It is the third working day of the close. The bank still has unmatched lines from the second week of the month, the accrual for goods received but not yet billed lives in someone’s spreadsheet, and the management pack will go out with commentary written late at night by the person who also prepared the numbers. The close is rarely slow because the accounting is hard. It is slow because of matching, chasing and explaining, three kinds of work that AI handles well if the controls around it are designed first.
Finance teams are already moving. Gartner’s 2025 AI in Finance Survey found that 59% of finance leaders use AI in their finance function, against 58% in 2024 and 37% in 2023, and that adoption has levelled off, with complexity, data and talent cited as barriers. That figure covers the finance function broadly, not the close specifically, but the barriers it names are exactly the ones a close project meets.
What does AI in the month-end close mean?
AI in the month-end close is the use of matching models, anomaly tests and language models to prepare, check and explain a period’s numbers, while people approve every entry that changes the books. It is useful to split the close into three kinds of work, because AI plays a different role in each:
- Preparing: matching bank lines, reconciling balance sheet accounts, drafting accruals and adjustments.
- Checking: testing every manual entry, cut-off, valuation movements and unusual balances before the period is locked.
- Explaining: drafting variance commentary and the narrative of the management report.
The steps of a sound close do not change. Our month-end close checklist for Odoo sets them out; this article is about where AI fits inside them.
Which close tasks can AI do, and which must stay human?
The decision rule we apply: AI may propose anything, may check everything, and may change the books only through a draft that a named person approves. The matrix below applies it task by task.
| Close task | What AI does | What a person does | Evidence kept |
|---|---|---|---|
| Bank reconciliation | Proposes matches and lists unmatched lines, as described in automating bank reconciliation with AI | Approves matches and every write-off | Reconciled and unreconciled items at period end |
| Accruals for goods received not billed | Drafts the accrual from receipts with no bill, with a reversal date | Approves the estimate and the reversal | The receipt records the draft cites |
| Balance sheet account reconciliations | Compares ledger balances with subledgers and lists breaks | Resolves breaks and signs the reconciliation | Signed reconciliation with its support |
| Manual journal entry review | Scores every manual entry for timing, user, roundness and posting after cut-off | Investigates the high scores | Score, reviewer and outcome per entry |
| Inventory valuation close | Compares unit cost movements and drafts valuation adjustments | Approves valuation entries | Recomputed cost per item moved |
| Variance commentary | Drafts explanations that cite accounts and records | Edits and owns the narrative | Record references behind every figure |
| Period lock | Checks that preconditions are met: no unapproved drafts, no open breaks | Sets the lock date | Lock log and the checklist sign-off |
How does AI change the close calendar?
The largest gain does not come from doing close tasks faster on day 1. It comes from moving them out of the close window altogether.
During the month
If bank matching and manual entry scoring run daily, the close starts with a short list of exceptions instead of a month of raw transactions. The same scoring that supports AI-based fraud detection doubles as a close control: a round manual entry posted late on the last evening of the month is a classic finding in MFT’s continuous-audit work, and it is far easier to question on the day it is posted.
The first days of the close
AI drafts accruals, reclassifications and valuation adjustments, each with the records it relied on. The accountant’s job shifts from building entries to reviewing them, which is faster only if every draft shows its evidence.
Reporting
Once the trial balance is approved, AI drafts the variance commentary and the management report narrative. The controller edits. Nothing in the pack is a number the controller cannot trace.
Who approves what AI drafts?
Speed is worthless if it comes from skipping review. The approval design we use with clients has four rules:
- Every draft carries the record references it was built from. A draft without them is returned, not reviewed.
- Approval is tiered by the client’s ISA 320 materiality threshold: below it, one named approver; above it, a second approver. Our article on setting the materiality threshold explains how to fix it.
- The person who requested or ran the AI task never approves its output, which is segregation of duties applied to AI agents.
- Corrections are made by reversal, never by deleting or editing a posted entry, so the trail of what was proposed, approved and corrected survives.
A close is fast when the evidence is ready, not when the approvals are skipped.
Can AI write the financial report commentary?
It can write the first draft, and it often writes it better than a tired controller at midnight. The danger is specific: a language model can produce a fluent sentence with a number that does not exist in the books. A practical tie-out test catches this before the report leaves the finance team:
- Highlight every figure and percentage in the draft commentary.
- Trace each one to the approved trial balance or the records the draft cites.
- Strike out any figure that cannot be traced, and any sentence whose cause is asserted without a record behind it.
- Keep the traced draft with the close file as evidence of review.
MFT’s agents apply the same rule automatically under the evidence-or-silence principle described in our article on AI provenance and the audit trail: a finding that cites nothing, or a number absent from the cited records, is rejected and shown struck through, never as a fact. The human tie-out remains, because the controller signs the pack, not the model.
What does Odoo 20 change for the close?
According to Odoo’s 20 release notes, several accounting changes touch the close directly:
- A bank reconciliation summary report listing reconciled and unreconciled transactions at period end, plus the option to run Auto Reconciliation manually.
- Bank consistency: entries affecting bank accounts must originate from bank transactions.
- Under perpetual valuation, accrual entries from the inventory valuation closing report post directly to the valuation account, and cost of goods sold is updated retroactively when cost changes after delivery. See the Odoo 20 inventory valuation changes.
- Exchange gains and losses consolidated into one line per invoice, and depreciation models that accept rates and decimals.
- Payment statuses renamed: "In Process" becomes "Paid" and "Paid" becomes "Reconciled", which affects saved filters and close reports.
- AI-powered financial insights: questions to the AI assistant about receivables, cash flow or reconciliation issues, answered solely from accounting data in your database, including a review of suspected reconciliation errors.
All AI features in Odoo 20 run on IAP credits billed per request and not included in any plan, so a daily AI routine has a running cost to budget. Lock dates, reconciliation models and Spreadsheet dashboards exist in recent Odoo versions; confirm how they behave in your 20 database before you build the close around them. The full AI map is in Odoo 20 AI features for accounting.
What can go wrong when AI enters the close?
- Automating a broken close: if the chart of accounts or cut-off rules are unclear, AI reproduces the confusion faster.
- Accepting matches in bulk: approval becomes a click, and the control disappears while the log says it happened.
- Drafts piling up: unapproved drafts at lock date mean the period closes on numbers nobody owns. Make "no open drafts" a lock precondition.
- Evidence in the wrong place: Odoo 20 keeps stored AI conversations up to 30 days after the last message, so anything that supports a closing entry belongs in the close file, not in a chat.
Where should you start this month?
- Write your close calendar as a list of tasks with the working days each takes and who owns it.
- Mark the tasks that are mostly matching, chasing or explaining. Those are the candidates.
- Run AI in parallel for one close on two of them, without changing the official process, and compare its drafts with what your team posted.
- Set the approval tiers and the lock preconditions before the AI output counts.
- In the following close, move matching and entry scoring into the month, and keep commentary drafting for last.
This is the work our Virtual CFO agent does as a managed service: it reads the books through the API at query time, drafts analysis and adjustments that cite the records behind them, and leaves approval to a named person on the client’s side; nothing is auto-posted. For teams that also want an experienced finance lead around the close, our Virtual CFO service pairs that work with people who have closed books and signed reports themselves.