Taxbit embeds a cited, role-aware AI assistant with AWS technology

Taxbit has built Taxbit Assist, a cited and permission-aware AI assistant embedded in its dashboard, using Amazon Bedrock AgentCore. AWS says the assistant is designed to help compliance teams find product answers while respecting user roles, organization access and regional data requirements.
- Taxbit Assist is embedded in the Taxbit dashboard.
- The assistant provides cited answers and checks roles and organization access.
- It uses Amazon Bedrock AgentCore, Knowledge Bases and S3 Vectors.
- The agent has no read access to customer tax data in this version.
- More than 55 percent of users return with another question within 24 hours, according to Taxbit.
An assistant designed for regulated workflows
Taxbit is a digital-asset tax and accounting compliance platform founded in 2018. It helps enterprises including exchanges, payment services and brokers calculate gains and losses, collect tax documentation and file information returns at scale.
According to AWS, the knowledge required to use Taxbit's dashboard was previously spread across the public website, API documentation and help-center articles. Some product knowledge was not written down. Users asking questions such as “what does this form status mean?” had to search for an answer or file a support ticket.
Taxbit Assist is intended to answer questions inside the product dashboard and provide links to the sources used. More than 55 percent of users who ask Assist a question return to ask another within 24 hours, according to Taxbit's internal production analytics.
The design reflects four constraints identified by Taxbit: answers must be grounded and cited, the assistant must understand roles and organizations, it must avoid giving tax advice, and it must support regional data residency. AWS says Taxbit Assist is developed and operated in line with ISO/IEC 42001. Taxbit is also certified to ISO/IEC 27001 and ISO/IEC 27701 and holds a SOC 2 Type II report.
How Taxbit Assist handles access and data
Taxbit Assist is a Strands Agents SDK application running on the Amazon Bedrock AgentCore runtime. A streaming chat API sends each question to the runtime with the user's verified email, organization ID and dashboard roles. Amazon Bedrock Guardrails screens the incoming message and masks personally identifiable information in place. If a message is blocked, the model is not invoked.
The agent retrieves passages from an Amazon Bedrock Knowledge Bases knowledge base, reranks them server-side and uses Amazon Bedrock models to reason over the results. Before recommending a workflow, it checks whether the user can access the relevant feature and has the required role. Answers stream back to the dashboard with source links, while AgentCore memory carries the conversation across turns.
The agent has eight deliberately narrow tools: knowledge-base search, organization entitlement and dashboard-role checks, four tools for creating, listing, updating and closing the caller's own support tickets, and a current-time tool. In this version, it has no read access to customer tax data, and its only writes are to the caller's own support tickets.
AWS says the model never handles identity directly. The trusted calling service supplies the user's identity and binds it into each tool before the model sees it. The model can decide whether to check permissions, but cannot choose whose permissions to check.
Retrieval is central to the system
Taxbit chose Amazon Bedrock AgentCore so a small team would not have to build and operate the agent loop, conversation-state store and session isolation itself. AgentCore provides a managed runtime, built-in memory and per-session isolation, according to AWS.
The knowledge base uses Amazon Bedrock's managed Retrieval Augmented Generation capability, backed by Amazon S3 Vectors. Taxbit uses Cohere Embed English v3 for embeddings, producing 1,024-dimension float32 vectors compared by cosine distance.
Content comes from three sources. Scheduled AWS Lambda functions crawl the public website, API documentation and help-center articles. Other Lambda functions read product source code and generate Markdown documentation about dashboard routes, permission gates, form eligibility logic and reporting configuration. Taxbit also adds hand-written articles and approved corrections.
Content is stored in a single Amazon S3 bucket under separate prefixes, with Amazon EventBridge scheduling staggered nightly and weekly crawls. AWS says Taxbit reduced retrieval competition by excluding most marketing pages from the answer-generating corpus. Its sitemap has roughly over 400 URLs, but only about over 70 belong in that corpus, while the /blogs/ path alone accounts for more than 300 marketing pages.
What the architecture means for users
For compliance teams, the main change is access to product guidance within the dashboard rather than through separate searches or support requests. The cited-response design also gives users a source they can open and read, while the permission checks are intended to prevent recommendations for workflows or pages they cannot access.
The system keeps inference geography-aware through cross-Region inference profiles. AWS says EU deployments keep inference within the EU and US deployments stay in the US. Taxbit also supports bring your own key accounts, where conversation history and other stores containing client data are encrypted with the customer's own AWS Key Management Service key.
Every request emits OpenTelemetry traces to Amazon CloudWatch, which feeds Taxbit's evaluation tooling. This gives the company a way to measure answer quality on live production traffic as the assistant operates for regulated customers.


