Should the report say that AI was used?

The working papers must. The report itself should say so where the organisation's policy or a regulator expects it; Deloitte's revised report added exactly that line. Either way, the record behind the report should show which sentences came from a model and what the auditor changed.

Will every finding start to sound the same?

Only if nobody edits. A draft from test results is a structure to fill, not a voice to adopt. The cure is the one that always worked: the reviewer reads the observation as the owner will, and sends back anything that reads like a template.

Can AI rate an observation?

It can explain a rating that rules produced from the risk and control scores on the work paper. It should not pick the rating from the text, because the text is the part that was written to persuade. Rules first, model explains, auditor decides.

What if the model writes something wrong and the auditor misses it?

The same thing that happens when a junior writes something wrong and the senior misses it: the review failed, and the record shows where. The difference is that with the suggestion and the decision logged, you can measure how often it happens and tighten the review where it does.

Can AI write audit observations and reports?

AI can draft an observation and it cannot find one. Given the test results, the control as designed and the evidence, a model will write a usable first version of the condition, the criteria and the recommendation, and a first guess at the consequence. It cannot tell you the cause, because the cause comes from a conversation the model was not in. It can draft an executive summary from the findings the report lists, one sentence per claim, each pointing at a finding. It cannot sign anything, and the auditor who accepts the draft owns every word of it. That is the whole answer; the rest is how to make it hold.

What a model can draft, part by part

The structure of an observation is in writing audit observations people act on: criteria, condition, cause, consequence. Here is how a model does against each part when it is given the work paper and not the internet.

Condition. Good. "In 4 of 12 payments sampled, the approver was the requester" is a sentence a model writes well from a test case that says 4 failed of 12, with the sample and the note. It is also the sentence an auditor writes slowest, because it is boring and it has to be exact.

Criteria. Good, when the desired control is on the work paper or the RCM row. The model quotes it. When the criteria are a policy clause, the auditor still has to find the clause.

Cause. Poor, and it should be. A model given test results will produce "the control was not consistently performed", which is the condition wearing a hat. The real cause ("the approval limit was never updated after the reorganisation") is in the owner's head, and the auditor gets it by asking. Treat a drafted cause as a placeholder that the conversation replaces.

Consequence. Partial. The model can name the category of risk. The number, the regulatory exposure and what already happened are the auditor's.

Recommendation. Good as a draft, if the prompt asks for a change someone could finish ("remove the two lapsed approvers and update the matrix"), not a wish ("strengthen the control environment").

The management response is a separate case. The owner should write it, in their words, and a draft for them to edit is reasonable, as long as the auditor is not the one asking for it. An auditor who drafts both sides has written a conversation with themselves. What happens to the response afterwards is in tracking management action plans.

Reports: one claim, one citation

An executive summary is where a model's fluency is most dangerous, because it is read by the people with the least time and the most authority. The rule that makes it safe is mechanical: every sentence in the summary cites a finding the report lists, and a sentence that cites nothing is removed before a person sees it. The model is not asked to be interesting. It is asked to be traceable.

The case for that rule was made in public in October 2025, when Deloitte refunded part of an A$440,000 fee to the Australian government after its 237-page report was found to contain a quotation from a court judgment that did not exist and references to academic papers that were never written. The revised version added a disclosure that generative AI had been used. The substance of the report was reportedly unchanged. The credibility was not.

An audit report has a stricter test than a consulting report, because Standard 14.6 requires documentation that lets a competent person repeat the work and reach the same results, and Standard 15.1 makes the final communication the function's own. A summary sentence whose only source is a model fails both.

The record a reviewer needs

Whoever reviews the report needs four things in the file for each piece of drafted text:

The suggestion as the model produced it.
The decision: accepted, edited or rejected, by whom and when.
The difference: how much of the suggestion survived, measured rather than remembered.
The provenance: which model, which prompt version, which records the model was shown.

With those, "did the AI write this finding?" has an answer in seconds. Without them, the honest answer is "we think the auditor edited it", which is not an answer the audit committee should accept.

Two things follow. A report's sign-off must be tied to the text that was signed: if a summary section changes after sign-off, the sign-off is void until it is done again. And if the prompt behind a section changes after the report was published, the file should say so, because a reader re-running the process would get different words.

What to ask for in a prompt, and what not to paste

People searching for "ChatGPT internal audit report prompt" are usually asking the right question in the wrong tool. A good drafting request contains the test results with the sample and the failures, the control as designed, the risk the control addresses, the rating scale with definitions, and the house structure. It does not contain names, email addresses, the evidence files, or anything from an investigation. If the only way to give a model those inputs is to paste them into a public chatbot, the answer is to not do it; see whether ChatGPT is safe for internal audit work. Inside a system that holds the work paper, the model can be shown the test results and nothing else, which is both safer and a better prompt.

How AuditEase Pro does it

In AuditEase Pro, "Draft from test results" on an observation drafts the condition and criteria, the cause, the consequence and the recommendation, one field at a time, from the ticked test cases (their result, sample and note) and the work paper's risk, control and desired control. Each drafted field goes through accept, edit or reject, and nothing is saved by itself. Treat the drafted cause as the starting point for the conversation, not a substitute for it. Business owners can ask for a draft of their own response to an observation, in their turn, and the audit team cannot ask for it on their behalf. A rating suggestion comes from rules over the work paper's scores, with the model explaining the result, and with no scores there is no suggestion.

On the report, "Draft from findings" writes the executive summary and the opinion from the observations the final report lists, keeps only sentences that cite one of them, drops the rest and lists what it dropped. The report records the runs behind each section, a reproduce check reports a changed prompt as drift, and accepting a drafted section saves through the same step as any wording change, so the head of Internal Audit's existing sign-off no longer matches and must be given again. Every run sits in the ledger with the model, the prompt version, the tokens and the decision, and every decision is an audit event with the measured difference between the suggestion and what was saved. The post on how audit teams use AI and where it fails sets out the rules this follows.

All of it is off until an administrator names a provider, which can be a model you host. The live demo has it on, against a model Hikmah hosts, with fictional audits, test results and findings to draft from, so you can see the accept step and the citations rather than read about them.