How to review manual journals before they are posted with an AI agent

Manual journals are where most accounting errors and many frauds enter the ledger, and approvers rarely have time to look closely. An agent can read every parked journal before approval: test its support, account combinations, amounts, timing, description and preparer rights, and give the approver a short risk note. The approver still decides and posts. The agent cannot approve.

What is manual journal review, and who owns it?

Most ledger entries come from subledgers: sales, purchases, payroll, fixed assets. Manual journals are the rest: accruals, reclassifications, corrections, allocations, top-side adjustments. They are typed by a person, often at period end under time pressure, and they can go to any account. That flexibility is why they are needed and why auditors test them hard.

The financial controller owns the journal policy: who may prepare, who may approve, what support is needed and above what value a second approver is required. Approvers are usually senior accountants or the controller. The agent sits between preparer and approver, as a reviewer that reads everything. The wider argument for this order is in finance agents should draft journals before they post them; here the journals are drafted by people.

What are the steps, and at what depth?

1. Agent, depth 1: picks up each parked journal. As soon as it is submitted for approval.
2. Agent, depth 1: checks support. Is a document attached; does it support the amount and accounts; is it final, not a draft; does its date fit the period.
3. Agent, depth 1: checks accounts. Unusual combinations, such as revenue debited against a balance sheet account, postings to suspense, equity, intercompany or rarely used accounts, and accounts the preparer does not normally touch.
4. Agent, depth 1: checks amounts. Round amounts, amounts just under an approval threshold, and unusual size for the account.
5. Agent, depth 1: checks timing. Posting dates at or after period end, entries made outside working hours, back-dated entries, and accruals with no reversal date.
6. Agent, depth 1: checks rights. The preparer is allowed to prepare to these accounts; the approver is allowed to approve this value; preparer and approver are different people.
7. Agent, depth 2: writes the risk note. Clean, or a short list of findings with evidence, at the top of the approval screen.
8. Person, approver: decides. Approves, rejects or sends back, with a reason where the note had findings.
9. Person, controller: reviews overrides. Journals approved despite findings are listed weekly.

Which controls matter, and what evidence does each leave?

Every journal reviewed. Not a sample. Evidence: a risk note stored with every journal.

Approval stays human. The agent has no approve or post right on manual journals. Evidence: the ledger's user rights and the approver identity on every posted journal.

Segregation tested, not assumed. Preparer and approver are checked against each other and against the policy. Evidence: the rights check result. The principles are in segregation of duties when an agent posts journals, and they apply equally when the agent only reviews.

Override reasons. An approver who posts a journal with findings records why. Evidence: the reason on the journal.

Weekly override review. Evidence: the controller's sign-off on the override list.

What does a risk note look like?

Take an illustrative journal of 249,000 reclassifying costs from an expense account to a capital project, prepared on the last day of the quarter at 9pm and submitted to a manager whose approval limit is 250,000. The agent's note: amount just under the approver's limit; posted after hours on the last day of the period; support is an email from the project manager, not a capital approval; expense-to-capital reclassifications by this preparer are rare. The approver asks for the capital approval, which shows only 150,000 was approved, and sends the journal back.

What should you measure?

• Journals with findings, by finding type.
• Journals sent back or rejected after findings.
• Overrides, by approver.
• Post-close corrections of manual journals, before and after.
• Audit findings on journal testing.

Where does it usually go wrong?

Notes nobody reads. The risk note sits in an attachment the approver never opens. Put it on the approval screen, findings first.

Too many findings. Every journal has three findings and approvers stop reading. Tune thresholds so findings are rare enough to matter, and review the tuning with the controller.

Agent posts by the back door. The same agent also has a posting capability for recurring templates and the boundary blurs. Keep review and posting separate capabilities, as set out in letting an AI agent post journals in your general ledger.

Support that is circular. A journal supported by a spreadsheet the preparer built from the same ledger proves nothing. The agent should flag support whose only source is the account being adjusted.

Only manual journals reviewed. Subledger interfaces can also be manipulated. Extend the same tests to unusual subledger postings once the manual review is established.

Where Pryme fits

In Pryme, journal review is a depth 1 and 2 capability certified through Attest on your own past journals, including ones that were later corrected. Approving and posting sit outside the reviewer's role scope, so the verdict for any attempt is refused. Every risk note and decision is filed. The finance bounds are in AI agents for finance.