The AI month-end close, step by step

Most articles about AI and month-end promise the close will "run itself." It won't, and if a vendor tells you otherwise, ask who signs the balance sheet. What AI can genuinely do today, with tools your firm probably already pays for, is compress the first-pass work at almost every step: sorting reconciling items, drafting accrual workings, writing the first version of variance commentary, and turning your close process into a checklist someone else can follow.

This is a tool-agnostic walkthrough. Everything here works in Claude, ChatGPT, Copilot or Gemini, because the technique is the same: feed the model your raw material, ask it to structure, group or draft, and review everything before it touches the ledger or a client. For the deep end, we ran a detailed look at Claude handling first-pass close work for a $20M manufacturer, integrations and all. This is the on-ramp version: no integrations, just copy-paste and judgment.

One rule before any of it: never paste identifiable client data into a consumer AI tool. Use placeholders ([CLIENT], [PERIOD], [AMOUNT]) or a business-tier product where your firm controls data handling. We've covered what happens to client data in AI tools separately; read it before rolling any of this out to a team.

Before the close: turn your process into a real checklist (Day -5 to Day -1)

Most close processes live in one senior person's head, plus a spreadsheet nobody has updated since a staff member left. The highest-leverage AI task in the whole close happens before the period ends: get the process out of heads and into a checklist with owners, dependencies and deadlines.

I will describe our month-end close process for a client conversationally, in no particular order. Turn it into a checklist: a table with task, owner (bookkeeper, accountant, reviewer or client), day due (Day 1 to 5), dependency, and source system. Flag any steps that seem to be missing compared to a standard small business close (bank recs, payroll, GST/VAT/sales tax, fixed assets, accruals, review, period lock). Ask me clarifying questions first. Here is the process: [DESCRIBE IT].

The "flag any steps that seem to be missing" line matters: comparing your checklist against the standard shape is a cheap way to catch the step that quietly disappeared when the person who owned it left. Once it exists, the checklist lives in your practice management tool (FYI, AccountKit, XPM, wherever your firm runs work) as a recurring job template.

Also in the pre-close window: chase the inputs. Missing bank statements, unanswered queries, the loan statement the client never sends. Chase emails are a solved AI problem, and if you give the model two of your real ones to copy the tone from, the drafts stop sounding like a robot wrote them.

Human review point: the checklist itself. AI drafts and stress-tests it; a senior person decides what the close actually includes for each client. That call is scoping, and scoping is judgment.

Day 1 to 2: reconciliations and rec triage

AI does not do your bank rec. Ledger software has matched transactions automatically for a decade. What eats the first two days of a close is the residue: the list of unmatched and reconciling items someone has to think about, one line at a time. That is triage work, and triage is pattern-grouping, which is exactly what language models are good at:

Here is a list of unreconciled items from a bank reconciliation, with dates, descriptions and amounts: [PASTE LIST]. Group them into: (1) likely timing differences, (2) possible duplicates, (3) likely coding errors, (4) needs investigation. For each "needs investigation" item, list the two most likely explanations to check first and what evidence would confirm each. Do not invent items or guess amounts, and say plainly when a description is too thin to classify.

The output is not the answer. It's a worked list, ordered so a human starts with the items most likely to be real problems instead of grinding through the noise top to bottom. On a messy rec that reordering alone can save an hour, and the same prompt works for supplier statement recs, clearing accounts and intercompany balances.

Human review point: every classification. The model suggests where to look first; the accountant confirms what each item is. Nothing gets adjusted, written off or reclassified on the model's say-so.

Day 2 to 3: accrual and journal prep

Accruals have two parts: the estimate, which is judgment, and the working, which is structure. AI helps with the second and must not be allowed near the first. Use it to standardise the workings that support each accrual, so every file in the firm documents estimates the same way:

I need a standard accrual working. Context: monthly accrual for [EXPENSE TYPE], estimated on [BASIS: e.g. prior-quarter average, contract value, pro-rata of an annual invoice]. Produce a working paper layout with: the basis, the calculation with each input labelled and its source stated, the prior three months of the same accrual for reasonableness, the journal (DR/CR, accounts, amount), and the reversal noted for next period. Use only figures I give you: [FIGURES]. If an input is missing, mark the cell TBC rather than estimating it.

