15 FAQs, no RAG: Claude Haiku 4.5 bot for tax office

The project was intended as a small, working proof of concept for selling generative-AI adoption support to professional-services firms, a field the developer says is hard to win trust in without real examples. The tax-accounting domain was picked because almost every office has a front-line inquiry desk, and the constraint of never giving a wrong answer is easy to grasp. The system prompt steers the model to rephrase only what is in the FAQs; anything outside them, such as individual tax calculations or specific numerical simulations, is treated as a case for a human, with the bot directing the user to a contact form. The developer deliberately kept the scope narrow to two choices — answer from the FAQ or hand off — and reports no significant stumbling blocks, attributing that to the simple architecture and limited scope rather than to any inherent ease. The stated plan is to decouple FAQs from code into a database, add an admin interface for office staff, and move toward embedding-based retrieval as volume grows, with the trigger being whether all FAQs can still be passed in every time without practical trouble.