Reading time: about 8 minutes
AI-supported internal audit software is good at work that is repetitive and checkable: preparing and reconciling data, sampling, spotting anomalies, continuous monitoring of controls, and a first draft of working papers and the report. What must not be handed to the machine: professional judgment, risk assessment, the audit conclusion and opinion, and independence and objectivity, because a human owns those. There is a third boundary that is easy to forget: audit data, findings, control weaknesses and whistleblower reports must not go into a public AI model. Below we lay out what to automate, what not to, and where that data has to stay, with a decision table you can apply to your own audit function in a couple of minutes.
The starting point is simple: AI does not change what the auditor is accountable for, only how much time goes into the mechanical parts. The new Global Internal Audit Standards, which became effective on 9 January 2025, keep professional judgment, objectivity and accountability at the centre, and no tool takes those over. Meanwhile adoption is rising fast: a Wolters Kluwer and Internal Audit Foundation survey found that 39% of audit teams already use AI and another 41% plan to within a year. The question is no longer whether, but what and where.
What "AI internal audit software" actually means
It is not one product but a layer of AI added to the tools audit already has: working papers, GRC systems, spreadsheets of data and the repository of prior audits. In practice AI enters in two very different roles that must not be confused.
The first role is AI as the auditor's tool, speeding up the auditor's own work: reading documentation, collating data, drafting a first version of a finding. The second is AI as the subject of the audit, meaning AI systems used across the company that audit has to assess for risk, controls and compliance. This piece is about the first role, supporting the auditor's work, because that is where the question "what can we automate" usually lands. The second is a separate topic, where the boundaries are set by the AI Act and internal model-assessment frameworks, among others.
The distinction matters, because a tool that helps an auditor write cannot at the same time assess its own work. When those roles blur, independence disappears, and independence is the foundation the whole audit function stands on.
What you can genuinely automate
AI delivers the most value where the work is tedious and repetitive and the output can be verified. Five areas recur in almost every audit function.
- Data preparation and reconciliation. Pulling extracts from several systems, bringing them to a common format, matching items, flagging differences. Those are hours of spreadsheet work that AI cuts to minutes, and the auditor checks the result anyway.
- Full-population testing instead of a sample. A classic audit tests a sample because you cannot go through everything by hand. AI-supported analytics lets you run 100% of transactions and surface the outliers, so the sample stops being a constraint.
- Anomaly detection. Unusual entries, duplicates, payments just under an approval threshold, accounts changed right before a transfer. The model catches patterns the eye misses across thousands of rows, but the auditor decides whether an anomaly is a finding.
- A first draft of working papers and the report. This is the most common use today: in Gartner's 2026 survey, 60% of teams use AI to draft issues, ratings and reports. AI turns notes into a coherent draft, the auditor verifies it, edits it and owns it.
- Continuous monitoring and knowledge. Instead of a once-a-year audit, steady watch over selected controls and an alert when something drifts. Plus search across policies, procedures and prior reports so the auditor finds context faster.
The common denominator is this: AI prepares the material, the human judges it. Automation covers gathering and first-pass processing, not the conclusion. That distinction leads straight to the second list.
What you must not automate
Some activities are off-limits not because the technology is immature, but because delegating them breaks the essence of audit. Four of them have to stay with the human.
- Professional judgment and skepticism. Deciding whether something is material, whether management's explanation is credible, whether an anomaly is an error or fraud, is judgment grounded in context the model does not have. AI suggests, the human decides.
- Risk assessment and the audit plan. What to focus on and what to leave out is a decision about allocating limited resources against the company's risks. AI can feed data into that decision, but it cannot make it for the audit committee.
- The conclusion, opinion and rating. The final audit conclusion, including a control rating, is a statement the auditor signs and the board relies on. A generative model can write a convincing sentence that is false, so the conclusion must not be delegated.
- Independence, objectivity and accountability. An auditor cannot assess a control they designed themselves, and likewise AI cannot be both author and reviewer of a finding. The signature on the report belongs to a human.
The same Gartner report shows the scale of the tension: while 93% of audit leaders report some AI use, only 38% have a formal strategy for it, and just 12% apply it to quality assurance reviews. In other words the tools arrived faster than the rules that say where not to use them. That is a governance gap, not a technology one.
