Squaring the bank isn’t a month-end job: it’s matching every movement to its invoice, and that gets automated

That’s the case that breaks most automatic reconciliation, because nearly all of it pairs one-to-one on an exact amount. Here it looks for the combination: given a deposit, it tests it against the sets of that customer’s open invoices that sum to the amount —including the possibility that it arrives with an early-payment discount or a withholding—; given an invoice, it accepts that several movements cover it and leaves it open as partial until the rest arrives, rather than calling it paid or calling it a discrepancy. When more than one combination fits equally well, it doesn’t pick at random: it escalates with both options, because a match that’s right by luck dirties the ledger and surfaces three months later.

Nothing explicit, but there are signals, and in most cases they’re enough. The movement carries an originator, an amount, a date and a description, however dirty. Against that we cross what that customer has outstanding, how they’ve paid historically —whether they round, whether they lump, whether they pay at 30 or 60 days— and the variants of their name, which is what fails most by hand: the payment comes in under a holding company, a director, or a trading name that doesn’t match the record. What we don’t do is force an assignment to keep the scoreboard clean: if the signal is weak, the movement stays identified as pending with its candidates, which is useful information, rather than falsely reconciled, which is an error someone pays for at close.

No. This is built on what you already use: the statement comes in through your bank’s file or API and the result gets written into your accounting (Sage, Holded, Odoo, SAP and similar). If you already have a treasury tool you like, we connect it rather than replace it; what we automate is what that tool doesn’t do —match the awkward cases against your real ledger and escalate with context instead of leaving a list of “unreconciled”. What we remove is the step where a person downloads the statement and ticks off lines.

It’s designed assuming it will, because in reconciliation the silent error is the expensive one. Three things contain it. It starts in proposal mode: it matches, but a person approves, and autonomy gets released per movement type only once the measured accuracy earns it. Everything it applies leaves a trail —which movement, against which document, by which rule or signal— so reverting is one step, not an investigation. And the confidence threshold is yours: below it, it escalates instead of deciding. The honest comparison isn’t against a perfect process: it’s against squaring by hand at nine at night on the 3rd, which also gets things wrong and leaves no record of why it ticked that line.