AI bank reconciliation
Also called auto-reconciliation, bank matching, bank feed matching, transaction matching, automatic reconciliation
How well does AI bank reconciliation match payments to invoices? What the research shows, what it costs, and the mismatches to watch for.
Automatic matching works well on simple, one-to-one payments, but we found no independent measurement of the time it saves. Reconciliation software proposes which invoice, bill or rule each bank line belongs to, and newer tools auto-match the lines they are confident about. In a study of 10 companies' records, a trained model matched 96% of payments correctly, against 71% for picking the closest amount. The hard cases are single payments covering several invoices. Vendors claim 30 to 70% time savings.
At a glance
| Status | In pilots |
|---|---|
| What it does | Pulls in bank feed lines and proposes a match to an open invoice, bill or bank rule, auto-confirming high-confidence matches and leaving the rest for a person. |
| Typical cost | Usually bundled in accounting plans: a R100 per month add-on to a R450 plan, or included from R795 per month, incl. VAT (vendor list prices, checked Oct 2026) |
| Data it needs | A live bank feed, open invoices and bills with accurate amounts and references, and your history of past matches and bank rules. |
| Setup effort | Low: connect bank feeds and switch it on; the effort is clearing old unreconciled items and getting customers to quote invoice numbers. |
| Evidence | Thin. Two academic matching studies, one co-written with an accounting software company, and a master's field trial; analyst surveys of finance leaders; time-saving figures are vendor claims. |
What is AI bank reconciliation?
Bank reconciliation is the job of ticking off every line on the bank statement against something in the books: a customer payment against an invoice, a debit order against a bill, a card fee against an expense account. Done by hand at month end, it is slow. Left undone, your debtors list is wrong, because customers who have paid still look overdue.
AI reconciliation proposes those matches for you. Cloud accounting systems have offered rule-based suggestions for years. The newer layer uses machine learning: it learns from how you reconciled similar lines before and, in at least one product, from how other users reconciled similar transactions. When it is highly confident it reconciles the line itself; when it is not, it leaves a suggestion for a person to confirm.
How does it work?
The software compares each bank line with candidates in the ledger and scores how likely each pairing is, using:
- Amount and date: the obvious signals, though many invoices share round amounts.
- Reference and description: customers type invoice numbers, names or nothing at all.
- History: how this payee was matched last time, and any bank rules you have set.
- Other users: some products also draw on how other businesses reconciled similar lines.
The result is an automatic match, a suggestion, or nothing. Some products let you switch automatic matching on or off for each bank account.
Where a bank has no direct feed, statements can be captured from PDF instead; one capture vendor charges three credits per statement page, against one per invoice. Card machine and online payment payouts, which arrive as one lump net of fees, usually need a clearing account before individual sales will match.
What it needs to work
It depends on the ledger already holding the invoices and bills each bank line should match, with references that customers actually use.
- A direct bank feed for every business account and credit card, or PDF statements captured monthly where no feed exists
- Every sales invoice and supplier bill raised in the accounting system before it is paid, each with a unique number
- Customers asked to quote the invoice number as the payment reference, printed on every invoice
- Bank rules for recurring items such as bank charges, debit orders, rent and payroll
- A clearing account for card machine and online payment payouts, so fees and gross sales can be separated
- A clean starting point: old unreconciled lines and stale open invoices cleared or written off before automatic matching is switched on
Most use cases in this guide succeed or fail here, on the data underneath them, before any AI is involved.
What does the evidence show?
The matching step has been studied. Researchers at Kyoto University, working with accounting software company freee, trained models on 21,122 invoice and payment records from 10 companies. Their best method matched 96% of payments correctly, against about 71% for picking the invoice with the closest amount. The study covered only one payment per invoice.
An RMIT University team, using three years of real bookkeeping data, reported that a graph-based model beat an industry-standard tool, with the biggest gains where one payment covers several invoices. A 2025 master's thesis designed and tested a mix of rules and machine learning in a Business Central bank feed module.
What is missing is independent evidence of hours saved. CPA.com's 2025 report notes vendors claim 30 to 70% time savings on bank reconciliation and similar tasks. Across finance AI generally, 91% of respondents to a 2025 Gartner survey of 183 finance leaders, not focused on small firms, reported low or moderate initial impact.
AI reconciliation vs manual reconciliation
| Compared on | AI reconciliation | Manual reconciliation |
|---|---|---|
| Speed | Confident lines reconcile with no clicks; a person handles the exceptions | Every line opened, searched for and matched by hand |
| Error profile | Systematic: plausible but wrong matches on round amounts or grouped payments | Random: missed lines, wrong customer, items parked in suspense |
| Cost model | Usually bundled in the accounting subscription or a small add-on | Bookkeeper time, often a month in arrears |
| Best for | Many repeat transactions with clear references | Few transactions, or messy payments that need judgement anyway |
Status and evidence
Rated pilots rather than production. Automatic and suggested matching is already built into standard small-business accounting plans, so adoption is not the gap; independent evidence of results is. Evidence is thin. The strongest matching study was co-written with an accounting software company and covered only one payment per invoice. The other studies measure matching accuracy rather than time or errors inside businesses, and the time-saving figures in circulation are vendor claims.
Where does it go wrong?
- Confident wrong matches. A payment for the right amount from the wrong customer gets auto-matched, and the real debtor drops off your chase list. Review a sample of automatic matches each week, not just the leftovers.
- One payment, many invoices. A customer pays several invoices in one transfer, or short-pays one. Grouped matches are harder, which is why current research focuses on them.
