Train finance reviewers around the close calendar. Teach how the agent drafts journals, reconciles accounts and writes variance commentary before a quiet month. Calibrate on last month's real entries. Shadow one full close with every review checked by the financial controller. Sign off reviewers who catch the planted errors and can show the evidence behind each approval.
What does the finance team review?
A finance agent typically reconciles bank and control accounts, matches intercompany balances, drafts accruals, prepayments and recurring journals, prepares close checklists, and writes first-draft variance commentary for management accounts. Recurring journals below a threshold may post at depth 3. Everything else is drafted at depth 2 for a person to approve. Payments are outside its scope, or held at depth 4 for named approvers.
Reviewers therefore handle four kinds of output. Drafted journals, where they check the logic and the support. Reconciliations, where they check that unreconciled items are real and explained. Variance commentary, where they check the story matches the numbers. And a sample of posted recurring journals.
Finance has a particular constraint: segregation of duties. The person who prepared or requested an entry cannot approve it, and the agent's drafts do not change that. Training must make clear who may approve what.
The financial controller owns the programme. The finance director agrees the sign-off standard. External auditors will ask how the review works, so the records matter.
What are the steps?
- Learning, mid-month: reviewers complete lessons on how the agent drafts: the sources it uses, the bounds for posting, and why it holds.
- Agent, depth 1: builds a practice set from last month's journals, reconciliations and commentary, with any corrections made.
- Calibration, before close: reviewers assess the set alone, then compare. The controller seeds three entries with deliberate errors.
- Shadow, one full close: reviewers approve live drafts; the financial controller re-checks every approval.
- Agent, depth 2 and 3 in live work: drafts entries and commentary; posts recurring journals within threshold.
- Sign-off: the financial controller assesses shadow approvals and the seeded-error results.
- Ongoing: post-close review of a sample and calibration before each quarter end.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Lessons | Nothing | none | Reviewers |
| Practice set | Lists last month's entries and corrections | 1 | Controller selects |
| Calibration | Nothing | none | Reviewers, controller |
| Shadow close | Drafts entries; posts recurring within threshold | 2 to 3 | Reviewer approves; controller re-checks |
| Sign-off | Nothing | none | Financial controller |
| Post-close sample | Lists posted entries | 1 | Senior reviewer |
What must reviewers learn to check?
Support for each journal. The agent attaches what it used: an invoice, a contract, a calculation. Reviewers check the support exists, matches, and comes from a reliable source. A supplier email is external; a contract in the contract system is sealed.
Period and account. Right month, right account, right cost centre. Cut-off errors are the most common agent-drafted mistake in early weeks, especially around invoices dated at month end.
Reconciling items. An unreconciled difference with an explanation is only acceptable if the explanation is true. Reviewers trace a sample to source and look for items carried forward month after month.
Variance commentary. The agent explains movements from the numbers it sees. It may not know about a one-off contract, a pricing change or a reorganisation. Reviewers check the story with the budget holder where the explanation matters.
Recurring journals. Posted automatically, so easily forgotten. Sample them; check they still should recur.
Segregation. Approve only what you are entitled to approve, and never your own requests.
Seeded errors make calibration concrete. Before close, the controller plants three faults in the practice set: an accrual booked to the wrong period, a reconciling item explained as timing that has in fact been outstanding for four months, and variance commentary that blames volume when the ledger shows a price change. Reviewers who find all three are ready to shadow. Reviewers who miss one learn more from that miss than from any lesson.
Which controls leave evidence?
- Lessons before shadow close. Evidence: completion dates.
- Seeded errors used in calibration. Evidence: the seeded list and results.
- Every shadow approval re-checked. Evidence: the controller's name.
- Segregation enforced. Evidence: approval rights in the general ledger match the role matrix.
- Recurring journals sampled. Evidence: the post-close sample log.
What should you measure?
- Seeded errors caught per reviewer.
- Post-close adjustments to agent-drafted entries.
- Audit findings on journal review.
- Reconciling items older than three months.
- Days to close, to confirm review is not slowing it unduly.
Where does it usually go wrong?
Training lands at close. Nobody learns during close. Schedule lessons mid-month.
Commentary taken at face value. Well-written variance commentary is persuasive. Train reviewers to check it against what they know.
Recurring journals never questioned. A journal that should have stopped keeps posting. Sample quarterly at least.
Approval rights drift. People move roles but keep rights. Review rights against sign-off monthly.
The agent's threshold is mistaken for materiality. The posting threshold decides what the agent may post alone. It says nothing about whether a smaller entry is right, and small recurring errors add up over a year.
Claims and operations teams follow the same four stages on different work; see how to train the claims team to review AI agent decisions and how to train the operations team to review AI agent decisions. The reviewer onboarding checklist for the finance team and the AI agent competency matrix for the finance team give the detail. The full programme is in training teams for AI agents. Pryme Intelligence Academy's Operator track includes a module on usage and billing, which finance reviewers often find useful when they also own the agent's cost line.
Questions people ask
Should the agent post journals at all?
Recurring, low-value journals with fixed logic are reasonable at depth 3 once certified. Start with everything drafted and widen with evidence.
What do external auditors want to see?
Who reviewed each entry, what they checked, and that the reviewer was entitled to approve it. The agent's record plus your approval log usually answers this.
How do we handle a reviewer who misses seeded errors?
More calibration and another shadow close. They do not approve alone until they pass.