AI for bank reconciliation: the finance-team process

What you’ll end up with

A process you can run every week or month without reinventing it. Your accounting software does the matching it’s good at. A general AI assistant helps with the leftovers, the queries and the spreadsheet. You approve, post and keep the evidence.

By the end you’ll have on file: the software’s matches, a stripped list of leftovers, a comparison workbook, a completed checklist and a signed-off reconciliation. Nothing in that list needs bank data sitting in a chatbot.

The steps use Xero, because that’s where the menu names below come from. The process works the same in other accounting software; the buttons will be called something else.

Before you start

Three things need to be true first.

Your bank feed or statements are imported. Every transaction for the period is entered, not just the ones you’ve got round to coding. And you have an AI tool your organisation has approved for this kind of work, because ICAEW’s line is to keep client and confidential internal data off public AI tools and to review outputs with professional scepticism.

Also check the opening balance, and check the statement balance in Xero matches the actual bank. Xero’s own guidance puts both before you reconcile anything. If they’re off, nothing downstream will be right, and no prompt will fix it.

Steps

1. Let the software do the matching first

Imported statement lines show on the left of the Reconcile tab. For each one you either match it to an existing transaction or create one. Xero suggests a match by comparing the statement line with the transactions in the account, and if several match it looks for the one with the closest due date. If the suggestion is right, accept it.

Nothing suggested, but you know the transaction is there? Use find & match to search for it.

Bank rules are for recurring or similar lines that you won’t create a bill or invoice for. Set one up and Xero suggests the same coding each time. If a line relates to an existing transaction, don’t use a rule: use find & match.

If you’re on the Grow plan or above, or you’ve added it to a smaller plan, there’s automatic bank reconciliation too. A subscriber or administrator turns it on per bank account. JAX then reviews future statement lines (not historical ones) and reconciles the ones that meet its high-confidence criteria. Each gets a label: Rule, Match, Memory or Prediction. Hover over the label to see JAX’s reasoning. Review them on the Reconciled transactions screen, and if one’s wrong, fix it; JAX learns from the correction.

Checkpoint: everything the software could match has been matched or accepted by you, and you’ve reviewed any automatic reconciliations.

2. Export the leftovers

If you have the administrator role, you can list the unreconciled lines: in the Accounting menu, select Bank accounts, click the menu icon at the top right, then select Uncoded statement lines. Choose the bank account and dates, click Run, then Export and pick CSV (PDF works too, but you want the CSV).

The export has a Your Comments column. Keep it. It’s where you’ll note what each line turned out to be.

Checkpoint: one CSV covering the bank account and period you’re reconciling, with a row count that matches what you saw on screen.

3. Strip it before AI sees anything

This is the step people skip, and it’s the one that matters most.

Make a copy of the CSV. In the copy, keep the date, an amount band (say “under £100” or “£1,000 to £5,000”) and the description text. Delete everything else. No account numbers, no sort codes, no names of customers, suppliers or staff. If a description has a person’s name in it, replace it with [PERSON].

You don’t need the exact figure for AI to say “that looks like a subscription”. A band is enough. The exact numbers stay in your accounting software, where they belong.

Checkpoint: read the stripped file top to bottom. Could anyone identify a person or an account from it? If so, strip it again.

4. Ask AI what each leftover might be

Ask for questions, not answers. You want a list of what each line might be and what evidence would settle it.

Below are bank statement lines I can't yet identify. Each has a date, an amount band and the description text only.

For each line, give me: (1) up to two things it might be, (2) the question I should ask or the document I should look for to confirm it, and (3) whether it looks related to another line in the list.

Do not tell me how to code anything. Do not assume any suggestion is right. If a description is too vague, say so.

[PASTE STRIPPED LINES HERE]

It may spot that several lines look like the same payee under slightly different reference text. That’s a pattern to chase, not a coding decision.

Checkpoint: every line has a next action (find the invoice, ask someone, check the contract). No line has been coded on the strength of the AI’s guess.

5. Draft queries to budget holders

Some lines you can’t clear without the person who knows the spend. Let AI turn your rough notes into clear questions.

