AI for Financial Reconciliation

The problem

Bank reconciliation is trivial one movement at a time and enormous in aggregate. An operator looks at a bank movement, recognises what it is, and assigns it to an accounting entry — then does it again, several hundred times, every month. Most of those decisions are mechanical, and the person making them has been making the same ones for years.

What I built

A model that learns from the ledger already sitting there: years of reconciled movements are a labelled dataset that costs nothing to collect. It proposes an entry for each new movement with a confidence beside it — high confidence is accepted at a glance, low confidence goes to the operator instead of being guessed at. Every correction feeds back in, so the model follows the business rather than freezing on the day it was trained.

The decision I would defend hardest is where it runs: inside the reconciliation screen the operators already used, not in a separate tool with a better model that nobody would open.

Where it stands

In production at Tecnos Group, on real books. Batches that took hours of manual matching became minutes of review, and the operator kept the last word on every entry — which is both the correct design for a ledger and the reason they were willing to use it.