AI-assisted accounting automation

The problem

Bookkeeping generates a long tail of repetitive work: finding which bank transactions still need an invoice, digging the invoice out of an email account, matching it to the right transaction, categorising it, and handling EC reverse-charge VAT cases.

The solution

A substantial FreeAgent integration covering invoices, supplier bills, banking, bank transaction explanations, expenses, contacts, projects, tasks and timeslips — and, on top of it, an automation layer that finds transactions needing invoices, searches for and ranks likely matches, retrieves invoices from Gmail and Outlook, attaches them to the transaction, updates the explanation and assigns the category. The whole thing is exposed through an MCP-style interface so an AI agent can drive the workflow safely rather than being handed raw API access.

Technical implementation

An E-PATH tool, built to remove the repetitive parts of running the company’s own books — and a working demonstration of joining a third-party API, two email providers and an AI agent into one dependable workflow.

Technical implementation

The foundation is a thorough FreeAgent integration. Not a thin wrapper over a couple of endpoints: it covers invoices, supplier bills, banking, bank transaction explanations, expenses, contacts, projects, tasks and timeslips, which is what makes real automation possible on top of it.

The automation layer does the work a person would otherwise do by hand:

• find bank transactions that still need a matching invoice
• search for candidate invoices and rank them by how likely each match is
• retrieve invoices from Gmail and Microsoft Outlook where they arrived as email attachments
• attach the invoice to the accounting transaction
• update the explanation and assign the correct accounting category
• handle EC and reverse-charge VAT cases rather than leaving them to be fixed later

Everything is exposed through an MCP-style interface. That matters: an AI agent gets a set of well-defined, reviewable operations against the accounting system instead of open API credentials, so the model works inside a boundary somebody designed on purpose.

The ranking step is the interesting part. Matching an invoice to a transaction is rarely exact — amounts differ by currency conversion, dates drift, and supplier names on a bank feed rarely match the name on the invoice. Producing a ranked set of candidates with a confidence signal turned out to be far more useful than trying to force a single deterministic answer.