AI Journal Entries and Accruals: How to Draft Them With AI and What to Check Before You Approve

A common pattern on working day two: the accruals schedule is copied from last month, the new lines are typed in from emails and purchase orders, the old ones are released by eye, and the journals are keyed one by one. It works, it is slow, and the mistakes it produces are the boring kind, such as an accrual that was invoiced three weeks ago and never released.

Those boring mistakes are exactly what AI is good at catching, and the drafting is exactly what it is good at doing. The question finance teams keep asking is how to trust the output. This guide answers that with a method you can use this month.

Where AI Fits in Month-End Journals

Think of the AI as a capable new preparer who was not in any of last month's conversations. It can read support quickly, apply a consistent method and write clear narratives. It cannot know that a contract was renegotiated on a phone call, or that the CFO wants a particular cost treated differently this year.

That makes the good uses fairly clear:

• Rolling forward recurring accruals and prepayments, and releasing ones that have been invoiced.
• Drafting standard journals from support: payroll, depreciation, prepayment releases, recharges.
• Writing the narrative and the reviewer note for each entry.
• Checking a batch of manual journals before posting, for balancing, valid accounts and unusual amounts.

The poor uses are the judgement calls: provisions, estimates, anything that depends on a conversation. Those stay with people. Anthropic's finance plugin for Claude includes a journal entry skill built around this split; our guide to Claude skills for finance covers how to install and adapt it.

Rolling Forward Recurring Accruals

Accruals are the best first job, because the inputs are documents you already have and the output is easy to check. Give the assistant last month's schedule, the open purchase orders or goods received not invoiced, and the purchase invoices posted since month-end:

You are preparing month-end accruals for a UK company.

Inputs: last month's accrual schedule, this month's goods received not invoiced report, and purchase invoices posted since last month-end.

1. For each existing accrual, say whether it has now been invoiced. If so, propose releasing it and quote the invoice.
2. Propose new accruals from goods received not invoiced, with supplier, amount, nominal code and the source line.
3. List anything you could not match, and anything where the invoice differs from the accrual by more than £500.
4. List every assumption you made, separately.

Do not estimate any amount that is not in the inputs.

The output is a proposed schedule with every line traced to a document. The reviewer's job shifts from rebuilding the schedule to challenging the proposals, which is a better use of their time.

Drafting a Journal From Support

For standard journals, give the support and ask for the entry in the format your team posts, plus a narrative a reviewer will understand without opening the support:

Draft the journal for this prepayment release using the attached invoice and our prepayments schedule.

Return: date, period, each line (nominal code, description, debit, credit), totals, a reversal date if one applies, and a two-sentence narrative for the reviewer.

Confirm that debits equal credits and that every nominal code exists in the attached chart of accounts. List any assumptions.

Five Checks Before You Trust an AI-Drafted Entry

1. It balances and the accounts exist. Debits equal credits and every code is live in your chart. Your ERP will usually enforce this, but check before upload rather than after.
2. Every amount traces. Each figure points to a document or a recalculation you can repeat. No figure appears from nowhere.
3. Period and reversal are right. The posting period, and the reversal date for accruals, match your close calendar.
4. The narrative stands alone. A reviewer can understand the entry from the description without opening the support.
5. Assumptions are listed, not buried. Any assumption about an amount, a date or a treatment is a reason to stop and confirm with the person who knows.

The same claim-by-claim discipline works for commentary; our guide to checking AI variance commentary applies it to the board pack.

Controls, Segregation of Duties and Audit

Nothing about AI changes the basic control: the person or system that prepares an entry should not be the one that approves it. Treat the AI as the preparer, keep a named approver, and record both. Most ERPs already support this through journal workflow.

Keep the prompt, the inputs and the draft with the posted journal, the same way you would keep a preparer's workings. If your auditors ask how an entry was produced, you can show them. Vendors are consistent on this point too: Anthropic's published finance agents state that they do not post to a ledger and that all outputs are staged for human sign-off.

If you connect an AI tool directly to your ERP, make the connection read-only to begin with. Posting rights are a decision for the finance director and your auditors, after months of evidence, not a default. Our AI governance framework for finance sets out the wider controls.

Frequently Asked Questions

Our AI for Finance Leaders course covers month-end workflows like this with marked practice, and the first two modules are free. Our AI Opportunity Blueprint finds which close tasks to automate first, and our AI consulting team can set up the workflow with your controls.