Login Book a demo
← Back to blog

Automating bank reconciliation with AI: what to ask your software vendor in 2026

Francesco Domizio 7 min read

Bank reconciliation is the accounting task that looks easiest to automate and breaks most often in practice. The reason is simple: traditional rules assume that one payment matches one invoice. In the reality of any SME, a single transfer covers five invoices, one supplier withholds tax, another adds a fee, and a customer pays a 30% deposit.

This article separates what AI in bank reconciliation does well today from what remains review work, and gives you the eight concrete questions to evaluate a vendor without being sold smoke.

Why bank reconciliation is still broken in 2026

Traditional accounting software reconciles with three rules: equal amount, close date, description containing some text. It works on 60-70% of movements. The other 30-40% lands in a queue the accountant reviews monthly by hand.

The real problem is that the 30% represents 80% of the time. A simple movement (“payment of €1,250 to a utility”) reconciles in a second. A complex movement (“transfer of €4,873 to Stock Supplies covering four invoices with two early-payment discounts applied to two of them”) can take 15 minutes. And it is the complex ones that pile up.

AI changes the game in those cases. Not by magic. By context.

The 4 cases that break traditional reconciliation

1. Partial payments

The customer pays 50% on signing and 50% on delivery. There is one invoice and two bank movements. Traditional reconciliation sees “amount does not match” and sends to a human. Serious AI understands that the sum of the two movements against the right customer matches the invoice and proposes the partial reconciliation automatically.

2. Grouped payments

A customer pays six invoices with a single transfer. The bank reference says “transfer 230102”, which does not help. Traditional reconciliation finds no invoice with that single amount. AI looks at customer history, age of open invoices, and proposes the group of invoices whose sum matches. If there is ambiguity (two possible combinations), AI escalates with both options instead of sending the human to investigate from scratch.

3. Bank fees

A €1,000 transfer arrives as €996 after fees. Exact-amount reconciliation does not catch it. AI trained on your history learns that transfers from Bank X systematically arrive €4 short and applies the adjustment to the correct GL account without asking permission every time.

4. Withholdings and taxes

The customer pays the invoice minus a withholding tax. The bank movement amount never matches the invoice amount, but the difference is predictable. Serious AI identifies the withholding by supplier and percentage, marks the movement as reconciled, and leaves the withholding entry proposed.

How AI improves each case

The common pattern is context. Rule-based traditional AI has no memory of the customer or supplier. AI with history knows which supplier applies early-payment discounts, which adds intermediary fees, which fragments payments. It learns from your corrections, not because someone programmed a rule.

The reasonable minimum threshold in 2026 is 90% automatic match on your usual bank movements. Below that, you are still working too much in review. Above 95%, the remaining cases are genuinely complex and human review adds value.

Bank connection: PSD2 vs CSV vs proprietary APIs

Reconciliation quality also depends on how bank data arrives. Three options:

1. PSD2 (Open Banking). The European regulation requires banks to expose movements via secure API. It is the standard for modern reconciliation. Connection is direct, data arrives near real-time, and you do not download anything monthly.

2. CSV / national bank file standards. The traditional method. You download the statement and upload it to the software. It works but introduces hours-or-days delays, and each bank’s format varies enough that parsing is constant maintenance work.

3. Proprietary APIs. Some banks offer richer APIs alongside PSD2, with extended descriptions and internal references not exposed in the PSD2 minimum.

If your reconciliation software only supports CSV, you are starting with a data-quality deficit. If the sales rep tells you “but PSD2 is limited”, what they hide is that their product has not integrated it yet.

The 8 questions to evaluate AI bank reconciliation software

Print this list. Ask in order at every demo.

1. What percentage of my movements reconciles automatically?

Have them test on a real month of your statements, not their catalog demo. Expect 88-95% on standard SMEs.

2. How does it handle partial payments?

Specific question: “if a €10,000 invoice is collected in two tranches of €4,000 and €6,000, does the system propose automatic partial reconciliation or force me to mark it manually?”. Expect the first.

3. How does it handle grouped payments?

Inverse question: “if a customer pays 5 invoices with one transfer, does the system propose the group or only reconcile if the amount matches one invoice?“.

4. Does it learn from customer or supplier history?

Concretely: “if I correct manually that Bank X transfers arrive €4 short due to fees, will the system apply that rule automatically next time?”. If no, it is not AI; it is disguised if-then rules.

5. Does it handle withholdings automatically?

“If a customer pays an invoice with a withholding tax, does the system recognize it as withholding and leave the entry proposed?”. The right answer is yes.

6. Does it connect via PSD2 or only CSV?

If only CSV, ask when they will fix it. If PSD2 is available, ask which banks: large national banks are mandatory, smaller and cooperative banks are nice-to-have.

7. How is each reconciliation documented?

Audit will require traceability. Every match should log: which movement, which invoices, what confidence level, who approved (human or AI), and what data the AI used to decide. If they tell you “the log is internal”, be wary.

8. What happens when it gets it wrong?

Direct question: “if AI mis-reconciles an invoice, how do I correct it and how does it learn from my correction?”. Right answer: there is an explicit feedback loop. Wrong answer: “mark as wrong and redo”. That is not learning.

Edge cases no software solves well (yet)

Honesty: there are three cases where AI in 2026 is not at the level of a competent accountant.

1. Internal transfers between own accounts. If you transfer between your accounts and then pay a supplier, AI may confuse the source of the funds. Current solution: mark internal accounts explicitly.

2. Aggregated platform settlements (Stripe, PayPal, Amazon). The monthly aggregate deposit mixes dozens of individual transactions with fees and refunds. AI reconciling against your Stripe is close but not all the way. Solution: native Stripe/PayPal integrations rather than going through the bank.

3. Multi-currency intra-EU operations. An invoice in USD paid in EUR introduces FX differences AI may confuse with discounts. Still requires review.

If a vendor says “we do everything, including Stripe and currencies, no intervention”, be skeptical. They are promising it, but the technical reality of 2026 says these cases still need human judgment.

How Calitem approaches it

Calitem reconciles automatically by connecting to your bank via PSD2. AI matches partial payments, grouped payments, fees, and withholdings, the four cases where traditional reconciliation breaks. Each decision retains its confidence level, the data it used, and the audit trail to defend in an inspection.

Where we do not arrive: internal transfers without marking the internal accounts, reconciliation against payment platforms (in roadmap), and FX differences in multi-currency operations (also in roadmap). If any of those three is central to your operations today, say so in the demo and we explain what we do and what we do not.