Reconciliation is three jobs that usually land on the same person in the same week: deciding what each transaction is, finding what it corresponds to, and noticing what looks wrong. All three are rule-heavy, high-volume, and only occasionally require judgment.
That shape is what makes them automatable — and it's also why the automation has to be honest about its limits. AI here cleans and organises the inputs. It does not invent the numbers. Classification is applied from your own chart of accounts, matching is deterministic, and anything uncertain is raised rather than resolved quietly.
The challenges
- Coding depends on who's coding
- The same supplier lands in two different accounts depending on the month and the person. Nothing looks broken; margins drift anyway.
- One payout, dozens of orders
- Marketplace and processor settlements arrive net of fees, batched and delayed. Unpicking them by hand is the largest recurring time cost most finance teams carry.
- Anomalies are found late or not at all
- A duplicate invoice, a fee that moved half a percent, a missing payout — each is invisible in aggregate and expensive over a year.
- Rules live in someone's head
- The logic that makes reconciliation work is institutional memory, and it leaves when the person does.
- Volume outpaces attention
- Past a certain transaction count, everything gets reviewed less carefully, and the errors that get through are the ones nobody was looking for.
How it works
- Account code classification
- Every line coded against your chart of accounts, learning your conventions rather than a generic taxonomy — and flagging low-confidence cases instead of guessing.
- Intelligent matching
- Transactions paired across bank feeds, payment processors, sales channels, and the ledger, including many-to-one settlements net of fees.
- Line-item anomaly detection
- Duplicates, fee drift, missing payouts, quantities and prices outside their normal range, surfaced with the reason they were flagged.
- Confidence thresholds you set
- You decide what auto-posts, what needs review, and what stops the process. The defaults are conservative.
- Rules that become explicit
- Every decision the system makes is a rule you can read, change, and hand to a new hire.
The measure of a reconciliation system isn't how much it matches. It's whether you trust the small pile it couldn't.
What changes
- Consistent coding
- The same transaction is treated the same way in March and in November, regardless of who is looking at it.
- Reconciliation stops being an event
- It runs continuously, so the month-end is a review rather than a reconstruction.
- Exceptions surface early
- A fee change is caught in the week it happens, not in the quarter it's noticed.
- The logic outlives the person
- Your reconciliation rules exist in a system rather than in someone's habits.
Get started
Run a month through it
Give us one month of transactions and we'll show you the match rate, the coding, and — more usefully — what it flagged and why.