One question keeps surfacing in practitioner discussions on Reddit: what evidence has actually satisfied an auditor when AI sits inside a control? We follow those conversations to see where teams get stuck, and this piece adds an answer in the form of a list an auditor can test.

By the end, readers will have five types of evidence to prepare, a way to agree them with the auditor before fieldwork starts and a worked evidence list for one AI-assisted control.

Audit evidence for AI-driven processes is the set of records that lets an auditor reconstruct what a system decided, under which configuration and at what point a person took over. When a person approves an invoice, the auditor tests the person's decision. When a system approves it, the auditor tests the records behind the decision.

What an auditor can test when a system decides instead of a person

An auditor cannot interview a model about why it approved a payment. The auditor can only test records. Three points about those records carry most of the weight: what each decision log contains, how the log is kept and what follows when it does not exist.

The log of each decision

A decision log is a record, created at the time of the decision, of what the system received, what it decided and what happened next. Each entry needs a timestamp, a unique case reference, the inputs the system used, the outcome and the reason or score behind it. It also needs the identifiers of the model and prompt version in force. Without those identifiers, the auditor cannot tell which approved configuration produced a given decision. Where the system calls other tools or systems, those calls belong in the same record.

Retention and tamper-evidence

A log is only evidence if it survives and cannot be changed unnoticed. Retention is the period the records are kept, and it must cover the whole period under audit plus the time until the audit is complete. Tamper-evident means that any alteration or deletion would be detectable. Write-once storage, hashed log batches and restricted write access with a reviewable access list all achieve this. Evidence of who can edit or delete the log matters as much as the log itself.

Missing logging is a finding

If decisions are not logged, or the logs cannot be queried, the auditor has nothing to test, and the absence is itself a finding. Raise it early, before fieldwork, because logging that starts today cannot evidence last quarter.

For a practitioner view of the logs AI agents need, read ISACA's article on auditing agentic AI workflows. It is a contributor article and not ISACA guidance. It names the questions to settle at the outset: what log data is available, what retention period applies and whether the store is tamper-evident.

Configuration, model updates and escalation: the records that show the control held

Logs show what happened. Four further sets of records show whether the system was running as approved and whether a person stepped in when needed.

  • Approved prompt or model configuration: the prompt, thresholds and settings in production, held under version control with a dated approval from a named owner. The auditor compares the live configuration with the approved version.
  • Third-party model update records: the vendor's notices of model changes during the period, the date each took effect and the results of a regression test. A regression test reruns a fixed set of cases to confirm that behaviour has not shifted.
  • Escalation evidence: documented triggers for handing a case to a person, test cases showing the handover happens when a trigger is met and log entries that pair each escalation with the reviewer's decision.
  • Gaps and bypasses: any period when the AI was switched off, overridden or run outside the approved configuration, with the reason and the manual control used instead.

Escalation records need the most care. An auditor will look for cases that should have escalated but did not, so the control owner should run that query before the auditor does.

How to agree the evidence list with the auditor before the audit starts

The evidence list should be settled before fieldwork, not during it. Four steps work in sequence. First, describe the control in one sentence, including where the AI acts and where a person takes over. Second, list the records that would prove the control operated. Third, send the list to the auditor and ask what they would test and in what format they need the records. Fourth, run a trial extract for one week and check that the fields, version identifiers and timestamps are present.

Agreeing the list early also settles who holds the evidence. If the vendor holds the logs, the control owner needs a contractual or technical route to extract them on request. Once agreed, the list becomes the evidence request and the test steps. Ideagen's internal audit solution shows how audit evidence and control testing sit together inside an audit platform.

A worked example: the evidence list for one AI-assisted control

Take a UK finance team where an AI model approves supplier invoices of up to £5,000 that match a purchase order and routes everything else to a person. The control objective is that no invoice above the threshold, or without a matching purchase order, is paid without human approval. The evidence list is as follows.

  • Decision log extract: one record per invoice, with timestamp, invoice reference, purchase order reference, amount, outcome (approved or escalated), reason code, model identifier and prompt version.
  • Log integrity and retention: the retention rule, evidence that the log store is write-once or hashed and the list of users with edit or delete rights, reviewed in the period.
  • Approved configuration: the prompt and the £5,000 threshold as held in version control, the dated approval from the finance controller and a comparison showing production matched the approved version throughout the period.
  • Model update records: the vendor's update notices for the period, the date each took effect and regression results on a fixed set of test invoices, with the reviewer's sign-off.
  • Escalation evidence: the written triggers (over threshold, no purchase order, possible duplicate), test invoices that meet each trigger and show escalation, and a query listing any invoice over £5,000 approved without an escalation record.
  • Reviewer records: for each escalated invoice, the reviewer's name, decision and time.
  • Bypass record: any time the AI was switched off or overridden, with the manual process that replaced it.

Each item gives the auditor a record to open and a question to answer.

Four evidence problems to check for before the auditor does

These patterns recur in AI-assisted controls, and each has a direct fix.

What the team says Evidence that answers it What the control owner gains
The vendor dashboard shows activity but nothing can be exported Export decision records to a store the control owner holds, with a fixed list of fields The auditor can test the control without waiting on a vendor request
Nobody can say which prompt was live in a given month Version the prompt and configuration, with an approval record for every change The exact configuration can be shown for any date in the period
A model update changed behaviour and nobody noticed Log vendor update notices and rerun a fixed test set after each one The team finds the change before the auditor does
The process escalates but nothing records the handover Log the trigger, the handover and the reviewer's decision together Human oversight is shown with records rather than recollection

Start with the log and agree the list before fieldwork

An auditor can test a system that decides only through the records it leaves behind. The decision log, its retention and tamper-evidence, the approved configuration, the vendor update records and the escalation trail form that set, and missing logging is the gap to raise first. Agree the list early, trial the extract and the audit has far less to find.

To continue the discussion with other practitioners, join Ideagen's Reddit community.

Explore internal audit solutions

Get more value, more audits and more flexible workflows from your internal audit software.