Bank reconciliation with AI: matching and exceptions

What bank reconciliation with AI means

Bank reconciliation with AI means software proposes the matches between bank statement lines and ledger entries, including the hard ones: a payment covering several invoices, a card settlement net of fees, or a reference typed differently. It also drafts entries for charges and sorts exceptions. It matters because people then review proposals and investigate exceptions instead of ticking every line.

What AI does not change is the reconciliation itself. The book balance and the statement balance must still agree after timing differences, every adjustment still needs evidence, and a reviewer still signs. Our guide on bank reconciliation step by step covers that method with a full worked reconciliation. This guide covers the automation layer on top of it: the data you feed in, the rules that should run before any model, where AI adds value, and the controls that stop a clever matcher from hiding a problem.

Rules first, AI second

Most bank lines in a typical business match on simple rules, and rules are cheaper, faster and fully explainable. A matching engine should run deterministic rules first and send only the leftovers to a model. Using AI for lines that a rule can match adds cost and a small risk of error for no benefit.

The leftovers are where AI helps: remittance text written by a person, a customer paying three invoices with one transfer, a reference with the invoice number mistyped, or a supplier refund that looks like a customer receipt. A model reads that text the way an experienced bookkeeper would, and it can explain why it paired two items, which a rule cannot.

• Rules do well: exact amount and exact reference, recurring direct debits, fixed bank fees, transfers between the business's own accounts.
• AI does well: free-text remittance, partial and combined payments, near-miss references, classifying unmatched lines by likely cause.
• Neither should decide: whether an unexplained payment is legitimate, whether to write off a difference, or whether to post into a closed month.

Bank statement formats: CSV, PDF and CAMT.053

The quality of automated matching is set before any matching happens, by the statement format. CSV exports have no common standard: columns differ by bank, debits may be negative numbers or a separate column, dates may be day-month or month-day, and European exports often use a decimal comma. A single misread column turns every amount into nonsense, so fix the mapping once per bank and test it.

PDF statements are designed for people. Extracting them, with AI or otherwise, is an interpretation of a printed page, so never match from a PDF extraction until it passes an integrity check: the opening balance plus credits minus debits must equal the closing balance printed on the statement, to the cent. If it does not, a line was missed or misread.

Structured ISO 20022 statements are the better input where your bank provides them. Deutsche Bank's overview describes camt.053 as the end-of-day account statement that replaces the older MT940, camt.052 as the intraday report that replaces MT942, and camt.054 as the notification used for batch bookings and returns. These messages can carry standard references such as the end-to-end identifier the payer set, structured remittance information and ISO bank transaction codes, and camt.054 gives the detail behind batch bookings so that the individual payments can be reconciled. When the payer's reference arrives in its own field, far fewer lines need AI at all.

In Canada the same principles apply to the CSV and spreadsheet exports many businesses use: map columns once per bank, check the date convention, and run the balance check before matching.

Matching rules, in the order they should run

Order matters because each pass removes lines from the pool and makes the next pass safer. A fuzzy match run first can pair a line that an exact rule would have matched to a different, correct item. Every match should carry its rule name and, for AI matches, a stated reason and a confidence level, so the reviewer knows how each pair was made.

• Pass 1, exact: same amount and the invoice or payment reference found in the statement text.
• Pass 2, amount and date window: same amount, same counterparty, within a few days, and only one candidate.
• Pass 3, one-to-many: one receipt equals the sum of several open invoices of the same customer.
• Pass 4, many-to-one: one settlement equals a group of card or gateway sales less fees, checked against the processor's settlement report.
• Pass 5, own transfers: a debit in one of the business's accounts and a credit of the same amount in another.
• Pass 6, AI proposals: fuzzy references, name variants and remittance text, each with a reason and a confidence level.
• Pass 7, classification: remaining lines sorted into likely causes (bank charge, interest, unknown receipt, unknown payment) for a person.
• Never: matching on amount alone when more than one open item has that amount.

Worked example: AI proposals reviewed by a person

A Canadian distributor reconciles its operating account for September. Rules match most lines; six are left for the AI pass. The assistant returns a proposal for each, with its reason, and the accountant reviews them one by one. All amounts are in Canadian dollars.

