | Approval, ownership, version control, and retirement |
| Meeting action synthesizer (Prompting) | Turns approved meeting notes into decisions, action items, owners, dates, open questions, and follow-ups. | Minutes preparation time; action completeness | Human confirmation before distribution |
| Research and standards assistant (RAG + web agent) | Searches approved external and internal sources for regulations, standards, industry issues, and audit guidance relevant to a question. | Research time; citation quality | Whitelist sources, freshness checks, no unsourced claims |
| Learning coach for auditors (RAG + prompting) | Creates role-based learning paths, quizzes, practice scenarios, and explanations using approved training material. | Completion, assessment improvement, time to competency | Do not invent policy or technical guidance |
| Staffing and skills matcher (Analytics + LLM) | Matches engagement requirements with auditor skills, certifications, experience, availability, and development needs. | Skill fit; staffing cycle time | Avoid discriminatory proxies; manager approves assignments |
| Audit calendar optimizer (Optimization + LLM) | Schedules engagements using dependencies, business blackout periods, available staff, required specialists, and reporting milestones. | Schedule conflicts; rescheduling effort | Constraints and priority rules must be transparent |
| Budget and effort forecaster (Analytics / ML) | Uses historical engagement data and current scope assumptions to estimate hours, external support, travel, and technology needs. | Forecast accuracy; budget variance | Account for scope novelty and data-quality limitations |
| External provider due-diligence assistant (RAG + prompting) | Structures comparisons of co-source providers, specialists, tools, statements of work, competencies, conflicts, and security requirements. | Selection cycle time; completeness of assessment | Procurement, independence, legal, and security reviews remain required |
| Template and communication generator (Prompting + RAG) | Drafts standardized announcements, agendas, status updates, evidence requests, and internal communications from approved templates. | Administrative time saved; consistency | No external send without approval |
| Portfolio status narrator (Analytics + LLM) | Converts structured project data into concise weekly or monthly narratives on progress, slippage, resource pressure, and decisions needed. | Reporting time; issue visibility | Numbers must reconcile to source systems |
| Lessons-learned curator (RAG + agent) | Captures reusable lessons, procedures, prompts, pitfalls, and examples from completed engagements and routes them into governed knowledge bases. | Reuse; duplicate work avoided | Remove sensitive content and assign knowledge owners |
This category often contains the strongest first pilots because many outputs are reversible and easy to review. An audit-manual assistant, prompt library, or meeting action synthesizer can prove value without giving the system authority over assurance conclusions.
Governance: using AI creates a new control surface inside Internal Audit
Once Internal Audit adopts AI, it is no longer only an observer of enterprise AI governance. It becomes a user of the technology itself. That creates its own risks: confidential information in prompts, stale sources, excessive permissions, vendor dependence, agents with too much authority, poor traceability, or outputs that sound persuasive without being supported.
The IIA's Artificial Intelligence Auditing Framework, 2nd Edition focuses principally on understanding and assessing AI risks, governance, management, and controls in the organization. The same discipline is useful when designing governance for Internal Audit's own AI use.
| Use case | What it does | Value signal | Key guardrail |
|---|---|---|---|
| Internal Audit AI use-case inventory (Agent + RAG) | Maintains a catalog of AI-enabled audit use cases, owners, tools, data, risk tier, status, controls, and value measures. | Inventory completeness; review timeliness | Discovery must include shadow or locally created use cases |
| AI use-case risk tiering (Rules + LLM) | Scores proposed internal-audit AI uses using data sensitivity, decision consequence, autonomy, external exposure, tool access, and reversibility. | Time to approve; consistency of risk classification | Risk criteria approved by governance; override documented |
| Approved-tool and data-use assistant (RAG) | Answers whether a tool and intended data type are permitted for a proposed audit task, citing policy and exceptions. | Fewer policy breaches; decision time | Policy source must be current; ambiguous cases escalate |
| Prompt governance and versioning (RAG + workflow) | Stores approved prompts, owners, versions, test results, permitted contexts, and retirement dates for repeatable use. | Reuse; failure rate; review cadence | Change control and testing before material updates |
| Output-validation gate (Rules + agent) | Requires defined verification steps before AI-generated analysis, summaries, or drafts can enter workpapers or reports. | Validation completion; defects caught | Gate cannot be bypassed for high-risk outputs |
| Model and vendor change monitor (RAG + agent) | Tracks material changes to approved models, enterprise assistants, APIs, features, terms, retention settings, and sub-processors. | Changes assessed before use; time to reassess | Use authoritative vendor and enterprise change sources |
