Skip to content
Implementa.
InfrastructurePlaybook··8 min

An AI usage register for internal audit: what you'll be asked for and don't have yet

The AI usage register internal audit (or your external auditor) will ask for isn't your agent's log or your list of tools. It's the list of consequential processes an AI touches, plus proof that someone reviewed what came out. The thesis: auditors don't audit the AI, they audit the control around it. The five columns, what you can skip, and what to do before your next close.

Senior AI Infrastructure Implementer

AI Infrastructure Pod

An illustrative scenario. Quarter-end close at a mid-sized company. The financial controller has spent months using her spreadsheet's built-in AI assistant to reconcile provisions and draft the notes for the management pack. It works, it saves hours, and nobody has banned it. At the audit planning meeting, the audit team asks a new question: “Where in the close does an AI tool get involved, and how do you review what it produces?” Silence. Not because anyone's hiding anything, but because nobody has ever written it down.

The thesis in one sentence: the register your auditor will ask for isn't a technical log or a license inventory. It's the list of consequential processes an AI touches, the human control covering each one, and proof that the control actually ran. Auditors don't audit the AI. They audit the control around it. And if the control isn't written down, as far as the auditor is concerned it doesn't exist.

What an AI usage register for internal audit is (and why it isn't your log)

An AI usage register for internal audit is a living document, one row per use, that answers three questions: where an AI is involved in a process that ends in a number, a legal obligation or a decision about a person; who owns what comes out; and what evidence shows someone reviewed it before it took effect. It doesn't store conversations or prompts. It stores control facts.

It gets confused with three neighbors, and the confusion is expensive because each one answers a different question. The agent inventory tells you which systems exist. The technical log tells you what each system did and how long you need to keep it. Decision traceability lets you rebuild one specific decision end to end. The usage register sits on top: it's the layer an auditor reads in ten minutes to decide where to dig.

Why you'll be asked for it now, not in five years

Because the standard that sends your financial auditor into your systems already exists, and it doesn't need to name AI. ISA 315 (Revised 2019), the IAASB's international standard on risk assessment, requires the auditor to understand the company's IT environment and the IT applications relevant to financial reporting. An assistant that proposes a provision is one of those applications, whether you bought it as “AI” or it arrived quietly inside your ERP.

In the US, the regulator of public-company auditors is already looking. The PCAOB published its staff observations on generative AI in July 2024: at companies, using it in accounting and financial reporting lagged behind using it in operations and customer-facing work, and audit firms stressed the need for strong supervision of its use. “Lagging behind” is a countdown, not a reassurance.

Auditors themselves are adopting the tools, so they know the failure modes. In Spain, for instance, a report backed by the national audit regulator (ICAC) and the accounting academics' association ASEPUC, reported in June 2025, found 61.5% of auditors already use AI in their daily work, while only 45% had received specific training. And on the internal audit side, IIA Spain published guidance in May 2025 on auditing AI applied to business processes, including a sample audit work program. Once there's a work program, there are questions at the next audit.

The five columns your register needs

You don't need a tool. You need a table with five columns, one row per use, not per tool. The same assistant used in the close and in marketing is two rows with two different risk profiles.

ColumnWhat you write downWhy the auditor looks at it
Process and outputThe process and the number, document or decision it produces: “bad-debt provision”, “draft note to the financial statements”, “claims prioritization”Tells them whether the use touches financial reporting, a legal obligation or a person
Tool and modeWhich tool, whether the AI is embedded in other software or standalone, and whether it suggests or executesAI that executes with no human step is a different risk level from AI that suggests
OwnerThe person who owns the process, not whoever installed the toolNo owner, no one to ask for the evidence
Control and evidenceWhat human review happens, when, and where it's recorded: sign-off, ticket, reconciliation, second reviewThis is the column that actually gets audited
ChangesGo-live date and every relevant change: new version, new use, new input dataToday's control doesn't prove last quarter's control

The column that matters is the fourth. “Maria reviews it” isn't a control. “Maria compares the proposed provision with the aged receivables by bucket and signs off the reconciliation in the system before the journal entry” is, because it leaves a trail an outsider can check. If you want the detail on designing that step, it's in the guide on human-in-the-loop automation.

What you don't need to keep (even if someone's selling it)

The reflex when you hear “audit” is to keep everything: every prompt, every answer, every conversation. It's expensive, it piles up personal data for no reason, and it doesn't answer the question that will actually be asked. The auditor won't read ten thousand conversations. They'll sample outputs and ask for the control evidence behind each one. What you need to be able to show is the output, the review, and the gap between what the AI proposed and what was approved.

That gap is gold, and almost nobody keeps it. If the AI proposed a provision of $48,000 and $52,000 was booked, the note explaining the correction proves the control worked better than any report could. A register where the AI is always right and nobody ever corrects anything doesn't signal reliability. It signals that nobody's looking. It's the same distinction that separates a performance dashboard from auditable evidence.

Where the uses that aren't in the register sneak in

The missing uses usually aren't AI projects. They're features that showed up on their own: the assistant switched on in the office suite, the “smart suggestions” module in the ERP after an update, the browser extension someone installed to summarize contracts. The Institute of Internal Auditors, in its piece on AI governance, talks about a living inventory of AI systems, models and embedded tools. The word that matters is “embedded.”

Two questions bring them to the surface. To every owner of a process with financial consequences: “which part of this month's work did a tool start and you finish?” And to IT: “which AI features were switched on by default in the last year's updates?” Banning doesn't work; as we've argued before, shadow AI is your best list of use cases. Writing it down does.

When you DON'T need this register

If no AI use in your company touches a number that ends up in the accounts, a legal obligation or a decision about a person, a formal register is bureaucracy. A team using AI for marketing drafts that a human always reviews before publishing doesn't need five columns; it needs a sensible usage policy. You also don't need it if you have no audit, internal or external, and no large customer requires one by contract. But watch the word “none”: it's the usual answer before anyone asks the two questions above.

What to do before your next close

  1. Ask the two questions to the owners of close, treasury, billing and payroll, and to IT. Write each answer down as a row, even if it's incomplete.
  2. Fill in the control column only with what leaves a trail today. Where the control is “I look at it,” write exactly that: it's your to-do list, not a failure to hide.
  3. For each use that touches numbers, decide the minimum evidence: who signs, where, and what you keep of the gap between proposed and approved. The guide on governance and control of automation works as a template.
  4. Schedule a quarterly review of the register, tied to the close, with the changes column as item one. And if a row belongs to an agent that already executes in accounting, its technical sheet goes in the documentation of your automations.

If you already run an AI accounting agent or are about to build one, this register isn't separate paperwork: it's page one of its design. An agent designed knowing what the auditor will ask gets audited in an afternoon. One designed without that in mind gets audited over a quarter, and sometimes it doesn't get audited at all: it gets switched off.

The line for your next meeting: “Nobody's going to ask whether we use AI. They'll ask where, who answers for it, and how we know. Let's have the table before they ask for it.”

This article describes control practices and auditing standards in general terms and is not accounting or legal advice. The figures on auditors' AI use refer to Spain; the PCAOB observations refer to the United States.

Shall we get it shipping?

If this resonated, 30-minute conversation with no commitment. We tell you what fits, what doesn't and the approximate price.

See cases
An AI usage register for internal audit: what you'll be asked for and don't have yet · Implementa