AI in the Month-End Close: What to Hand the Model, and What to Keep

In a group’s month-end close, AI earns its place on the repetitive checks: matching, tying out, comparing both sides of an intercompany flow, flagging what does not add up, and preparing proposals. It should not decide, lock a period or post. And it does not fix a slow close. A close that drags in a multi-entity group usually waits on intercompany flows nobody agreed and charts of accounts that drifted apart. Settle those first, and the model finally has something stable to check.

Why doesn’t AI fix a slow close?

Because the model reads what the entities recorded, and a slow close is rarely a reading problem. If two subsidiaries book the same intercompany flow in two different ways, an AI finds the difference faster than a person does. It cannot decide which side is right. That is an arbitration, and it needs someone with the authority to settle it.

Business Central shows where those decisions sit. A consolidation can bring together companies with different charts of accounts, but only once someone has mapped each posting account to an account in the consolidated company, through the Consol. Debit Acc. and Consol. Credit Acc. fields. And Microsoft’s documentation is explicit about the next step: processing consolidation eliminations is a manual process. The mapping and the elimination rules are choices people make. A model can apply them. It cannot make them.

When a group’s close is slow for these reasons, the work that unblocks it is root cause consulting, not a new tool. AI comes after.

What does a month-end close in Business Central consist of?

A list of optional tasks, not a mandatory procedure. Microsoft’s own wording is that Business Central does not force you to close periods. Its period-end overview lists, among others: setting posting periods, making general ledger adjustments, posting recurring journals, posting all open sales and purchase documents, reconciling receivables and payables to the general ledger, posting fixed asset depreciation, processing intercompany transactions and completing tax statements.

Two details matter before any AI goes near it:

• Accounting periods do not control posting dates. The Allow Posting From and Allow Posting To fields do, on General Ledger Setup for the company and on User Setup per user, where the user setting overrides the company one.
• Since the 2026 release wave 1, those limits can be date formulas. Microsoft warns that formulas are evaluated against the work date, which users can change, so they are less restrictive than fixed dates set by an administrator.

Read that list once more and sort it. Some tasks are decisions. Some are checks that repeat every month with the same rules. Some are postings. AI belongs in the second group only.

Which close tasks can a model prepare?

The checks that follow a rule someone can state. The pattern is always the same: read, compare, flag, propose.

• Tie-outs. Receivables and payables totals against the general ledger, per entity, with the gaps listed.
• Open documents. Sales and purchase documents still unposted at the period end, recurring journals not yet run.
• Both sides of intercompany. The Intercompany Transactions report lists posted intercompany sales, purchases and journal entries between internal entities. Comparing what entity A booked toward B with what B booked from A is exactly the kind of tedious, rule-based work a model does well.
• Entries outside the window. Postings dated outside the expected period, or made by users whose posting window was left open.
• Bank reconciliation. Business Central already ships this one. Reconcile with Copilot runs the standard automatch first, then uses AI on the unmatched lines, based on dates, amounts and descriptions. Nothing it proposes is saved until a user selects Keep it.

That last point is the shape to copy for every other check. On our own Business Central, an AI agent handles an internal check of the same kind: it reads the month’s time entries, groups them, flags the lines that do not add up and prepares the corrected lines. Saving them is a separate step, run on purpose by a person. The model prepares. A person commits.

What must stay with a person?

Everything that is a decision or a commitment:

1. Closing the posting window. Who may still post after the cut-off, and until when, is a control, not a convenience. It lives in General Ledger Setup and User Setup, and a person sets it.
2. Arbitrating intercompany differences. The model shows that two entities disagree. Someone decides which one corrects its entry.
3. Eliminations. Manual in Business Central, and a matter of judgment on what counts as internal.
4. Estimates. Accruals, provisions and adjustments carry judgment that has to be signed by a name.
5. Posting. The model writes proposals into a journal or a holding area. A named user posts them, under their own permissions.

The detail of that write path, and what an auditor will ask to see, is in our article on AI governance in finance.

Why must the model’s “all clear” be checked by a second route?

Because the most dangerous output of a close check is an empty list. An agent that filters on the wrong field, or on the wrong company, reports “no difference found” with the same confidence as a real clean result. In a close, that silence is what gets a period locked with an error inside it.

So every result the model returns gets compared with a number obtained another way: the count of intercompany lines it checked against the totals of the standard report, the tie-out it reports against the trial balance. If the two routes disagree, nobody locks anything. We cover this failure mode in detail in our article on AI agents on ERP data and the “nothing found” answer.

In what order should a group bring AI into its close?

In the order that makes each step checkable:

1. Settle the rules first. Intercompany flow conventions, account mapping for consolidation, who owns each balance. If nobody can state the rule, no model can run it.
2. Write each check down as it is done today. Which report, which filter, which tolerance, what counts as an exception. That written check is the specification the model follows.
3. Run the model read-only, next to the people. One entity, the same checks, and compare its output with the human result over several closes.
4. Then let it prepare proposals. Into a journal or a holding area, never straight into posted data. Posting stays with a named person.

None of these steps promises a number of days saved. What they do is move the repetitive part of the close onto something that does not get tired on day eight, while the decisions stay where they belong.

The Asio Services way

We start a close project with the decisions, not with the model. Which intercompany rules are settled, which are not, who owns each balance. Then we write the checks down, run them read-only next to the finance team, and only then build the write path, with a holding area and a named user. If your close is slow and you are being asked for an AI plan for it, start with a clarity session and we will tell you which part is a tool question and which part is not.

FAQ

Can AI close the books on its own in Business Central?

No, and it should not. AI can prepare the repetitive checks of a close, such as tie-outs, intercompany comparisons and bank matching proposals. Locking the posting window, arbitrating differences, eliminations and posting stay with named people.

Will AI make our month-end close faster?

It can take repetitive checking off the finance team. But a close that is slow because intercompany flows were never agreed, or because charts of accounts diverged, stays slow until those decisions are made. The model only checks faster what the entities recorded.

Does Business Central already have AI for the close?

Yes, for bank reconciliation. Reconcile with Copilot proposes matches after the standard automatch and can suggest general ledger accounts for residual lines. Nothing is saved until a user keeps the proposals.

Where should AI start in a group close?

With a read-only check that the team already runs by hand and can describe precisely, such as comparing both sides of intercompany transactions. Run the model next to the people for several closes before letting it prepare any entry.