Of the six, three proposals are accepted, one is re-matched, one is rejected and escalated, and one line gets no proposal and is parked. The model's reading of the text is good on every line; the two proposals it gets wrong were a coincidence of names and a payment it should never have been asked to justify, which is why a person reviews every AI match and why unexplained payments go to someone with authority to investigate.

• Credit 4,520.00, text NORTHWIND LTD INV1041/1043. Proposed one-to-many match to invoice 1041 for 2,712.00 and invoice 1043 for 1,808.00; 2,712 plus 1,808 is 4,520. The reviewer checks the customer's remittance advice. Accepted.
• Credit 11,640.55, card settlement. Proposed match to card sales of 12,000.00 for 28 September less fees of 359.45, with the entry Dr Bank 11,640.55 / Dr Card processing fees 359.45 / Cr Card clearing 12,000.00. The reviewer agrees the fee to the processor's settlement report. Accepted.
• Credit 1,980.00, text E-TRANSFER J MARTIN. Proposed match to invoice 1052 of Martin Consulting Inc. for 1,980.00. The reviewer sees that invoice 1049 of Jean Martin, a sole proprietor, is also open for 1,980.00 and finds her remittance e-mail. Re-matched to 1049.
• Debit 45.00, monthly plan fee. Proposed Dr Bank charges 45.00 / Cr Bank 45.00. Accepted, and turned into a rule so it never reaches the AI pass again.
• Debit 3,150.00, text BILL PAYMENT PACIFIC SUPPLY CO. Proposed match to bill 2231 of Pacific Supplies Ltd for 3,150.00. The reviewer finds the payment is in no approved payment run, the payee name differs from the supplier record and bill 2231 is still awaiting approval. Rejected and escalated to the controller and the bank as a possible unauthorised payment.
• Credit 600.00, text DEPOSIT, no reference. No proposal; three customers have open balances of 600.00. Parked as an unidentified receipt, Dr Bank 600.00 / Cr Unallocated receipts 600.00, with an owner and a date to clear it.

Working the exceptions

An exception is any statement or ledger line that no pass could match with confidence. Good automation does not hide them; it sorts them so the right person sees the right problem. Four causes cover almost everything, and each has a different owner and a different fix.

Two habits keep exceptions honest. Age every open item, and review anything older than one month by name, because a reconciling item that survives three statements is usually an error, not a timing difference. And never close a gap with an unexplained adjustment. A reconciliation that balances only because of a plug has hidden the problem the process exists to find.

• Timing: cheques issued but not yet presented and deposits recorded but not yet credited. They stay on the list and must clear on next month's statement.
• Missing from the books: bank charges, interest, direct debits and card fees. These become proposed entries.
• Errors: a transposed amount in the cash book, a payment posted to the wrong bank account, or occasionally a bank error. Each is corrected at source with evidence.
• Unknown or suspicious: receipts with no identifiable payer and payments nobody approved. These go to a named owner, not into an adjustment.

Proposed entries and who approves them

Some lines need a new ledger entry, not a match. AI is useful for drafting these because it can read the statement text and propose the account, but the entry remains a draft until a person approves it, and the approval rules should be stricter than for ordinary matches because a new entry changes the books.

A practical split: recurring, low-value items such as monthly fees and interest can be converted into bank rules once a person has approved the first instance, after which no AI is involved. Everything else is drafted with its evidence and approved individually. Unidentified receipts go to an unallocated receipts account with an owner, never straight to revenue or to a customer chosen by guesswork, because a wrong allocation sends a false statement to two customers at once. Nothing is ever posted into a period that has been closed; a late-discovered item belongs in the current period with a note, or is handled as a correction under your error policy.

Controls over AI-assisted reconciliation

Automation moves the reviewer's job from checking every line to checking the design, the exceptions and a sample. That is only safe if the controls are explicit and written down, because an auditor testing the reconciliation will want to see how an AI match was made and who accepted it. Auditors may also confirm balances directly with the bank under ISA 505, and ISA 500 asks them to consider how reliable the information behind the reconciliation is.

