It can prepare them; a named person must answer them. An AI agent can read an auditor's request, pull the ledger entries, invoices, approvals and reconciliations that answer it, and draft a response with the evidence attached. What goes to the auditor is management's statement, so a named person in finance reviews, corrects where needed and sends it.
Why can't the agent reply to the auditor directly?
Because an answer to an auditor is a representation by management. It may later be relied on in the audit opinion, and someone must stand behind it. An agent can be accurate about what the ledger says, but it cannot take responsibility for whether the ledger is right, whether an estimate was reasonable, or whether a disclosure is complete.
There is also a separation reason. The audit tests the controls of the team that prepared the numbers. If the same agent that posts journals also answers questions about them without review, the auditor is effectively asking the preparer to vouch for itself.
What conditions decide how far the agent can go?
The kind of query. Requests for documents and facts, such as "provide the invoice and approval for this payment", are well suited to an agent. Requests for explanations involving judgement, such as "why was this provision reduced", need the person who made the judgement.
The source of the evidence. The agent should draw only from systems of record: the general ledger, the bank feed, the approval log, the contract repository. A spreadsheet someone emailed is not evidence the agent should attach without a person confirming it.
Completeness. If the agent cannot find something the auditor asked for, it must say so in the draft rather than send a partial set that looks complete.
What may the agent prepare, and what stays with a named approver?
| Query type | The agent prepares | The named approver |
|---|---|---|
| Sample requests (invoices, receipts, approvals) | Pulls each item, links it to the ledger entry, notes anything missing | Reviews the pack and sends it |
| Reconciliations | Produces the reconciliation from source data, flags unreconciled items | Explains reconciling items, signs off |
| Walkthrough questions on a process | Drafts a description from documented procedures and logs | Confirms it reflects practice |
| Questions about estimates and judgements | Gathers the supporting data | Writes the explanation |
| Follow-ups on errors found | Pulls the transactions and history | Responds, and decides on any adjustment |
| Questions about the agent itself | Supplies its certificate, limits and verdict log | Answers for how the agent is controlled |
The last row is increasingly common. Auditors want to understand automated controls, including agents, and how they are governed. What they are likely to ask for is covered in how to show an auditor that an agent is certified and, for EU deployments, EU AI Act documentation an auditor will ask for about your AI agents.
An illustrative example
Take a mid-sized manufacturer at year end. The external auditor sends a list of 40 sampled payments and asks for the invoice, purchase order, goods received note and approval for each.
The agent pulls the documents from the purchase-to-pay system and links each to its ledger entry. For 37 payments everything matches. For two, the goods received note is missing. For one, the approval came from someone above their approval limit, which the approval log shows as an exception approved by the finance director.
The agent drafts the response: 37 complete packs, two missing documents flagged, one exception explained with the log entry attached. The financial controller reads the draft, chases the warehouse for the two notes, adds a sentence on the exception, and sends it. The auditor gets a complete, honest answer in a fraction of the usual time, and the controller has stood behind it.
Which controls make it safe?
- Draft depth only. The agent cannot send anything to the auditor. Every response is held for a named approver.
- Sealed sources. Evidence comes from systems of record; anything else is marked for the approver to confirm.
- Gaps stated. Missing items are listed, never silently left out.
- A query log. Each request, the evidence pulled, the draft, the approver's changes and what was sent are filed, so the audit trail of the audit itself is clear.
Supplier queries follow a similar pattern with lower stakes; see whether an AI agent can answer supplier payment queries and how to answer supplier payment status queries. The step-by-step process for audit requests is in how to answer auditor queries with an AI agent.
Where Pryme fits
In Pryme, evidence reaches the agent stamped with its provenance: ledger and bank data as sealed, emailed documents as external. Sending to an auditor is not a capability a finance agent holds; drafts are held for the named approver, and every verdict is filed. More is in AI agents for finance: books you can audit.
Questions people ask
Can the auditor query the agent directly?
Some firms give auditors read access to evidence. Answers that carry management's view should still come from a person.
Will auditors accept evidence the agent gathered?
They assess evidence by its source and integrity, not by who fetched it. Evidence drawn from systems of record with a clear trail is usually what they want.
Does using an agent change what the auditor tests?
It may add work: the auditor may test the agent's controls as part of the control environment. Have its certificate, limits and logs ready.