If an agent is processing supplier invoices, your auditors will want to know how you know it is right. The evidence they will look for: who owns it, what it can do, who approved its output, how accuracy is measured and what the logs show. How to have it ready before they ask.
Auditors don't need to understand how a language model works. They need to understand the control environment around it, the same way they do for any process that affects the accounts. If an agent creates purchase invoices, the auditor's question is the familiar one: how do you know they're right?
What they'll ask
What AI is in use in financial processes? A list: which agents, which processes, since when.
Who owns each one? A named person accountable for its output (see who owns AI).
What can it do? Its permission set. Can it post, or only draft? Can it create suppliers or change bank details? (See permissions.)
Who approves its output? Evidence that a person reviewed and posted what the agent prepared, with segregation of duties intact.
How do you know it's accurate? Monthly measures: error rate, override rate, sample checks (see the five numbers).
What does the log show? Agent activity, approvals and configuration changes, retained for the audit period.
What happens when it goes wrong? An incident process, and examples of it being used.
How are changes controlled? Who can change what the agent does, and is that tested and recorded?
What about fraud? How a convincing fake invoice or a changed bank detail is caught (see prompt injection).
The evidence to keep ready
An AI register: each agent, owner, process, permissions, go-live date.
The approval trail in Business Central: who posted each agent-prepared document. Posted documents can't be edited, and Find Entries traces everything.
Agent activity logs and MCP call logs, and configuration audit through Microsoft Purview.
Change log entries for setup changes, including vendor bank accounts (change log).
Monthly accuracy reports, signed off by the owner.
The AI policy and evidence staff have seen it.
The common findings
Agent running under a broad permission set, or SUPER.
No evidence of review; "the agent posts and someone glances at it".
No one can say what the error rate is.
Configuration changed without a record.
No named owner.
None of these are AI problems. They're control problems, and they're easy to fix before the audit rather than after it.
Email this post as a PDF
A branded PDF of this post, sent to your inbox to keep, print or share. Copy in colleagues if you like.
We only use these addresses to send this PDF. Nobody is added to a mailing list. Mike at Business Boxed receives a copy. See our privacy notice.