• Separation: the person who can release payments does not approve the bank reconciliation.
• Balance roll: the reviewer re-performs statement balance, less outstanding items, equals book balance, rather than trusting the tool's total.
• Sample of automatic matches: each month, re-check a handful of rule and AI matches against source documents.
• Versioned rules and prompts: changes to matching rules, thresholds or the model are recorded and tested before use.
• Confidence floor: AI matches below an agreed confidence level are shown as suggestions only and cannot be accepted in bulk.
• No auto-posting: every new entry is approved by a named person, and approvals are logged.
• Timeliness: reconcile at least monthly, and daily for high-volume accounts, so exceptions are fresh enough to investigate.

Security and data protection

Bank statements are full of personal and sensitive data: names of customers and employees who pay or are paid, account numbers, and payment descriptions that can reveal salaries or medical providers. Under the GDPR in Europe you need a lawful basis, a processor agreement with any service that handles the data and a view on international transfers. Under PIPEDA, Canada's federal private-sector privacy law, you stay accountable for personal information you transfer to a service provider for processing and must protect it by contract or other means. Mask account numbers the model does not need, send only the lines left after the rules pass, and keep the statement file itself in your own systems. This is general information, not legal advice.

Remittance text is written by outsiders, so it is an injection route. A payer can put a sentence in a reference field that tries to instruct an AI matcher. Anthropic's guidance for indirect prompt injection is to deliver third-party content as data, tell the model where it came from, state in the instructions that such content never overrides them, and apply least privilege. In a reconciliation that means the matcher can read lines and propose matches, and it has no tool that posts entries or moves money.

Bank reconciliation AI tools: what to use when

Bank reconciliation AI tools fall into four groups: bank-rule engines inside accounting software, spreadsheets with an AI assistant, custom matching built on a model API, and ERP-native reconciliation with AI proposals. The right choice depends less on the model than on where your statements arrive and where the approved entries must end up.

For a spreadsheet route, as of September 2026, Claude for Excel answers questions about a workbook with cell-level citations, and Anthropic's Finance plugin for Claude Cowork includes a reconciliation command that compares ledger balances with bank or subledger balances; Anthropic states that its outputs should be reviewed by qualified professionals. For a custom build on the Claude API, the code execution tool runs Python in a sandbox with no internet access and can analyse CSV and Excel files uploaded through the Files API, which suits the arithmetic of matching better than asking a model to add numbers in its head. Structured outputs can force each proposal into a fixed shape, for example statement line, ledger items, rule or reason and confidence. Structured outputs guarantee the shape, not the correctness of the match.

Mind the data terms of each route. Anthropic's documentation lists the Files API and code execution as not eligible for zero data retention, with code execution container data kept for up to 30 days, and Claude for Excel as not currently eligible either. Anthropic states that by default it does not use inputs or outputs from its commercial products to train its models.

Bank reconciliation automation, step by step

Bank reconciliation automation works best when introduced in stages, with the old process running alongside until the new one has earned trust on your own data.

• Week 1: fix statement imports per bank and add the opening-plus-movements-equals-closing check.
• Week 1: write exact and own-transfer rules and measure how many lines they match.
• Week 2: add one-to-many and card-settlement rules; build a golden set of last quarter's lines with the correct matches.
• Week 3: run the AI pass on the leftovers against the golden set; record every wrong proposal and why.
• Week 4: go live in Draft mode only, with a named reviewer, a confidence floor and a monthly sample of automatic matches.
• Every month: turn repeated AI matches into rules, and review the exception list by age.

Bank reconciliation in Skyline Nexus ERP

In Skyline Nexus ERP, bank accounts are held in Treasury and synced into the chart of accounts, and the trial balance warns about any bank account that has not been synced. Treasury > Bank Reconciliation > New Reconciliation takes the account, the date range, the statement date, the statement reference and the statement ending balance, and shows the book balance and the last reconciliation alongside. Statements are imported as CSV, TXT, XLSX or XLS files of up to 10 MB, an Auto-match screen pairs lines, and the reconciler can mark items outstanding, add reconciliation adjustment lines, detect duplicates, merge items and print the result. For banks that deliver camt.053, use the bank's CSV or spreadsheet export as the import route.

AI bank matching with Claude, proposing matches and entries for approval, is being rolled out on the Skyline Nexus ERP roadmap, and direct bank feeds are being rolled out market by market. Tell us your bank and country and we will confirm your go-live date.