Month-end close AI: automate the work, keep sign-off

What month-end close AI actually does

Month-end close AI is software that prepares and checks close work so people can review it: it chases checklist tasks, proposes accruals from open orders and receipts, drafts variance commentary from ledger detail, tests cut-off and reviews reconciliations. It matters because most close time goes on gathering and explaining, while accountability for the numbers stays with the controller who signs.

The useful way to think about AI in the close is not as a robot that closes the books but as a fast, tireless preparer that never gets to approve. It reads more of the ledger than a person has time to read, it writes the first draft of every explanation, and it never gets bored of checking the same fifty reconciliations. It also has no idea what your business agreed with a supplier on the phone, it can be confidently wrong, and it cannot carry professional responsibility. Everything below follows from those two facts.

This guide is about the AI layer. For the order in which close steps must happen, see our guide on the month-end close checklist; for what reconciled means per account, see our guide on balance sheet reconciliation.

Where AI fits in month-end close automation

Close work falls into six kinds, and AI helps with each in a different way. Sorting your own task list into these kinds, before you buy or build anything, tells you where automation pays and where it only adds a review step.

Three modes keep this honest. Read means the AI looks and reports but changes nothing. Draft means it prepares an entry or text that waits in a queue for a named person. Auto means it acts without review, and in a close that should be reserved for reminders, status reports and flags.

• Orchestration: tracking who owes what, spotting blocked tasks and missing evidence. Mode Auto for reminders, Read for status.
• Data gathering: pulling trial balance movements, open purchase orders, goods receipts and last month's schedules into one place. Mode Read.
• Matching: bank lines to ledger lines, receipts to invoices, subledger totals to control accounts. Mode Draft for proposed matches.
• Estimates: accruals, prepayment releases, allowances. Mode Draft only, with the basis stated for each figure.
• Analysis: flux review against last month, last year and budget. Mode Read for the analysis, Draft for the commentary.
• Narrative: the close pack summary and management commentary, written from approved figures after the numbers are locked. Mode Draft.

Checklist orchestration: the quiet win

The least glamorous use of AI in the close is often the most valuable. A close tracker, whether a spreadsheet or a workflow tool, holds tasks, owners, due days, dependencies and links to evidence. An assistant that reads it each morning can tell the controller which tasks are late, which are blocked by an upstream task, and which are marked complete without any attached evidence.

Evidence completeness is where this earns its keep. A reconciliation marked done with no workpaper, a payroll journal ticked off with no approval, or an accrual schedule whose total does not agree to the ledger line it supports are all things a reviewer finds late on day four. An assistant finds them on day one, because it compares the tracker to the ledger rather than trusting the tick.

Keep the assistant's authority narrow. It may send reminders and post a status summary without review, because a wrong reminder costs nothing. It may not mark tasks complete, change due dates or reassign owners, because those are management decisions and the tracker is part of the close evidence your auditor will read.

Accrual suggestions, and a worked review

Accruals are where AI is most tempting and most dangerous, because every accrual is an estimate and a model is good at producing plausible numbers. The safe design is that the model proposes, states the basis and the release for each figure, and a person accepts, amends or rejects each line. Remember the distinction IAS 37 draws: accruals are liabilities to pay for goods or services received or supplied but not yet paid, invoiced or formally agreed with the supplier, so the question is always whether the service had been performed by the period end, not whether someone mentioned a cost.

Worked example, in euros. At 31 August the assistant reads open purchase orders, goods received notes without a matched invoice, supplier usage reports and the close mailbox, and proposes three accruals. The controller reviews each one against its evidence, and two of the three change. That is normal: the model reads documents well but does not know about a return note filed in another folder, or that a fee estimate is not an obligation.

• Proposal 1, temporary staff: no August invoice yet; the agency portal shows 92 approved timesheet hours at the contract rate of EUR 25.00, so 92 x 25.00 = 2,300. The reviewer checks the approved timesheets and the contract rate. Accepted at 2,300.
• Proposal 2, goods received not invoiced: a goods received note of 27 August for 400 units at EUR 12.50 has no invoice, so 400 x 12.50 = 5,000. The reviewer confirms the receipt has not already been booked to stock, then finds a return note of 29 August for 100 faulty units. Amended to 300 x 12.50 = 3,750.
• Proposal 3, legal fees: an e-mail from the company's lawyers gives a fee estimate of EUR 4,000 for a lease review. The engagement letter says work starts on 15 September. No service had been received at 31 August. Rejected: there is nothing to accrue.
• Posted, one journal for the accepted items: Dr Temporary staff costs 2,300 / Dr Inventory 3,750 / Cr Accrued expenses 2,300 / Cr Goods received not invoiced 3,750. Debits 6,050, credits 6,050.
• Release: the temporary staff accrual is reversed on 1 September; the goods-received accrual is cleared when the supplier invoice is matched to the receipt.
• Recorded for each line: what the assistant proposed, the evidence it cited, the reviewer's decision and the reason. Those reasons become rules for next month's prompt.

