Digital-asset compliance company Taxbit has embedded an AI assistant in its enterprise dashboard to answer product questions with citations. Taxbit Assist uses a curated knowledge base rather than customer tax records, and every suggested workflow is checked against the caller’s organization entitlements and dashboard role.
The system runs on Amazon Bedrock AgentCore. Incoming messages pass through a guardrail that masks personal information before the model is invoked. Retrieval combines public documentation, help articles, code-derived descriptions of product behavior and manually reviewed corrections. Source links accompany answers so users can verify them.
Identity is bound to the assistant’s tools by the trusted application rather than supplied as a model-controlled parameter. That prevents a prompt from changing whose permissions are checked. The agent has no read access to customer tax data; its only write actions can create, update or close the caller’s own support tickets.
Taxbit says more than 55% of users who ask the assistant a question return within 24 hours. That is an engagement measure, not proof of answer accuracy, but the architecture shows how a regulated product can narrow an assistant’s authority and make its claims traceable.