I'm reconciling a business bank account and need to ask budget holders about unexplained payments. Write a short, polite query for each item below. Each query should give the date, the amount band and the description text, ask what the payment was for and which cost centre or project it belongs to, and ask for any supporting document. Four sentences or fewer each. Plain UK English.

Items:
[DATE, AMOUNT BAND AND DESCRIPTION FOR EACH ITEM]

Add the exact amount when you send the email yourself, from your own work account. The AI saves you the wording, not the sending.

Checkpoint: every query is sent, logged in the Your Comments column and has a chase date. You read each one before it went.

6. Build a bank CSV vs ledger comparison in Excel

Export the ledger transactions for the bank account and period. Put the bank CSV on one sheet and the ledger export on another. Then give the AI the column headings, not the data.

I have two Excel sheets. Sheet "Bank" has columns: [BANK COLUMN HEADINGS]. Sheet "Ledger" has columns: [LEDGER COLUMN HEADINGS]. Dates are in [DATE FORMAT]. In "Bank", receipts are positive and payments negative; in "Ledger", [HOW AMOUNTS APPEAR].

Write formulas to add a "Status" column on the Bank sheet that: (1) uses XLOOKUP to find a ledger row with the same date and amount, (2) uses COUNTIFS to count how many ledger rows share that date and amount, and (3) shows "UNMATCHED" where there is no match and "CHECK" where there is more than one. Then add the same check on the Ledger sheet for rows with no bank match. Explain each formula in one sentence. Use cell references only, no typed numbers.

XLOOKUP returns the item for the first match it finds, which is why the COUNTIFS column matters: it tells you when there’s more than one candidate. Test on a handful of rows you can check by eye before you trust it on the lot.

Checkpoint: the workbook flags the unmatched rows on both sheets, and you’ve checked several matched and unmatched rows by hand.

7. Write the working-paper checklist with AI

Checklists suit AI well: structure and words, no numbers.

Draft a one-page checklist for a monthly bank reconciliation working paper in a UK finance team using [ACCOUNTING SOFTWARE]. Include: statement balance agrees with the software; Bank Reconciliation Detail report run; every unreconciled item explained; no AI suggestion posted without evidence; queries to budget holders answered; bank rules reviewed; preparer and reviewer sign-off with names and dates. Add a tick box and a notes space for each. Short, plain wording.

Adjust it to your own standards, then keep it with the workings.

Checkpoint: one checklist, completed for this reconciliation, not a template left blank.

8. Approve, post and keep the evidence

This part is yours. AI has no role in it.

Check each match and each coding decision against the evidence. Then run the Bank Reconciliation Detail report, which compares your actual bank balance with the balance of the bank account in Xero. If they don’t agree, check for missing, deleted or duplicated transactions.

Found a line reconciled to the wrong transaction? Unreconcile it, or use Remove & Redo, then reconcile it again correctly.

Then file the pack: the exports, the comparison workbook, the completed checklist, the replies from budget holders and the report. The preparer signs and dates; the reviewer signs and dates.

Checkpoint: the balances agree, every unreconciled item has an explanation, and both signatures are on the checklist.

Common mistakes

• Pasting raw statements into a chatbot. It’s the obvious one, and it’s why step 3 exists.
• Treating the AI’s guess as the coding. A suggestion is a lead. The evidence is the invoice, the contract or the budget holder’s reply.
• Using a bank rule where an invoice exists. If a payment is reconciled as a spend or receive money transaction instead of against the invoice or bill, the invoice still shows as awaiting payment and you can end up with a duplicate. Match it instead.
• Accepting automatic reconciliations without looking. JAX shows its method and reasoning for each one so you can review it. Use that, and correct mistakes so it learns.
• Letting AI do the totals. It writes formulas; the ledger and the bank supply the numbers. Never type in a total an AI gave you.

When to keep AI out of it

Approving a reconciliation. Posting a transaction. Making a payment. Anything with bank details or names in it.

That’s the short list. The AI isn’t the decision-maker and it isn’t a vault, so it gets descriptions, headings and drafts, and nothing else. The approval and the sign-off stay with you.