How to prepare the year-end audit request list with an AI agent

Auditors send a list of what they need; the agent turns it into a plan. It maps every request to a source and owner, drafts the schedules it can build from the ledger, gathers documents, checks each item ties to the trial balance and tracks status. Named owners review every item, and a person sends everything to the auditors.

What is the audit request list, and who owns the response?

Before the year-end audit, external auditors issue a request list, often called a "prepared by client" or PBC list. It can run to hundreds of items: reconciliations, schedules, contracts, board minutes, confirmations, samples of invoices, payroll reports and access listings. Each has a due date. Late or wrong items extend the audit and raise fees.

The financial controller usually owns the response, with a named owner for each item. Owners prepare or find their items and someone in finance checks them before release. Most of the effort goes into finding things, rebuilding schedules in the auditor's format and chasing colleagues. The finance bounds for this work are in AI agents for finance. Much of what is asked for is the output of each month-end close, which is why a well-run close shortens the audit.

What are the steps, and at what depth?

1. Agent, depth 1: parses the request list. Each item becomes a task with a description, due date and the auditor's reference.
2. Agent, depth 2: proposes owner and source. Using the previous audit's list and responses, it suggests who owns each item and where it came from. The controller confirms.
3. Agent, depth 2: drafts schedules from the ledger. Account analyses, roll-forwards, accrual listings, fixed asset movements, aged debtors and creditors, built in the format requested. Many reuse workpapers already produced, such as the accruals from running month-end accruals with an AI agent.
4. Agent, depth 1: gathers documents. Contracts, minutes, policies and statements from the document store, by request.
5. Agent, depth 1: checks each item. Totals tie to the final trial balance; documents are the signed, final versions; the period is right; personal data is limited to what was asked.
6. Person, item owner: reviews and approves. Confirms the item is complete and correct.
7. Person, controller or delegate: sends to the auditors. Through the agreed channel.
8. Agent, depth 1: tracks status. Requested, in progress, approved, sent, queried, closed. Sends daily status to the controller and reminders to owners.
9. Agent, depth 2: drafts responses to queries. The owner reviews and sends.

Which controls matter, and what evidence does each leave?

Tie to the final trial balance. Every schedule agrees to the trial balance version the audit is based on. Evidence: tie-out per item with trial balance version.

Final, signed documents only. Drafts of contracts or minutes are not sent. Evidence: document version and signature status per item.

Data minimisation. Payroll and customer items include only the fields and records requested. Evidence: the check result per item and the owner's approval.

A person sends. Evidence: the send log with the sender's identity, time and channel.

Complete trail. For each item: request, owner, source, draft, checks, approval, send, queries. Evidence: the task record.

What evidence does the controller have at the end?

A single index: every request, who approved it, what was sent, when, and any queries with their answers. Next year, that index is the agent's starting point for owners and sources. It also answers questions from the audit committee about how the audit went and where the delays were.

If the agents themselves are in audit scope, as they may be where they prepare entries, auditors will ask how they are controlled. Preparing AI agents for an audit covers the evidence they expect: role scopes, certification results and the verdict record.

What does one item look like from request to close?

Take an illustrative request: "Analysis of other creditors and accruals at year end, with support for items over 25,000." The agent proposes the payables lead as owner, builds the analysis from the ledger, ties its total to the trial balance, and attaches invoices or contracts for the eleven items above 25,000. Two items have no support on file; it lists them for the owner. The owner finds one contract and reverses the other accrual, which was stale. The analysis is re-tied, approved and sent by the controller.

What should you measure?

• Items delivered by due date.
• Items returned by auditors as incomplete or wrong.
• Queries per item, by area.
• Audit adjustments, by area. Areas with many adjustments need better close work, not better audit preparation.
• Hours spent by finance on the audit, compared with the previous audit.

Where does it usually go wrong?

Schedules from the wrong trial balance. A late adjustment changes the trial balance after schedules were built. The agent re-ties every schedule when the trial balance version changes.

Drafts sent as finals. A draft contract or unsigned minute goes to the auditors and creates questions. Version checks stop this.

Too much data. A payroll report with every employee's full record is sent when the auditor asked for a sample of five. The agent should cut to the request, and the owner confirms.

Chasing replaces fixing. The agent's reminders make lateness visible, but some items are late because the underlying work was never done. Treat persistently late items as close weaknesses.

Whether an agent can prepare the audit pack for external auditors in full follows the same split: the agent assembles, owners approve, a person sends.