AI in the product

Five places a language model does work inside the engine. In every one of them the model proposes or answers, a person decides, and the ledger never moves on the model's word.

Answering questions

Ask how to do something, or ask about the account in front of you: the open screen is part of the question. The model picks which lists to read and how to group them; our own code reads the rows and does the sums, so every figure links to the records behind it. It reads with your permissions and never changes anything.

Reads as you, and every figure is cited

Triage of payments going out

Every direct debit and direct credit is classified by rules first: is this an established series, has the payee's bank account just changed, how large is it against its history. The model then writes a plain-English explanation of each finding for the approver, and the approver decides. If the model is unavailable, the rules decide without it.

The approver reads and decides

Bank statement ingestion

Standard formats, ISO 20022 camt.053 and SWIFT MT940, are parsed outright. For one bank's own CSV or spreadsheet export, the model reads a sample of rows and drafts a format profile: which column is the date, the amount, the reference, the sign. A person reviews the profile once, and every later statement from that bank imports on it.

Drafted once, reviewed once, then reused

Matching money to accounts

A statement line that no matching rule recognises goes to the matching assistant, which proposes the account it belongs to with a confidence. Confirming is what moves the money, and it also teaches the matcher, so the next line from the same payer is proposed the same way.

Proposed by the model, confirmed by a person

Designing an account type

You describe a new financial instrument in plain language, and the designer reads the configuration model, drafts the account type, prototypes the capture screens in a sandbox and forecasts it to maturity. When the numbers are right it goes live at runtime, as configuration.

Drafted by the model, proven by the forecaster

The rules that hold for all five

A person decides

No posting, payment, match or configuration reaches the ledger on the model's output alone. The confirm, approve or deploy step is a person's, and it is on the event.

Rules first, model second

Where a rule can answer, it does. The model explains, drafts or proposes only where a rule cannot, and a rules-only fallback takes over when the model is unavailable.

Switched on per tenant

Each use is a capability your organisation turns on or leaves off. When it's off, every proposal comes from a rule or from a person.

The minimum leaves the building

The model sees only what the task needs, whether that's an instruction's amount, payee and rule findings, a sample of statement rows, or a description of an instrument. Today that model is Anthropic's Claude, behind one client.

Your own agent, too.

The model above is ours. Your organisation may well have its own agent, and the MCP door, arriving in the next release, lets it read the ledger over the Model Context Protocol with a person's permissions and a record of every call. Anthropic's GL Reconciler wants a read-only subledger to reconcile your general ledger against, and the door is one. Our founder tells the story in the journal: let your own agent read the ledger, and context-sensitive help, on steroids.

The tedious half of the back office, reading statements, explaining exceptions and matching money, gets faster, while the accountable half stays with your people and on the audit trail.