| Access and permission reviewer (Analytics + rules) | Reviews which users, agents, connectors, repositories, and business systems an AI use case can reach and flags excessive access. | Excess access removed; review completion | Least privilege, segregation, periodic recertification |
| Human-approval orchestration (Agent workflow) | Places mandatory approval checkpoints before sensitive actions such as sending requests, changing records, closing issues, or publishing reports. | Unauthorized actions prevented; approval latency | Approval authority must be real, informed, and logged |
| AI activity and traceability monitor (Analytics + agent) | Checks logs for prompts, retrieved sources, tool calls, outputs, approvals, errors, and exceptions for governed use cases. | Traceability coverage; exceptions investigated | Logs protected from alteration and access restricted |
| AI incident intake and classifier (Document AI + agent) | Captures and categorizes hallucinations, data leakage, wrong actions, access failures, policy breaches, bias concerns, and near misses. | Incident capture rate; resolution time | Material incidents require human escalation and root-cause review |
| AI performance and quality monitor (Analytics / ML) | Tracks accuracy, citation quality, exception rate, user overrides, failure modes, latency, and cost for productionized audit use cases. | Quality trend; failure reduction; cost per task | Metrics must match the use case and risk, not generic model scores |
| AI value and ROI tracker (Analytics + LLM) | Measures time saved, coverage expanded, cycle-time reduction, quality improvement, user adoption, and incremental cost by use case. | Realized benefit vs. expected benefit | Do not equate hours saved with value if quality or coverage falls |
| Agent segregation-of-duties checker (Rules + graph analytics) | Detects when one agent or service can request, receive, test, approve, and close the same item without sufficient separation. | Conflicting permissions identified | Independent approval for high-impact actions |
| Sensitive-data prompt scanner (Classification + rules) | Detects or blocks restricted information before it is sent to a model or external AI service, based on approved data-handling rules. | Blocked policy breaches; false-positive rate | Complement, not substitute, user training and approved environments |
Governance should be proportional. A function does not need a separate committee for every prompt, but it does need clear rules on approved tools, data, permitted decisions, traceability, human review, change, and accountability.
How to choose the first five use cases
A large catalog becomes valuable only when it leads to choices. A practical prioritization model can score every candidate on five dimensions:
- Current friction. Does the task consume material time, create rework, or constrain assurance coverage?
- Source quality. Is the information sufficiently complete, current, and accessible for the system to work reliably?
- Consequence of error. Is an error easy to detect and reverse, or could it affect an audit conclusion, board communication, or system action?
- Measurable value. Can the function compare time, coverage, quality, cycle time, defects, or user satisfaction before and after?
- Governability. Are there approved tools, permissions, logs, owners, validation rules, and a way to stop the use case if it fails?
For a function at the beginning of the journey, a sensible sequence might be audit manual RAG assistant → prompt library → scoping assistant → evidence PDF assistant → report quality reviewer. These use cases have relatively controllable sources, reviewable outputs, and clear value metrics. More integrated applications can follow: agentic PBC intake, automated testing, continuous monitoring, and follow-up workflows.
The most sensitive cases should wait until governance is stronger. Examples include approving the audit plan, assigning final finding ratings, closing management actions, publishing reports, or allowing an agent to write to corporate systems without human approval.
IIA discussions in 2026 describe an evolution from prompting toward agentic capabilities and the possibility of bringing quality controls forward into the workflow rather than applying them only at the end. That direction is promising precisely when the function preserves traceability, criteria, source control, and professional review.
An internal audit function is not more mature because its AI does more. It is more mature when it knows what to delegate, what to verify, and what it should never stop deciding.
The strategic opportunity is larger than saving hours. If AI removes searching, classification, documentation, and coordination work, the capacity released can become broader coverage, deeper analysis, more stakeholder interaction, and more timely follow-up. That value appears only when technology extends auditor judgment instead of obscuring it.
Sources
- The Institute of Internal Auditors, Global Internal Audit Standards
- The Institute of Internal Auditors, Artificial Intelligence Auditing Framework, 2nd Edition
- The IIA, Internal Audit Upskilling for Critical AI Capabilities (2026)
- Internal Audit Foundation, Solving the Riddle: Harnessing Generative AI for Internal Audit Activities (2024)
- Internal Audit Foundation, Demystifying AI: Internal Audit Use Cases for Applying New Technology (2024)
- NIST, Artificial Intelligence Risk Management Framework
- ISO, ISO/IEC 42001:2023 — AI management systems
- COSO, Achieving Effective Internal Control Over Generative AI (2026)