Automating our financial systems
Agents read our receipts, match them to the bank and book them only when every check passes. What needs a person goes to one review queue. It runs our own finance every day.
11 of 563 in August 2026, with the old rules. 26 of the 43 lines that needed matching in September, the first full month with the agents.
26 of the 32 bank lines matched that month. A person matched the other 6.
The review queue in August 2026, and on 16 October 2026.
All of them, using the official Spanish chart of accounts.
Bank reconciliation was the first real agent work, and the hardest. Receipts come in every format, from restaurants and SaaS vendors in several countries and currencies, and the bank line rarely matches the receipt exactly: card lag, foreign-currency legs, fees.
When we measured it, the rule-based matcher auto-matched 2% of bank lines and left the rest in a review queue. That queue was the work we wanted agents to take.
Step by step.
- 01
Capture. Receipts arrive by email, upload, phone or chat, and an AI model reads them: supplier, amounts, taxes and currency.
- 02
Classification. Each expense gets its accounting category and the right VAT treatment, based on what the invoice actually shows.
- 03
Matching. Each receipt is paired with its bank movement, with foreign amounts converted at the official rate of their date. Every confirmed match teaches the system, so the next one from that supplier goes through on its own.
- 04
Judgement calls, like whether two names are the same business, go to Jev, a small model that gives a scored answer. Sums, taxes and dates are always checked in code.
- 05
Booking. An expense is booked on its own only after a series of checks passes. If anything doesn’t add up, it waits for a person as a one-click suggestion, with the reason.
Built with Receipt reading (OCR) · Bank feed · Payments · Official exchange rates · Spanish chart of accounts · Automatic checks before booking
Exchange rates need the document’s own date. Converting at the day the work ran looked right and was off by several percent.
We measured Jev for merchant resolution before trusting it. At the best threshold, 28 correct and 8 wrong. Not good enough to act alone, so that use stays off and only feeds suggestions.
Learning from confirmations beats smarter guessing. One confirmed alias unlocks every pending line from the same counterparty.
What we decided.
Adopted. It runs our own finance every day. Automatic booking is gated by checks in code, and every autonomous step sits behind a flag we switch on only after measuring it.
The same agents and checks fit any company whose finance still runs on manual matching and spreadsheets. For your finance team, that can mean:
- Receipts captured from email, phone or chat, read and classified without typing.
- Bank reconciliation, with one review queue for the doubtful cases.
- Invoices drafted from project milestones and time entries.
- Foreign-currency documents recorded at the exchange rate of their own date.
- Monthly close and tax reports prepared from data that is already checked.
- Nothing is booked on a guess. A failed check leaves a suggestion, not an entry.
- Arithmetic, tax and period checks run in code; models only answer scored questions.
- Closed periods are locked. A late receipt for a filed quarter stays a suggestion.
- The ledger is Spanish: Spanish chart of accounts and Spanish tax forms. Receipts can come from any country; the books follow Spanish rules.
Switch on the automatic confirm flows one by one as their measured precision holds.
Can AI reconcile receipts with bank transactions?
Yes, with rules around it. Exact cases are matched automatically, known suppliers are recognised from past confirmations, and an AI model only helps with judgement calls. Anything uncertain goes to a person.
How do you handle receipts in foreign currencies?
Each receipt keeps its original currency and is converted at the official exchange rate of its own date.
What happens when the AI isn't sure?
It doesn't book anything. The item becomes a suggestion in a review queue with the reason, for example a missing tax ID or a closed quarter, and a person confirms in one click.