AI Governance & Assurance · live demo
The matching cascade that leaves the model only the leftovers.
Bank reconciliation eats 20 to 50 hours of every close, and most of it is still VLOOKUP on amount, which breaks on split payments, bundled remittances, netted bank fees and duplicate amounts. This demo draws the trust boundary the way I believe every finance agent should: a deterministic four-rule cascade clears the pile on documentary keys, the model may only propose matches for the residual by quoting the exact remittance text (a code check verifies every quote against the source), and a named human accepts or rejects each proposal. Nothing the model touches is booked without that acceptance. Synthetic data · 3,000 bank lines · runs in your browser
The waterfall · four rules, in order
documentary key first, weakest key lastThe referee queue · residual only
Field structure follows ISO 20022 camt.053 statement entries and Xero's CSV bank statement import · the "model" here is a deterministic stub, so the run reproduces exactly · amounts in integer cents