AI Month End Close: How an Agent Matches the Bank Before Your Team Arrives
An AI month end close starts at the bank: an agent pulls every bank, card and payment provider feed at 5am, matches each line to the ledger in four passes from exact match to reading the remittance email, and leaves the controller a ranked list of 30 to 80 exceptions instead of 3,000 lines. It proposes and a person posts, and because it runs every morning rather than once a month, the bank is about 95% reconciled before month end even starts.
Why the bank is where most closes lose 2 days
In most finance teams I walk into, bank reconciliation happens once, on working day 1 or 2, for the whole month. One person sits with 2,000 to 4,000 lines across 3 to 6 accounts and ticks them off. The ledger's own auto-match catches the easy half. The other half takes 2 days, and it's the half where the mistakes live.
A duplicate supplier payment that lands on the 29th is the line a rushed rec misses. Until someone finds it, the ledger and the bank disagree by the full amount, and the board pack carries a cash figure nobody can explain. When the bank is late, every number that depends on cash is late with it: the runway figure, the forecast, the covenant check.
How I build the overnight bank matcher
1. Get every feed in by 5am, and check it adds up
The agent needs yesterday's lines from every account before anyone logs on. That means direct bank feeds where the bank offers them, daily statement files where it doesn't, and the payout reports from Stripe, Adyen or whichever card provider you use, since those carry the fee and refund detail the bank line hides. Before it matches anything, the agent checks that today's opening balance equals yesterday's closing balance on every account. If one doesn't, it stops on that account and says so. A missing day of lines will make every later match look wrong.
2. Match in four passes, strictest first
Each pass only sees what the pass before it left. The order matters, because a loose rule run first will grab lines a strict rule would have matched properly.
| Pass | What it matches | Share of lines I typically see | Auto-accepted? |
|---|---|---|---|
| 1. Exact | Same amount, date within 2 days, reference matches | 60 to 70% | Yes |
| 2. One to many | One receipt to several invoices, one payout to many sales net of fees | 15 to 20% | Yes, if the totals tie to the penny |
| 3. Tolerance | Right payer, amount off by bank charges or FX inside a set limit | 5 to 10% | Proposed, controller accepts in bulk |
| 4. Reading | Model reads remittance emails, bank narrative and open items, proposes a match with a reason | Most of the rest | Never |
Only pass 4 uses a language model. Passes 1 to 3 are plain rules, and I keep them that way because rules give the same answer every time and an auditor can read them. The model earns its place on the 5 to 10% of lines where the answer sits in an email from a customer saying "this covers invoices 4471 and 4480, less the credit note".
3. Rank the exceptions by value and age
Whatever is left goes into one list, sorted by amount and then by how many days it's been open. Each line carries the agent's best guess and a one line reason. On a 100 to 300 person company that list is 30 to 80 lines a day at the start, and it falls as the rules improve. The controller works it from the top at 8am. The 4 lines over £10,000 get looked at first, and the £3.20 card fee can wait.
4. Turn corrections into rules
When the controller fixes the same kind of proposal the same way twice, the agent drafts a new pass 1 or pass 3 rule and the controller approves it. Expect the rules file to grow fast for the first 3 months and the daily exception list to shrink with it. I don't let the model learn quietly. Every rule is written down with a name, a date and who approved it, and the CFO reads the list once a month.
5. Keep the agent's hands off the money
The agent has read access to the bank and the ledger. It writes proposals into a review queue and nothing else. A person posts. The CFO sets the tolerance in pass 3, usually £25 or 1%, whichever is lower. Any change to a supplier's bank details goes to a person with a phone call, every time, because that's where fraud comes in. I run it in shadow for 2 full closes first, comparing its answer with the controller's on every line, before it counts.
Why daily matters more than speed
The 5am run is the whole trick. A team that reconciles once a month does 30 days of work on day 1. A team whose agent matches every morning does 1 day of work on day 1, because the other 29 were done on the mornings they happened. Month end stops being a bank event. On a 10 day close that's 1.5 to 2 days back from the bank alone, and the cash figure in the board pack matches the bank on day 1. The wider close, accruals, commentary and chasing, is in how to automate month end close with AI agents.
From my operating seat
I'm Hayat Amin, an AI-native CFO, AI operator and AI coach for CFOs who teaches finance leaders how to build AI into finance operations. I spent twenty years in the C-suite, with three exits and three FT100 listings. Now I build and run AI agent systems inside companies, across finance, sales pipeline, onboarding and compliance evidence, and I coach CFOs one to one on the same work.
Inside one client I run, the matcher finishes before 6am and the controller gets a note at 7am: how many lines came in, how many matched, and the exceptions that need a decision today. The same matched balance feeds the live cash runway figure the CFO reads every morning, so nobody argues about which cash number is right. I've been through three exits, and a buyer's diligence team always asks for the bank recs early. A rec that runs every day with a log behind it is an easy thing to hand over.
How long does month end close take with AI?
For a 100 to 300 person company, 4 to 5 working days is a realistic target once the bank, accruals and chasing run on agents, against the 8 to 12 I usually find. The bank alone moves from 2 days to half a day. Most of the rest comes from inputs arriving on time, which is a chasing problem, not an accounting one.
Can AI do bank reconciliation in Xero or QuickBooks?
Both already suggest matches and let you write bank rules, and that covers the exact pass well. What they don't do is match one payout to 140 card sales net of fees, read a customer's remittance email, or work across 6 entities at once. I run the agent on top through the API, reading from the ledger and writing proposals into a review queue, never posting.
What does AI bank reconciliation still get wrong?
Three things in the ledgers I've worked on: refunds netted inside a payout, payments from a parent company for a subsidiary's invoice, and anything with a reference a person typed by hand. The agent gets these partly right with high confidence, which is the dangerous kind of wrong. That's why pass 4 never auto-accepts, whatever score it gives itself.
Where I come in
This is the work I do with CFOs one to one and inside finance teams. In the first session we pull a month of your bank lines and run them through the four passes, so you see your own match rate before anything is built. The other agents worth building beside it are in AI agents for finance. The map of the whole topic is at AI for CFOs, the end state is at the AI-native CFO, and the sessions are at coaching for CFOs. How many days does your bank rec take now? Start at meethayat.com and tell me.