Our product · FinTech & AI

PayFlow

Invoice follow-up that handles the reply — and is built so it cannot email a wrong figure.

Real product screens from a local run with sample customers and invoices — not real company data.

Type: Our own product (MVP)
Our role: Product, AI, engineering
Integrations: Xero, Gmail, Microsoft 365
Status: MVP — demos on request

The problem

What needed solving

Any accounting tool can send a payment reminder. The work starts when the customer replies: “we'll pay on Friday”, “we were only invoiced for three of the five shifts”, “this was paid last week”. Someone has to read it, decide what it means, update the records and plan the next step.

Handing that to AI raises a real risk for a finance team: a model that writes emails could quote the wrong amount or date to a customer.

Our approach

What we built

We designed PayFlow so the model interprets but never asserts a financial fact. Every amount, date and invoice reference in an outgoing message is bound from the ledger, and a pre-send validator re-reads each message and blocks it if any figure can't be traced back.

The model's output is only a classification. A deterministic, fully tested policy engine decides what happens next from a closed set of actions; anything uncertain lands in a human attention queue with the evidence highlighted. New organisations start in simulation mode and send nothing until a person switches sending on.