Where AI agents fit into a bookkeeper's month-end close

The industry is full of lists of AI agents that promise to do your books. Most of them describe tools, not procedures. What matters to a working bookkeeper is where an agent can genuinely help, where it must not be trusted, and how to keep the working papers defensible when software did part of the work.

What does an AI agent actually do in bookkeeping?

An agent is software that takes an action, not just an answer. A chatbot that explains a journal entry is not an agent. Software that reads the bank feed, proposes a coding to a supplier account, and posts the entry once you approve it, is closer to the real thing.

In practice, agents cluster around a few tasks: transaction coding from descriptions and history, matching bank lines to invoices, drafting accrual and prepayment journals from recurring patterns, and flagging anomalies such as duplicate invoices or odd amounts. These are all mechanical steps inside the close. The judgement around them stays with you.

The parts of the close agents help most

Coding is the strongest case. A well-trained agent applies your historical mapping consistently, which beats manual entry on volume. Bank matching is similar: it clears the obvious 80 percent and leaves you the exceptions.

Draft journals are useful too. If you accrue the same monthly software subscription, an agent can draft the accrual and the reversal from the pattern. You review the amount, the cut-off date and the account, then approve.

Anomaly detection earns its keep at cut-off. Invoices dated just after period end but delivered before it, duplicates, and round-sum entries are exactly the kind of thing pattern-matching catches well.

Work you must not delegate

Cut-off judgement is the clearest boundary. Whether goods were received before period end is a fact question that needs evidence: a delivery note, a goods-received record. An agent can flag the candidate transactions. It cannot decide the treatment.

Estimates, provisions and anything requiring policy choice stay with the accountant. So does anything involving a conversation with a client about what a transaction really was.

Reconciliations also need a human sign-off. An agent can tick and match, but the reconciliation is your assertion that the ledger agrees to the bank, and your name goes on it.

Keeping the evidence clean

If software proposed or posted an entry, the working paper should say so. Keep the approval trail: what the agent proposed, what you changed, and what you accepted. Most packages log this automatically; make sure it is retrievable at year end.

Review is faster if you sample deliberately. Check every large item, every round sum, and a random slice of the routine ones. That is the same discipline you would apply to a junior, because an agent is effectively a very fast junior with no memory of its own mistakes.

A sensible way to adopt

Start with one task, coding or bank matching, in one entity. Reconcile the agent's output against your own for two or three months until the error rate is known and acceptable. Then widen scope. Document the account mappings and the review procedure so the process survives staff changes and year-end questions.

The ledger still closes the way it always did: evidence, cut-off, reconciliation, review. Agents change the speed of the mechanical parts. They do not change who is responsible for the answer.