Six months building an AI bookkeeping SaaS for UK accountants — what we learned about category accuracy, HMRC compliance and trust.
Why accountants?
UK accountants spend a huge chunk of their week on a single task: categorising transactions. We saw an opportunity to take that off their plate with AI — without sacrificing accuracy or HMRC compliance.
Three hard-won lessons
1. Domain context beats model size A smaller model with rich UK-specific prompts beat a larger general-purpose model. We baked HMRC category guidance directly into the system prompt and grounded every decision in client-specific rules.
2. Confidence scores are non-negotiable Accountants don't want a black box. Every categorisation in Reconova surfaces a confidence percentage and a human-readable reason. Low-confidence rows route to a review queue.
3. Excel still rules We initially built a beautiful in-app review experience — and then watched users export everything to Excel. Now we treat Excel export as a first-class output, not an afterthought.
What's next
Reconova is now live with paying accountancy practices. The same patterns — domain-rich prompts, confidence-first UX, export-friendly data — apply to almost every AI software project we build.
Tags:#AI#SaaS#Case Study#Reconova
