The situation
A small accounting practice runs on a few experienced people and a lot of repetitive, low-margin work. Client enquiries pile up, end-of-financial-year turns into a workload spike, and hours disappear into researching the same questions and drafting the same kinds of replies. None of it is hard; all of it is time the firm would rather spend on judgement, not lookup.
What we built
A multi-agent co-pilot that sits in front of the firm's inbox. An inbound email arrives through a Postmark webhook and moves through a pipeline: a classifier reads intent, a retrieval step pulls relevant ATO guidance and the firm's own procedures, a drafter writes a reply, and a compliance step checks it before anything is shown to a human. The accountant reviews and sends — they are never out of the loop.
Around that core sit document tools: bank-statement and receipt extraction with ABN verification, built to take the dreariest data-entry work off the desk.
How it works
Retrieval is a RAG system over ATO rulings and firm SOPs, embedded into pgvector on Postgres. Different models are used for different jobs — a cheaper model to classify, a stronger one to draft. Personally identifying information is redacted before anything leaves the perimeter, and every draft carries a compliance check. The review dashboard is a Next.js app; the agents run as serverless functions.
Why it matters
The point of AI in a practice like this isn't to replace the accountant — it's to hand back the hours lost to triage and lookup, while keeping a human on every client-facing decision. The system is now in pilot, built so the firm can trust it precisely because it never acts alone.