Will AI replace internal auditors?

Not the judgement. It is already replacing some of the typing, the searching and the first pass over a data file. The IIA's Vision 2035 survey found most practitioners expect new technology to change entry-level work, and most also see a risk of relying on it with too little human involvement. Both are true at once.

Do we need a policy before anyone on the team uses AI?

Yes, and it fits on one page: which tools are allowed, what may never be pasted into them, who decides what a draft becomes, and how the use is recorded. Without it, people use consumer chatbots quietly, which is the worst of the options.

What should a small team start with?

Something low-risk with a visible payoff: search across past findings, a data test on a ledger extract, or a draft management response that the owner edits. Leave the executive summary and anything that reads evidence until the record-keeping is in place.

Does the IIA allow AI in audit work?

The Global Internal Audit Standards do not name AI, but Standard 10.3 makes the chief audit executive responsible for the technology the function uses, and Standard 14.6 requires documentation that another auditor could repeat. AI is allowed where those two still hold.

How do internal audit teams use AI today, and where does it fail?

Most internal audit teams that use AI use it for four things: drafting text an auditor will edit, finding things in their own audit history, running first-pass tests on data files, and preparing planning material. It fails in three predictable ways: it states facts it was never given, it reads records a person should not see, and its words get accepted without anyone deciding. The teams that get value from it are the ones that have written down who decides, what is recorded and what stays out.

That is the short answer. The rest of this post is the detail, with the numbers behind it.

Where the profession actually is

The IIA's Internal Audit: Vision 2035 research (July 2024, more than 7,000 practitioners and stakeholders) found that 93% expect new technology to produce better insights for recommendations, while fewer than half (48%) were involved in any AI-related activity. The Internal Audit Foundation's follow-up, Demystifying AI, reads the same survey more bluntly: 77% reported low or no implementation of AI in their function, only 18% reported using AI for internal audit activities, and 78% agreed there is a risk that organisations will rely too heavily on new technology with too little human involvement.

So the typical function is early, interested and wary. That is the right place to be, as long as it does not stay there by default.

What AI is good at in audit work

Drafting text a person will edit. The first version of a finding's condition and criteria from test results, a management response for the owner to rewrite in their own words, closing-meeting minutes from the notes, a plain-English summary of a long policy. The model is fast at the blank page. The auditor is better at the final page. The post on drafting observations and reports with AI goes through it part by part.

Searching your own history by meaning. "Have we raised this before?" is a question every team asks and few can answer from a folder structure. Embedding-based search finds the finding from three years ago that used different words for the same control gap, and a repeat-finding hint while an observation is being typed is often the most useful thing AI does in an audit system.

First-pass tests on data. Benford's law on a payments extract, exact and near duplicates, outliers, weekend postings, gaps in sequences. None of this needs a language model: it is classic statistics, it is deterministic, and it runs where the data is. It produces a list to look at, not a finding.

Classification and hints. A suggested rating with the reasons, from the risk and control scores that already sit on the work paper. A filter built from a sentence ("open issues in Finance due this quarter"). Small, bounded, easy to check.

Preparing planning material. Draft questionnaire questions for an entity, candidate RCM rows from a process description, test steps for a control, a first engagement memo from the risk scores, the RCM and prior findings. All of it as proposals a person ticks or discards. The post on AI in the risk assessment and the annual plan goes through this in detail.

Where it fails

It states facts it was never given. A language model completes text; it does not check it. In 2023 a New York court sanctioned two lawyers for filing a brief with six invented cases that ChatGPT had produced. In October 2025 Deloitte refunded part of an A$440,000 fee to the Australian government after a report was found to contain a fabricated court quotation and references to papers that do not exist. Neither was a failure of intelligence. Both were a report that nobody traced back to its sources before it went out. An audit report has exactly one defence against this: every sentence that asserts something points at a work paper, and a sentence that points at nothing is cut.

It reads what a person may not. An assistant that searches "everything" searches the draft observations management has not yet seen, the evidence from the investigation two people know about, and the HR audit. If the scope check is applied after retrieval, a well-phrased question leaks the lot. Scope has to be decided before the model sees anything: start from the records this user may open, then search those.

It does not know the cause. The part of an observation that decides whether anything changes is why the gap exists, and that comes from a conversation, not a document. A model will happily write "the control was not performed consistently", which restates the condition and produces a training session as the action plan. See writing observations people act on.

It leaves no trail unless you build one. If a draft arrives in a text box and the auditor saves it, a year later nobody can say which words the model wrote and which the auditor did. Standard 14.6 of the Global Internal Audit Standards requires engagement documentation such that "an informed, prudent internal auditor" could repeat the work and reach the same results. That needs the suggestion, the decision and the difference between them on the record.

It leaks when it is convenient. The fastest way to use AI is to paste the evidence into a consumer chatbot. That is also the way audit material ends up in a provider's training data or somebody else's chat history. The post on whether ChatGPT is safe for audit work covers what to keep out and the contained options.

The rules that keep it defensible

The Standards do not mention AI by name. They do not need to. Standard 10.3 says the chief audit executive "must regularly evaluate the technology used by the internal audit function" and must collaborate with IT and information security to implement it properly. Standard 14.6 covers the documentation. Between them they set the rules any AI feature has to meet:

A person decides. A suggestion fills a field or a draft. It never changes a status, signs off a report or closes an issue.
Everything is on the record. Who asked, which model and prompt version, what it suggested, what the person kept, and how different the two are.
Scope first. Retrieval begins from what the person may see.
Personal data stays out. Names and email addresses never go to a model that does not need them; most drafting does not.
Off until someone turns it on. Nothing leaves the deployment unless an administrator has named a provider and knows what is sent to it.
Capped. A budget and a per-person limit, so an experiment cannot become an invoice.

A team that can show these six has an answer for the audit committee's first question, which will be "how do we know the AI did not write the finding?" If the organisation's own AI use is the bigger question, how to audit an AI system covers that side.

How AuditEase Pro does this

AuditEase Pro is an internal audit system with optional AI built on those six rules, and it is worth being exact about what that means today.

The AI is off in every deployment until an administrator names a provider: Amazon Bedrock, Anthropic, Azure OpenAI, or any OpenAI-compatible server such as Ollama or vLLM, which is how a customer keeps the text on their own network. Every suggestion goes through an accept, edit or reject step, and a test in the codebase checks that no AI code can call the workflow, the sign-off or the publish step. Each run is a ledger row with the kind, the prompt version, the provider and model, the tokens, the cost, the user and the record, and each decision is an audit event with the measured difference between the suggestion and what was saved. Search and retrieval start from the records the user may open. Prompt builders never read a person field, and email addresses and users' full names are replaced before any text leaves. There is a monthly budget and a daily cap per person.

What it drafts: the parts of a finding from ticked test results, management responses, ratings with reasons, questionnaire questions, RCM rows, test steps, closing-meeting minutes, a QA pre-check and an executive summary that keeps only sentences citing a listed finding. What it finds: search by meaning, repeat findings, and answers to questions about the audit history with citations. What runs with no model at all: data tests on uploaded files (Benford, duplicates, outliers, dates) and an overdue-issue prediction that an administrator has to turn on, because it profiles people.

The live demo runs with the AI features switched on, against an open-weight model hosted on Hikmah's own infrastructure, so the accept step, the citations and the ledger can be seen rather than described.