- Learning bad habits. The tool learns from how lines were reconciled before. If past items were forced into suspense or the wrong account, it repeats that.
- Old backlog. Months of unreconciled lines and stale open invoices create false candidates. Clear or write off the backlog first.
- Processor payouts. Card machine and online settlements arrive net of fees in one amount and will not match individual invoices without a clearing step.
- Borrowed patterns. Suggestions based on other businesses' behaviour may not fit your chart of accounts.
Numbers it should move
Ask the demo business
Northgate Supply is a fictional wholesale distributor of catering and packaging supplies: 28 staff, R46.6m in sales over the last year. Its data is generated for this demo, to 30 September 2026.
R1.20m is overdue across 30 customers. The three largest balances: Lighthouse Hospitality (R229 273), Silvertree Conference Centre (R174 389), Baobab Lodges (R104 409).
How this was worked out
Unpaid invoices as at 30 September 2026, grouped by days past their due date. Terms are 30 days from invoice. Amounts exclude VAT.
Where the data usually lives
Where the records this use case needs are usually kept in a small business. It is not a list of recommended tools.
Frequently asked questions
How much does automatic bank reconciliation cost?
It is usually part of your accounting subscription rather than a separate product. In South Africa, Xero includes auto-reconcile on its Standard plan at R795 a month and sells it as a R100 monthly add-on to Starter, prices including VAT. In the UK, the same vendor includes it on Grow and higher plans and offers an add-on for lower tiers. Check what your current plan includes before buying anything.
Do I need an IT person to set up bank feeds and auto-matching?
No. Bank feeds are connected from inside the accounting software by signing in to your bank, and automatic matching is a setting you switch on per account. The work is bookkeeping rather than IT: clearing the backlog, setting bank rules and checking the first weeks of automatic matches. If your bank has no feed, PDF statements can be captured instead.
Can I trust automatic bank reconciliation without checking it?
Not entirely. Products auto-reconcile only when highly confident and leave the rest as suggestions, but confident is not the same as correct. In the strongest academic test we found, about 4% of payments were still not matched correctly, and payments covering several invoices are harder. Review a sample of automatic matches weekly, starting with large amounts and customers who pay several invoices at once.
Is it legal under POPIA and UK GDPR to let software read my bank transactions?
Generally yes, but data protection rules apply because bank lines name staff, sole-trader suppliers and customers. Where a software provider processes that data on your behalf, UK GDPR requires a written contract binding the provider to you. Company information is not personal data in the UK, but under POPIA it can be, and POPIA requires a written contract obliging the provider to keep data secure. This is not legal advice.
Why does auto-reconciliation keep suggesting the wrong match?
Usually because of the data, not the model. If customers do not quote invoice numbers, several invoices share an amount, or past lines were coded inconsistently, the tool has weak signals to work with. Some products learn from how you and other users reconciled similar lines, so old mistakes repeat. Fix payment references, set bank rules for recurring items and clear stale open invoices.
More in cash & finance
Sources
- 1Automatic Bank Reconciliation Using Machine LearningJapanese Society for Artificial Intelligence (JSAI 2017, Kyoto University and freee K.K.), 2017Study
- 2Enhancing Bookkeeper Decision Support Through Graph Representation Learning for Bank ReconciliationKeAi Publishing (The Journal of Finance and Data Science), 2025Study
- 3Hibridinis finansinių operacijų gretinimo algoritmas ir jo tyrimas (hybrid financial transaction reconciliation algorithm, master's thesis)Kaunas University of Technology, 2025Study
- 42025 AI in Accounting ReportCPA.com (AICPA subsidiary), 2025Survey
- 5Gartner Survey Shows Finance AI Adoption Remains Steady in 2025Gartner, 2025Survey
- 6Bank Reconciliation Software for Small Businesses | Xero UKXero, 2026Vendor
- 7Pricing Plans | Xero ZAXero, 2026Vendor
- 8Pricing - Data Entry Automation Software | AutoEntryAutoEntry, 2026Vendor
- 9What is personal data?Information Commissioner's Office, 2026Government
- 10Protection of Personal Information Act 4 of 2013Government of South Africa, 2013Government
- 11When is a contract needed and why is it important?Information Commissioner's Office, 2024Government
This entry is educational. It is not legal, financial or professional advice, and it does not recommend any vendor. Figures attributed to vendors are their own claims.
- 1Automatic Bank Reconciliation Using Machine LearningJapanese Society for Artificial Intelligence (JSAI 2017, Kyoto University and freee K.K.), 2017Study
- 2Enhancing Bookkeeper Decision Support Through Graph Representation Learning for Bank ReconciliationKeAi Publishing (The Journal of Finance and Data Science), 2025Study
- 3Hibridinis finansinių operacijų gretinimo algoritmas ir jo tyrimas (hybrid financial transaction reconciliation algorithm, master's thesis)Kaunas University of Technology, 2025Study
- 42025 AI in Accounting ReportCPA.com (AICPA subsidiary), 2025Survey
- 5Gartner Survey Shows Finance AI Adoption Remains Steady in 2025Gartner, 2025Survey
- 6Bank Reconciliation Software for Small Businesses | Xero UKXero, 2026Vendor
- 7Pricing Plans | Xero ZAXero, 2026Vendor
- 8Pricing - Data Entry Automation Software | AutoEntryAutoEntry, 2026Vendor
- 9What is personal data?Information Commissioner's Office, 2026Government
- 10Protection of Personal Information Act 4 of 2013Government of South Africa, 2013Government
- 11When is a contract needed and why is it important?Information Commissioner's Office, 2024Government