The quality test for a workpaper is simple to state. Another auditor opens it without speaking to whoever prepared it, repeats the work and ends up in the same place.

That is almost word for word Standard 14.6 of the Global Internal Audit Standards. The analyses, evaluations and supporting information must be documented such that an informed, prudent internal auditor, or similarly informed and competent person, could repeat the work and derive the same engagement results. The standard adds that the engagement supervisor reviews the documentation and that the chief audit executive reviews and approves it.

With small samples and manual tests, that standard was met almost by inertia. The sample sat in a spreadsheet, the criterion in the work program and the evidence in a folder. When the test is run by a script, an analytics tool or an AI model, what has to be kept changes.

What AI adds to the problem

Risk in Focus 2026, the survey by the European institutes of internal auditors of 879 chief audit executives in fifteen countries, Spain among them, records the warning of a CAE who was implementing AI into his own processes. As well as being properly documented, audit tests done with generative AI needed to be assessed to see whether the same input prompts created the same results and whether there was a human in the loop to review the main outcomes.

Internal auditors should see generative AI as a super-smart intern that should be constantly under review. That's because it looks smarter than it actually is and, as an intern, you have to recheck whatever it produces.

That is how the CAE put it, quoted without a name in the report.

What has to be written down

Standard 14.6 already lists the essentials in its considerations for implementation. The date or period of the engagement, the work program, the description of analyses with procedures and data sources, the population evaluated and the names of those who performed and supervised the work. With automation or AI, that translates into seven concrete pieces.

  1. The criterion, written before running and with the version applied.
  2. The population, with its origin, the number of records and how its completeness was checked.
  3. The cut-off date and the period it covers.
  4. The version of the work program and, if a tool runs the test, the version of the rule, the script or the instructions used.
  5. The evidence behind each exception, linked to the source record and not summarised.
  6. Who reviewed, what they reviewed and when. If a model prepared the output, the name of the person who accepted it.
  7. What could not be concluded. Missing records, systems that could not be extracted, unexplained exceptions. Writing it down protects the auditor and whoever reads the report.

Repeating is not getting the same text

A language model does not always return the same text from the same instructions. That is why what gets documented cannot be just the output. It has to be enough for another auditor to reach the same conclusion from the same evidence, even if the wording changes. If the conclusion depends on an answer nobody can reconstruct, the test is not finished.

There is a simple way to check this before closing the engagement. Someone on the team who did not take part picks up the workpaper, reruns the test on the same population with the same criterion and compares. If they reach the same result, the workpaper does its job. If they need to ask something, the answer to that question is what the workpaper is missing.

It is the same standard worth demanding from any tool before signing. The first of the eight criteria for evaluating an AI tool for audit is exactly that, every conclusion with its source.