Decision table: automate, hybrid or human only
The quickest way to draw the line is to line up a concrete audit task against the approach that fits it and the reason behind it.
| Audit task | Approach | Why |
|---|---|---|
| Pull and reconcile data across systems | Automate | Mechanical work, output easy to verify |
| Test the full population and flag outliers | Automate | AI goes wider than a sample, human judges the result |
| Draft a finding and a first report version | Hybrid | AI writes version zero, the auditor verifies and owns it |
| Classify an anomaly as a finding | Hybrid | Model shortlists, materiality judgment is the human's |
| Risk assessment and annual audit plan | Human only | A decision to allocate resources against company risk |
| Conclusion, opinion and control rating | Human only | A signed statement, plus model hallucination risk |
| Safeguarding independence and objectivity | Human only | Author and reviewer cannot be the same tool |
The pattern is clear: the closer to the data and the processing, the more you can hand to the machine; the closer to judgment and the signature, the more firmly it stays with the human. Hybrid is not a compromise, it is the default mode of AI-supported audit.
The closer to judgment and the signature, the less you can automate.
Where audit data has to stay
The third boundary matters as much as the split of tasks, and it is skipped far more often. The data audit works on is among the most sensitive in the company: findings, maps of control weaknesses, financial documentation, and sometimes whistleblower reports. It is exactly the knowledge that is most dangerous in the wrong hands.
If internal audit software ships that material to a public AI model, the company loses control of it: it cannot be sure how long it is kept, who on the vendor's side can see it, or which jurisdiction it sits in. For an entity under NIS2 that is not a detail, because the AI vendor then enters the supply chain being assessed, and the auditor asks plainly where data is processed and who can reach it. We laid this out more fully in the piece on where the real risk of sensitive data in AI sits, and in the guide to NIS2 registration by 3 October 2026.
The conclusion is not "do not use AI in audit" but "process audit data where it stays under your control". That is the core of private AI: the model runs on your infrastructure or in a dedicated, isolated instance, not on a shared, public API. Which variant is right depends on data sensitivity and regulatory status, which we compare in the guide to single-tenant versus shared cloud.
How to roll out AI in audit step by step
An order that lets you capture the automation without losing control of judgment and data.
- Start with one process. Pick a tedious, repetitive task, for example data reconciliation or full-population testing, and prove the value there before you widen the scope.
- Set the judgment boundary. Write down what AI may prepare and what stays a human decision. This is your internal version of the split in the table above.
- Settle where the data goes. Before you feed audit documentation into a tool, establish whether the data leaves an environment you control. If it does, that is the first problem to solve.
- Build verification into the process. Every AI output should pass through auditor review, and the trace of that review should stay in the working papers. Without it you cannot defend the result to an external auditor.
- Write the rules before you scale. A simple policy on when and for what AI may be used closes the gap Gartner sees in most teams: tools without a strategy.
The worst case is letting AI into the working papers without settling where the data goes and who owns the result. If you want to see which side of the line your uses fall on, our readiness mini-audit takes 10 minutes and leaves no data behind. The wider picture of when a private environment is worth it at all is in our piece on private AI for manufacturing.
Frequently asked questions
Will AI replace the internal auditor?
No. AI takes over the mechanical work: gathering data, sampling, anomaly detection, drafting the report. Professional judgment, risk assessment, the conclusion and the signature stay with the auditor, as the Global Internal Audit Standards effective January 2025 confirm.
What can you safely automate in internal audit?
Data preparation and reconciliation, full-population testing, anomaly detection, a first draft of working papers and the report, and continuous monitoring of controls. The shared trait: AI prepares the material, the human verifies it.
What must not be handed to AI in audit?
Risk assessment and the audit plan, the final conclusion and control rating, and safeguarding independence and objectivity. These carry the auditor's personal accountability, and a generative model can write a convincing but false sentence.
Can you paste audit documentation into ChatGPT?
Not into a public model. Audit data is among the most sensitive in the company, and sent to a public API it leaves your control and becomes a problem in the NIS2 supply chain. The safe route is a private or isolated instance where the data stays with you.
Is AI internal audit software compliant with NIS2?
It depends on the deployment model, not the feature itself. NIS2 requires you to control and document where data is processed and to assess your vendors. A tool that keeps audit data in an environment under your control makes that answer simpler.
Fryderyk, CortexMine. We write about private AI for NIS2-covered manufacturers, based on our own deployments and tests.
Prefer to talk it through? Book a 30-minute call with the founder, no pitch, just your case.