Much like transactions from any automated process. Auditors do not trust or distrust the agent itself; they test the controls around it: what it was allowed to do, how that was enforced, who approved its output, and whether a complete record shows every transaction it touched. Where those controls hold, agent-prepared transactions are tested like any others.
What changes for the auditor, and what does not?
The audit questions stay the same. Is the transaction real? Is it recorded at the right amount, in the right period, in the right account? Was it authorised? What changes is where the evidence comes from. When a person prepares a journal, the auditor looks at the support and the approval. When an agent prepares it, the auditor also needs to understand the agent: its scope, its limits, and how anyone would know if it went wrong.
That makes the agent more like a system than a member of staff. Auditors will treat its configuration, access and change history much as they treat an automated control in an accounting system.
What will auditors ask for?
- Scope and limits. What the agent is permitted to do, at what depth, for which processes and values.
- Enforcement. Whether those limits are enforced by the systems it acts on, or only by its own instructions.
- Change control. Who can change the agent's brief, sources and limits, and the history of changes in the period.
- Certification. Evidence that the agent was tested before it was given authority, and when that was renewed.
- Approvals. For anything the agent drafted, who approved it, with the approver independent of the request. This is the subject of who approves payments prepared by an AI agent.
- The verdict record. A complete, unaltered log of what the agent did, what it held and what it was refused. Why this matters is set out in auditors will ask for the verdict record.
How do auditors test it in practice?
| Agent role | Typical audit approach |
|---|---|
| Drafts; a person approves and posts | Test the person's approval as a manual control, with the agent's draft as support |
| Posts routine entries within limits | Test the limits and their enforcement once, then test a sample of entries |
| Matches or reconciles | Re-perform a sample of matches and review unmatched items |
| Flags exceptions | Test that flags were raised and resolved, and check for missed cases |
Where the agent posts at depth 3, auditors usually want comfort over the whole period, not just the sample. That comes from showing the limits did not change, or changed only with approval, and that the record is complete.
An illustrative example
Take a retailer whose finance agent codes and posts card transactions at depth 3 and drafts all other journals. During the year-end audit, the auditors ask for the agent's scope, its posting limit, the history of changes to both, and the full record of postings.
The finance team provides them. The auditors notice that the posting limit was raised mid-year. They ask for the approval and find it was signed by the financial controller with a written reason, after a certification run on the retailer's own transactions. They test a sample of 40 card postings and re-perform the coding, finding one miscoded item, which had already been caught in the monthly sample review. They then test the approval of drafted journals as a normal manual control.
The audit file notes the agent as part of the control environment. No additional findings arise. The approach to reconciling corporate card transactions gives the detail of how such a process is set up.
What stays with a person?
Everything that involves accountability to the auditor. Management representations are signed by directors. Explanations of judgements come from the people who made them. The agent can gather evidence, as covered in whether an AI agent can prepare the audit pack for external auditors, but answers to the auditors' questions about why something was done come from people.
Where Pryme fits
In Pryme, every verdict an agent reaches, allowed, held, refused or withheld, is filed in an append-only record with the bound that decided it, the certificate in force and the sources used. That gives auditors a complete population to test rather than a reconstruction. The wider finance controls are in AI agents for finance: books you can audit.
Questions people ask
Do auditors need to understand how the model works?
Generally not in technical depth. They need to understand what the agent may do, how that is limited and how its output is checked. The approval structure is usually set out in an approval matrix for agent-prepared payments or its equivalent.
Will auditors test more if we use agents?
Possibly at first, while they understand the set-up. Clear scope, enforced limits and a complete record tend to reduce testing in later years.
Can the auditors rely on the agent's own checks?
Only if those checks are tested like any other control. An untested check is information, not evidence.