The "mark the cell TBC" constraint does the safety work in that prompt. The failure mode of AI in accrual prep is a plausible-looking number filled into a gap; the fix is telling the model that gaps are the correct output. Same for recurring journals: AI can draft the standing schedule and flag figures that move beyond a threshold you set, but the trigger for posting is a person.

Human review point: every estimate and every posting. AI drafts the working; the accountant owns the number in it. If a number came from the model rather than a stated basis, it does not go in the file. It's the line we drew in our look at agentic AI in accounting: the further a tool moves toward acting on the ledger rather than drafting for review, the tighter your controls need to be.

Day 3 to 4: variance review and first-pass commentary

This is where AI saves the most senior time, because variance commentary is drafting, and drafting from known facts is what these tools do best. You supply the movement and the driver, the model supplies the sentences.

I am writing management commentary for a monthly reporting pack. Here are the variances worth commenting on, each with the movement and the driver I have confirmed: [ACCOUNT, PRIOR, CURRENT, DRIVER] per line. Draft one short paragraph per variance in a neutral reporting style: state the movement, state the driver, note any expected reversal I have mentioned. No speculation, no drivers I have not given you, no filler. Under 60 words each.

Run the review itself first, the normal way: scan the comparatives, confirm the drivers, decide what deserves a comment. The model then turns your findings into clean prose in one pass instead of forty minutes of wordsmithing. You can also invert it, as a second pair of eyes before the partner sees the pack:

Here is a summarised P&L and balance sheet for [PERIOD] with comparatives, anonymised: [PASTE]. List the ten movements a reviewer would most likely question, ordered by size and unusualness, each phrased as the question a partner would ask. Flag balances that look inconsistent with each other (revenue up while debtors and cash both fall, say). Do not explain the movements; I only want the questions.

That second prompt is a pre-review, not a review. It surfaces the question you would have been asked on Day 5, on Day 3, while there's still time to fix the answer. More working patterns like these are in our library of AI prompts accountants actually use.

Human review point: the drivers. AI must never be the source of a variance explanation. If you don't know why a number moved, the honest commentary is "under investigation", not a model's fluent guess.

Day 5: reporting, sign-off and the close retro

AI helps with both edges of the final pack: tightening commentary to a consistent voice across clients, and drafting the cover email. Sign-off is not on the list. The reviewer reads the file, the partner signs, the period locks, and no AI vendor's roadmap changes that, because the person signing is professionally accountable.

Then the step almost nobody does: a ten-minute retro.

Here is our close checklist and what actually happened this month, including what ran late and why: [PASTE BOTH]. Suggest specific changes: tasks to move earlier, dependencies to add, tasks to split or reassign, and any recurring blocker needing a permanent fix rather than a workaround. Rank by likely days saved next close.

Do that for three closes running and the checklist stops being documentation and becomes a system that improves itself, with you approving each change. That habit, more than any single prompt, is what the AI month-end close actually looks like in 2026: same steps, same sign-off, materially less grind between them. For where these tools fit across the rest of a firm, start at our guide to AI for accounting firms.

None of this is professional advice for any specific engagement; apply your own standards and your firm's review policies to anything AI touches.

Frequently asked questions

Can AI fully automate the month-end close?

No. Transaction matching and recurring journals were automated by ledger software before modern AI arrived. What AI adds is first-pass thinking work: triage, drafting and structuring. The judgment layer (estimates, materiality calls, sign-off) stays human because a person, not a model, is accountable for the statements.

Which close task should a firm automate with AI first?

Reconciliation triage. It needs no integration, no new software and no process change: paste the unmatched items, get back a grouped, prioritised list, work it top down. It saves time on day one and builds the review habit every other AI close task depends on.

Do we need our ledger connected to an AI tool for any of this?

No. Every prompt here works by copy-paste with anonymised data. Direct connections (APIs or MCP-style integrations) remove the copy-paste step and matter at volume across many clients, but they add data governance questions you should answer deliberately, not stumble into.

How do we stop AI inventing numbers in close workings?

Constrain the prompt, then verify the output. Tell the model to use only figures you supply, mark missing inputs as TBC rather than estimating, and say when it can't classify something. Then treat every number in AI output as unverified until someone traces it to a source.

What rules should exist before the team uses AI on close work?

Three at minimum: which tools are approved and at what data tier, what client data may be pasted into them, and which outputs require review before use (for close work, all of them). One page is enough to start; the point is that the rules exist before the habits form.