A close is a list of small, dull tasks done in the right order by people who know which numbers to distrust. Most small businesses do not have a close problem so much as an order problem: they reconcile the bank before cash application is finished, book accruals before the bills arrive, and review the P&L before anyone has looked at the balance sheet. Then they do it all again on day 12.
AI tools now claim to run much of this. Some of those claims hold up. Intuit says its Accounting Agent "automates bookkeeping and transaction categorization, and assists in reconciliation." Xero says JAX now matches transactions to bank feeds automatically, in real time, and that complex cases such as one payment split across sales and fees are coming later. Puzzle lists AI-powered reconciliations and an AI accuracy review on its Complete plan. Digits sells an "Agentic Close." The useful question is which day of your close each of these tools touches, and what you still have to check by hand.
The nine-day close
The plan below runs from three business days before month-end (day -3) to five business days after (day +5). A two-person finance team at a company with one or two bank accounts, a card program, and some inventory or deferred revenue can hit it. If you are a solo bookkeeper, stretch it; the order matters more than the dates.
| Day | Tasks | AI can help? |
|---|---|---|
| -3 | Send cutoff reminders to staff for expense reports and receipts. Ask department heads for known accruals (work done, not yet billed). | Yes: drafting and sending reminders, chasing missing receipts |
| -2 | Review open POs and unbilled work. Confirm payroll dates and amounts. | Partly: listing open POs; a person confirms what was received |
| -1 | Final invoicing run for the month. Freeze vendor master changes until close is done. | Yes: bulk invoice drafts from contracts or timesheets |
| +1 | Download statements. Finish cash application. Categorize remaining bank and card transactions. | Yes: categorization and matching are the strongest AI use case |
| +2 | Reconcile every bank, card, and loan account to a zero difference. Reconcile AR and AP subledgers to the GL. | Yes for matching; a person signs off on each reconciliation |
| +3 | Book accruals, prepaids, depreciation, deferred revenue, and payroll accruals. Record inventory adjustments. | Limited: recurring entries yes, judgment entries no |
| +4 | Flux review: compare each P&L and balance sheet line to last month and budget. Explain material swings. | Yes for first drafts of explanations; a person verifies each |
| +5 | Final review, close the period in the ledger, send the reporting pack. | Yes for formatting the pack; no for approval |
Where AI saves real time
Bank and card matching (day +1 and +2)
This is the clearest win. Matching a bank line to a recorded transaction is a pattern task with a right answer, and the ledger vendors have put most of their AI effort here. QuickBooks Online now lets you upload a statement for AI-powered reconciliation; Intuit's help article says the check can take up to an hour, and the goal is still the same as it always was: keep matching until the difference is $0.00. Xero's JAX does the matching against live bank feeds.
What you still do by hand: look at what the tool could not match, and look hard at anything it matched with low confidence. A reconciliation that ties to zero because someone forced a plug is worse than one that does not tie, since nobody goes looking for the problem.
Chasing receipts and documents (day -3 to +1)
Asking staff for receipts is a task nobody wants and everyone forgets. An agent that reads the card feed, spots transactions with no receipt, and sends a short message to the cardholder is cheap to run and saves hours of nagging. Xero lists document chasing as coming soon. You can build the same thing with a script and a mailbox.
First drafts of flux explanations (day +4)
Give a model last month, this month, the budget, and the transaction detail behind each line, and it will write a decent first draft: "Software expense up $2,100, driven by the annual renewal of X billed in October." That saves the controller from starting at a blank page. It does not save the controller from checking, because the model will also confidently explain a swing that is really a miscoding.
Where AI breaks
Accruals
An accrual is a guess about an expense you incurred but have not been billed for. The facts that drive it live in people's heads: the contractor who finished the job on the 28th, the utility bill that always arrives on the 9th, the legal work that is running over. A model can book a recurring accrual that repeats each month. It cannot know the job finished on the 28th unless someone tells it. Ask the people who know, on day -3, and book the entry yourself.
Judgment entries
Inventory reserves, bad debt reserves, revenue recognition on a contract with odd terms, capitalizing versus expensing a purchase: these are calls that your CPA may have to defend. Use AI to gather the support. Keep the decision and the entry with a person who will sign their name to it.
Categorization drift
AI categorization learns from your corrections. That is good until someone corrects it wrong, and then every similar transaction follows the wrong rule for months. Pull a sample each month of auto-categorized items in your largest expense accounts and spot-check them. Vendors quote high auto-categorization rates (Puzzle's pricing page says 85 to 90% on its Starter plan). The 10 to 15% left over is where the errors sit, and the errors in the 85% are the ones nobody sees.
The "close is done" claim
No tool knows that your close is done. It knows that its own checks passed. Closing the period in the ledger should stay a manual step, done by a named person, after the balance sheet review. Lock the period once closed so nobody can post into it by accident.
Controls to keep no matter what you automate
- One owner per task. Every line in the close list has a name next to it. "Finance" is not a name.
- Preparer and reviewer are different people on every reconciliation, when headcount allows. If you are solo, have your CPA review quarterly.
- Reconciliation support saved each month. Statements, rec reports, and the support for each journal entry, filed by period. The IRS says to keep income tax records for at least three years in the default case, and employment tax records for at least four years after the tax is due or paid, whichever is later. Many firms keep more.
- No vendor bank-detail changes during close without a call-back to a known number.
- Period lock after sign-off.
How to start
Write your close down as a list before you buy anything. Time each task for one month. Then automate the two that took longest and had the clearest right answer, which for most small businesses means bank matching and receipt chasing. Leave accruals and the final review alone. Measure the close again next month. If day +5 slips to day +8 after you add a tool, the tool added review work and you should know that.
Get the checklist
The Month-End Close Checklist ($19) is a working Excel workbook with this day-by-day task list from day -3 to day +5, owner and status columns, a reconciliation tracker, and an "AI can help" column so you can mark which tasks you have automated.
If you would rather have this built than build it, Agor AI sets up AR, AP and close automations for small finance teams. See Agor AI. Start with the $250 AI Automation Audit, credited toward a build.
This is part of a weekly series on AI for finance and back-office work. Get it by email.
This article is educational and is not accounting, legal or tax advice. Talk to your CPA before changing how you close your books or file.