How to build an AI assistant for an accounting, tax or finance business with RagLeap Core

Short answer: run RagLeap Core, load your service descriptions, published fees and document checklists, and use roles such as finance, collections_agent and tax_preparation_intake. It collects intake details and answers routine questions. It is not financial, tax or legal advice. The steps are in the general setup guide.

Which roles apply?

tax_preparation_intake, insurance_agent and compliance_officer are among the built-in sensitive-domain roles in the code (sensitivity.py at v0.11.0): they are forced to semi or off autonomy and get a reasoning mode and a grounding check, and the README says sensitive roles always need approval. finance, collections_agent and secretary are other built-in roles.

What would you upload?

Service descriptions, the fee schedules you publish, document checklists, deadlines and process descriptions, and your FAQ. Do not upload clients' financial records unless you have decided that your setup, your provider and your legal basis allow it.

How do people reach it?

Through WhatsApp, Telegram, Discord or a voice line (Twilio). A website chat widget is planned, not built. Language detection runs across every channel; short messages in closely related languages can be misdetected.

What should you check first?

Not advice. No accuracy claim is made. Do not use it to give financial, tax or legal advice.
Personal financial data. Check your provider terms and your obligations before any client data goes through a provider.
Autonomy is off by default; semi asks before acting; full is opt-in. Read real conversations before going live.
Your documents and each conversation go to the AI provider you configure. A local Ollama model keeps those steps on your machine, but local models on a CPU are slower and less accurate than Gemini.
Single-tenant: one deployment, one business, one document set. Named users and roles arrived in v0.12.0: four roles, personal keys and an audit log of who made each state-changing request (updated 2026-10-29: this page said “No named users or roles yet”). No single sign-on or multi-factor sign-in, and no independent security review yet.
No results for this field are claimed, and no compliance certification. Test with your own documents and questions.

Sources

The v0.11.0 README, defaults.py and the AI employees documentation.