Flux analysis and variance commentary

Flux analysis compares each account's balance or movement with last month, the same month last year and the budget, and asks for an explanation where the change crosses a threshold. A common rule is the greater of a fixed amount and a percentage, for example EUR 5,000 or 10 percent, set with materiality in mind. AI is well suited to the first draft because the work is reading ledger lines, grouping them by driver and writing plain sentences.

Two rules make AI commentary reliable. First, numbers come from a query or a spreadsheet formula, never from the model's own arithmetic, and the model receives them as data. Second, the model may explain a variance only from lines in the data it was given, and must cite the lines. Anything else is a hypothesis and must be labelled as one.

Example. Freight expense is EUR 25,000 for August against EUR 18,600 for July, an increase of 6,400 or 34.4 percent. The assistant cites three air shipments totalling 4,800 that replaced sea freight for a delayed order, and a fuel surcharge of 1,600 on the carrier's August invoice: 4,800 plus 1,600 equals 6,400, so the variance is fully explained. The reviewer confirms the three shipments against the dispatch log and asks sales whether the delayed order will recur in September, which turns a backward-looking explanation into a forecast point. This is the same discipline auditors apply to analytical procedures under ISA 520: an explanation is not evidence until it is corroborated.

Variance commentary goes wrong in predictable ways: a sentence that sounds right but has no ledger line behind it, a timing difference described as a trend, or a reclassification between accounts explained as real growth in both. Reviewers should scan for exactly those three.

Cut-off tests and reconciliation review

Cut-off errors move profit between months without breaking any balance, which is why they are hard to see and why a machine that reads every document near the boundary is useful. Reconciliation review is the second high-value target: the assistant does not prepare the reconciliation, it checks the one a person prepared, which keeps the preparer and reviewer roles clean.

In both cases the output is a list of exceptions for a person, never an adjustment. A cut-off exception may be a legitimate timing difference; a stale reconciling item may be a known dispute. The assistant's job is to make sure nobody has to discover them by accident.

• Cut-off, purchases: goods received notes dated in the last five days of the month compared with supplier invoices and accruals, flagging receipts with neither.
• Cut-off, sales: dispatch notes and delivery confirmations around the month end compared with invoice dates, flagging goods shipped in August but invoiced in September, and the reverse.
• Cut-off, services: contracts billed in arrears or in advance, checked for the right accrual or deferral.
• Reconciliation review: the reconciled balance agrees to the trial balance at the same date, to the cent.
• Reconciliation review: reconciling items older than 60 days, round-sum items and any unexplained difference presented as a plug.
• Reconciliation review: preparer and reviewer names and dates present, and the reviewer is a different person.

Close pack narrative and management reporting

The close pack is the set of statements, key schedules and commentary that goes to management once the period is closed. AI drafting helps here, provided the order is right: the numbers are final and locked first, and only then does the model write about them. A narrative drafted while journals are still going in describes numbers that will change, and nobody reliably updates the sentences.

Give the model the locked trial balance, the approved flux explanations and last month's pack, and ask it to write in your house style, to use only figures from the supplied data and to mark any sentence that interprets rather than reports. The reviewer then ties every figure in the text back to the statements. A simple tie-out rule works: if a number in the narrative cannot be found in the pack, the sentence is rewritten or removed.

Keep interpretation human. Whether a margin decline is a pricing problem or a mix effect, and what management should do about it, is a judgement the finance lead owns. The model can lay out the arithmetic of price and volume; it should not be the voice that tells the board what the numbers mean.

Roles and sign-off in an AI-assisted close

The rule that holds everything together: the person who accepts an AI proposal becomes its preparer of record, and a different person reviews and approves it. The assistant is a tool used by the preparer, like a spreadsheet, and it never appears as the approver of anything.

Auditors will ask how AI-prepared entries were controlled. ISA 500 asks them to consider the relevance and reliability of information used as evidence, and ISA 230 requires documentation an experienced auditor can follow. A close where each AI proposal, the evidence it cited and the human decision are retained is far easier to rely on than one where entries simply appeared.

