Once an agent touches the books, auditors will ask for its verdict record: what the agent did, under which bounds and certificate, relying on which facts, and who approved it. Activity logs and model transcripts will not answer those questions. Design the record before the agent drafts its first journal, because it cannot be rebuilt afterwards.
What will auditors ask?
External auditors test whether the financial statements are fairly stated and whether the controls that produce them work. When an agent is part of the process, their questions follow the same logic, applied to a new kind of actor:
- Which entries, reconciliations or payments did the agent prepare or execute?
- What was the agent permitted to do at the time, and was it within those permissions?
- Was the agent's authority based on evidence, and was that evidence current?
- What information did it rely on, and where did that information come from?
- Who reviewed or approved its work, and when?
- Were any actions refused or held, and what happened to them?
- Can the record be relied on, or could it have been altered?
Each question needs an answer for a sample of transactions chosen by the auditor, not by the firm. That means the record has to cover every action, not a curated selection.
Why aren't ordinary logs enough?
System logs record events: a user logged in, an API was called, a record changed. Model transcripts record what the agent said. Neither records a decision in the form an auditor needs.
An auditor looking at a journal prepared by an agent wants to know why it was allowed. An event log might show that the agent's service account posted an entry at a certain time. It will not show what depth the agent held, whether its certificate was valid, which limit applied, or that a named controller approved the entry. Piecing that together from several systems, after the fact, is slow, error-prone and hard to defend.
A verdict is different. It is the decision made at the moment of action about whether the agent may proceed, recorded with the reasons. Allowed, held, refused or withheld, each verdict names the bounds that were checked and the one that decided. That is the unit an auditor can test.
What should each verdict entry hold?
| Field | Why the auditor needs it |
|---|---|
| The action attempted | Identifies the entry, reconciliation or payment |
| Agent and capability | Which agent, doing what |
| Depth and certificate status at that moment | Shows the agent's authority was current |
| Each bound checked and the deciding bound | Shows why the action was allowed, held or refused |
| Facts relied on, with provenance | Shows the basis for the figures and whether sources were reliable |
| Person who approved or rejected a held action | Shows human review where required |
| Timestamp and integrity check | Shows sequence and that the entry has not been changed |
Provenance deserves emphasis. An accrual based on a sealed source, such as a purchase order in the procurement system, rests on firmer ground than one based on an external source, such as an emailed estimate. Auditors already weigh evidence by its source; the record should let them do so for agent work.
How does the record support controls testing?
Auditors test controls by sampling. For a manual journal review control, they might pick a set of journals and check that each was approved by someone other than the preparer. With an agent in the process, the same test needs the verdict record: for each sampled journal, did the agent prepare it, was it held, and who released it?
A complete record makes this efficient. The auditor can sample from the full population and get a clear answer for each item. An incomplete one forces the auditor to widen testing, rely on other evidence or report a control deficiency. For a firm adopting agents in finance, a well-kept verdict record is one of the most direct ways to keep audit costs and findings under control.
It also helps the auditor understand the record-to-report process as a whole. If agents prepare a significant share of entries, the auditor needs to understand where agent work enters the process and where people check it, and the record shows that.
Who prepares the audit pack?
An agent can do much of the assembly. It can pull the verdicts for a sample, attach the supporting documents, compile reconciliations and format everything as the auditor requests. Whether an AI agent can prepare the audit pack for external auditors is a fair question, and the answer is yes for preparation.
A named person in finance should review the pack before it goes out and answer the auditor's follow-up questions. When an investigation is needed, such as a query about a specific transaction, the record should be exportable intact, as described in how to export an AI agent's verdict record for an investigation.
Does this change the auditor's role or the accountant's?
It changes the evidence, not the accountability. The firm's directors are still responsible for the financial statements; the finance team still owns the controls; the auditor still forms an opinion. What changes is that some of the work being audited was done by agents, and the evidence of how it was controlled lives in a new kind of record.
That is also the realistic answer to whether AI agents will replace accountants. Agents can prepare a growing share of the work. People remain the approvers, the owners of judgement and the ones who answer to auditors. The verdict record is how those people show their oversight was real.
What should finance teams do now?
Before any agent works on the ledger, agree the record format with the people who will rely on it: the financial controller, internal audit and, ideally, the external auditor. Confirm that every action, including refused and held ones, produces a verdict. Make sure the record is append-only and exportable. Then test it: pick ten agent-prepared entries and see whether someone unfamiliar with the setup can answer all seven auditor questions from the record alone.
Where Pryme fits
In Pryme, every action an agent attempts produces a verdict filed in an append-only record, naming the bounds checked, the deciding bound, the certificate in force, the facts relied on with their provenance class, and any person who released a held action. Records can be exported for audit or investigation with their integrity intact. This supports the aim of AI agents for finance: books you can audit.
Questions people ask
How long should verdict records be kept?
At least as long as the accounting records they relate to, under the retention rules that apply to the firm. Many firms align them with their existing financial record retention schedule.
Do auditors need access to the agent itself?
Usually not. They need the record of what it did and the evidence of its authority. Access to replay certification results may help for some tests.
What if an auditor finds an error in an agent-prepared entry?
Treat it as any other audit finding: correct it, find the cause and assess whether other entries are affected. If the cause was the agent, it should trigger a review of its certification.