Taxbit, a company that handles crypto tax and compliance work for exchanges, payment services and brokers, has built an AI assistant into its customer dashboard. The assistant, called Taxbit Assist, gives answers drawn from a curated set of sources and limits what each user can do. Taxbit says more than 55 percent of users who ask a question come back with another within 24 hours.
Why Taxbit built it
Taxbit was founded in 2018. Its customers use the dashboard to calculate gains and losses, collect tax documents and file information returns. The knowledge needed to use it well was spread across the public website, API documentation and help-center articles. Some of it was never written down. A question like "what does this form status mean?" meant digging or opening a support ticket. The assistant runs on Amazon Bedrock AgentCore, an AWS platform for running AI agents.
Keeping answers trustworthy
Much of the work went into choosing what the assistant reads. The Taxbit website has roughly 400 pages, but only about 70 were kept as sources. More than 300 were marketing blog pages. Left in, they would compete with the help center as the cited source for compliance questions. Taxbit calls removing them one of its highest-value changes, and it needed no work on the AI model itself.
The assistant also reads documentation generated from Taxbit's own product code. It covers dashboard routes, permission rules, form eligibility logic and reporting settings. Because it is regenerated from the code, it stays current. A third source is hand-written articles and approved corrections.
Privacy and permissions
The AI model never decides whose permissions to check. A user's email, organization and roles come from a trusted calling service and are attached to each tool before the model sees them. Input is cleaned so a prompt injection, meaning a trick message meant to fool the AI, cannot pose as that trusted information.
- In this version, the assistant cannot read customer tax data.
- Its only writes are the caller's own support tickets.
- EU deployments keep AI processing in the EU, and US deployments keep it in the US.
- Customers who bring their own encryption key get conversation history encrypted with it.
What it means for other businesses
Taxbit's post offers a practical pattern for adding an AI helper to business software. Pick a narrow set of trusted sources. Keep identity out of the AI's hands. Limit what the helper can touch. Taxbit also says it measures answer quality on live traffic, though the details are not covered in the material available here.
The takeawayBefore adding an AI helper to your tools, limit its sources, its access and its control over who counts as the user.