• Preparer: the accountant who runs the assistant, accepts or amends its proposals and posts the entries.
• Reviewer: a second person who checks the entries, the evidence and the reasons for any amendments.
• Approver: the controller or finance manager who signs off the close and locks the period.
• Owner of the automation: the person who maintains prompts, rules and thresholds, and who should not also approve the entries the automation prepares.
• Evidence kept per item: the assistant's proposal, the documents it cited, the model and prompt version, the decision and the reason.

Claude for month-end close (as of September 2026)

Anthropic publishes a Finance plugin for Claude Cowork, marked as made by Anthropic, described as streamlining journal entries, reconciliation, financial statements and variance analysis. Its commands include journal-entry for preparing accruals, reconciliation for comparing ledger balances with subledger or bank balances, income-statement and variance-analysis. The plugin page states that all outputs should be reviewed by qualified financial professionals before use in reporting or filings. On 5 May 2026 Anthropic also announced agent templates for financial services, including a month-end closer that runs the close checklist, prepares journal entries and produces close reports, available as plugins in Claude Cowork or Claude Code, or as cookbooks for Claude Managed Agents.

For accountants who close in spreadsheets, Claude for Excel is an add-in for the Pro, Max, Team and Enterprise plans that answers questions about a workbook with cell-level citations and can trace how a number was derived. Anthropic's own limitations list says it is not recommended for audit-critical calculations without verification, and it warns that files from external sources can contain hidden instructions.

Teams that build their own close automation on the Claude API can send many commentary or accrual-review requests as one Message Batch, which Anthropic prices at 50 percent of standard API prices, with most batches finishing in under an hour. Current models include Claude Opus 5.5, Claude Sonnet 5 and Claude Haiku 4.5. Our guide on automating accounting with the Claude API covers the build itself.

Security and data protection in the close

Close data is among the most sensitive a business holds: unpublished results, payroll, customer balances and supplier terms. Treat every AI step as a data flow that needs the same controls as any other system that touches the ledger.

On the data protection side, Anthropic states that by default it does not use inputs or outputs from its commercial products, such as Claude for Work and the API, to train its models, that for Claude for Work customers it acts as a processor on the customer's behalf, and its API documentation describes Anthropic as the data processor for the Claude API. Under the GDPR you still need a lawful basis, a processor agreement and a view on international transfers; that is general information, not legal advice, so involve whoever handles data protection for your organisation.

• Read-only by default: the assistant gets a ledger extract or read-only reports, not a login that can post.
• No posting without approval: drafts land in a queue; a named person posts them.
• Prompt injection: supplier e-mails and PDFs in the close mailbox can contain instructions aimed at the assistant. Treat their text as data, never as commands.
• Data minimisation: payroll commentary does not need employee names; aggregate by department before the data leaves the ledger.
• Logging: keep who ran what, on which data, with which model, and what was accepted.
• Closed periods stay closed: the automation must not be able to post into a locked month.

Choosing month end close automation software

Month end close automation software ranges from task trackers with reminders, through reconciliation matching tools, to ERPs where the close controls are part of the ledger. Whatever you consider, test the controls first and the AI second: an assistant layered on a ledger that lets anyone post into last month is automating a weak close.

• Period locking that refuses back-dated postings, including automatic ones from sales and purchases.
• An approval workflow for journals above a threshold, with the approver recorded.
• An audit trail with who, when and the old and new values of every change.
• Comparative reports (previous period, previous year, budget) you can export for flux review.
• AI proposals that show their evidence and wait for approval, never silent postings.
• An open way to connect your own AI assistant, such as an MCP server, with permissions no wider than the user's.

Month-end close in Skyline Nexus ERP

Skyline Nexus ERP gives the close the controls an AI layer depends on. Fiscal periods can be closed and locked, and posting into a closed or locked period is refused, whether the entry is manual or comes from a sale or purchase. The year cannot be closed while any journal in it is unposted. Journals at or above the approval threshold go to an approver when approval is switched on, accruals are posted as manual journals and reversed with Reverse Journal Entry and a reversal date, and the Audit Trail records who created, approved, posted or reversed each entry with old and new values.

For flux review, the Profit and Loss and Balance Sheet in Skyline Nexus ERP compare with the previous period or the previous year, the Data Verification report compares the trial balance with ledger and point-of-sale transactions, and the Audit Pack exports the trial balance, journals and ledger detail to one Excel workbook. The in-app assistant, and Claude Desktop through the Skyline Nexus ERP MCP server with a personal token an administrator issues, can answer questions such as purchases this month or expenses by branch against last period from eight permission-checked live figures.

Claude drafting accruals, journals and close commentary for approval, and AI close runs inside the product, are being rolled out on the Skyline Nexus ERP roadmap; ask us for your go-live date.