AI bookkeeping assistant that reads receipts and prepares GST returns.
The client started as a two-person bookkeeping firm drowning in shoeboxes of receipts from their SME clients. They wanted to productise what they did by hand, without building something accountants wouldn’t trust.
The hard part wasn’t reading receipts. It was making the system honest about uncertainty, so a human could check the 6% that mattered instead of the 100% that didn’t.
Our approach
- Confidence as a first-class featureEvery extracted field carries a confidence score. Low-confidence items go to a review queue; high-confidence items can be approved in bulk.
- Local rules, not generic onesCategory suggestions learn from each company’s history, and GST checks validate registration numbers and the 9% rate before anything is filed.
- Ask the ledgerA chat panel that answers questions about the books using only the company’s own ledger, always showing the receipts behind each answer.
- Mobile captureA camera flow that reads the receipt before the shutter sound ends, so the entry is done before you leave the shop.
AI in the product, and in how we built it
Receipt extraction and the assistant run on a commercial LLM with structured outputs, wrapped in validation rules that reject anything that doesn’t add up. During the build we used AI coding tools for scaffolding, test generation and migration scripts. Every model prompt and evaluation set was written by our engineers and reviewed with the client’s accountants.
- Evaluation set of 2,400 real, anonymised receipts
- Extraction accuracy of 97.8% on totals and 91% on categories
- No customer data used to train third-party models
What we’d do differently
Our first version of the assistant answered too confidently. In the beta it told one user their August fuel spend was “normal” while a duplicated receipt was inflating it. We added a rule that every answer must cite its records, and the assistant now says “I’m not sure” far more often.
“Sampan pushed us to show the AI’s doubt instead of hiding it. That one decision is why accountants are willing to sign off on what it produces.”
Services on this project: Product strategy, SaaS, AI features, Mobile app.