Can AI do bank reconciliation? Deterministic matching, with AI explaining the exceptions.

The short answer

AI can help a great deal with reconciliation, but it shouldn’t make the match or the decision. The match should be deterministic: written rules that give the same answer on the same evidence every time, so you can show an auditor why two records were paired. AI is useful around that core: reading messy remittance text, investigating the items the rules couldn’t match, and drafting an explanation for a person to accept or reject.

Put simply: deterministic matching, and AI that explains the exceptions.

What “AI reconciliation” can mean

Vendors use the phrase for three quite different things, and it is worth asking which one you are being sold:

• Rules-based automation described as AI. Matching rules and tolerances, sometimes suggested from past matches. This is deterministic, and usually what you want.
• Statistical match suggestions. A model scores likely pairs, and a person confirms them. Useful for messy data, as long as nothing is matched without a person or a rule accepting it.
• Language-model agents. A model reads statements, emails and files, and proposes matches, explanations or actions. Powerful for investigation; risky if it is allowed to decide.

Why the match itself should be deterministic

• Repeatable. The same evidence under the same rules must give the same result. A language model can answer differently on a second run.
• Explainable. Every match should point to the rule and version that made it. “The model thought so” is not an audit trail.
• Governed. Tolerances are policy decisions. A rule that accepts a 0.50 fee difference should be written down, owned and logged each time it is used, not learned silently.
• Safe to re-run. When a rule changes, you should be able to re-run it over the same days and see exactly what changed.
• Resistant to injection. Remittance text and emails are data from outside your business. A model that acts on what it reads can be steered by that data.

Where AI genuinely helps

• Reading unstructured text. Pulling an invoice number or payout reference out of a free-text narrative, as a suggestion the rules can then check.
• Investigating. Gathering the provider’s record, the file line and the bank entry for an exception, and requesting a status lookup or a file re-fetch through the same controlled commands a person would use.
• Explaining. Drafting a plain-language explanation of what probably happened, with the evidence it relied on.
• Spotting patterns. Noticing that forty exceptions share one cause, such as a changed cut-off, and proposing a rule change for a person to review.

Where AI must not decide

Some actions change money, records or policy, and should always need a person with the right permission:

• closing an exception or accepting a settlement;
• writing off a difference;
• changing matching rules, tolerances or posting rules;
• moving money, or contacting a provider or customer on its own.

These aren’t limits on usefulness. They are what lets you use AI in a finance process at all, because every decision still has an accountable person behind it.

Why auto-match rate is the wrong headline number

A high auto-match rate sounds like success, but it says nothing about whether the matches were right, or what happened to the rest. Measure instead:

• unexplained items, by age and value;
• false matches found later;
• exceptions reopened after closure;
• time to close each business day, and whether it closed on complete evidence.

The goal is every payment matched or explained, with evidence, not the highest possible percentage matched automatically.

Questions to ask a vendor

1. If I re-run yesterday with the same rules, do I get exactly the same matches?
2. For any match, can you show the rule and version that made it?
3. Can the AI close an exception, accept a settlement or write off a difference? Who records those decisions?
4. Is every use of a tolerance logged?
5. What happens to an item the rules can’t match: is it owned, dated and evidenced?
6. Is every AI action recorded, with the AI as the actor?

One exception, investigated

Say the rules find a payout that the provider’s API reports as paid, but that is missing from the provider’s settlement file. It becomes an exception with a type, an owner and a due-by time. An AI agent then gathers what a person would: the API response, the webhook history, the settlement file for that day and the next, and the bank statement.

It notices that the payout was confirmed at 23:52 in the provider’s time zone, after the file’s cut-off, and that the next day’s file lists it. It drafts a short explanation with those three records attached and suggests matching against the next day’s file. A person reads it, agrees, and records the decision. If forty payouts show the same pattern, the agent can propose a change to the timing rule, and a person reviews and approves the new rule version before it is used.

At no point did the agent decide anything. It saved the investigation time, which is where most of a reconciliation team’s day goes.

What good AI output looks like

An AI explanation is only useful if a person can check it quickly. Ask for three things every time: the records it relied on, linked; the conclusion in one or two sentences; and the proposed next step, marked clearly as a proposal. An explanation without its evidence should be treated as a guess, however confident it sounds.

Where to start

If you want AI in your reconciliation, start where it is safest and saves the most time:

1. Get the deterministic core right first. Written rules, tolerances and expectations for your main providers and bank accounts, with every break raised as an owned exception.
2. Add AI to investigation. Let it gather the records for each exception and draft an explanation, and measure how often people accept its drafts.
3. Let it propose rule changes. When it spots a repeated pattern, it drafts a new rule version for a person to review, test on past days and approve.
4. Keep decisions with people. Closing, accepting and writing off stay human, with permissions and a record of who decided.

Teams that start the other way round, with a model matching directly, usually end up rebuilding the rules later, when an auditor asks why two records were paired.

How Flominzo Recon does it

In Flominzo Recon, matching never depends on a model. Versioned rules, tolerances and expectations decide every match, so the same evidence always gives the same result and every match shows the rule behind it. The Recon Agent reads, correlates and drafts an explanation for each exception, and can request a deterministic action, such as a status lookup or a file re-fetch, through the same controlled commands a person uses.

It never closes an exception, accepts a settlement or writes off a difference on its own, and every agent action is recorded with the agent as the actor. 100% reconciliation means every payment is matched or explained, with evidence: it ends either matched against independent evidence - the provider’s records, the settlement file and the bank statement - or as an open exception with an owner and a reason. None is silently assumed paid.

Questions

Related