Taxbit Unveils Taxbit Assist, a Grounded AI Agent Enhancing Compliance Workflows on Amazon Bedrock

Taxbit, a digital-asset tax and accounting compliance platform, launched Taxbit Assist, an AI assistant embedded directly in its dashboard to help compliance teams find authoritative answers quickly. Launched in production environments in both the US and EU since May 2026, it responds to user queries by retrieving grounded, cited content from a curated knowledge base.

The assistant operates on Amazon Bedrock AgentCore, leveraging managed agent runtime, memory, and session isolation features to streamline infrastructure and securely manage conversation state. Incoming user questions, associated with verified identity and roles, pass through Amazon Bedrock Guardrails that mask personally identifiable information and enforce safety policies before model invocation.

At its core, Taxbit Assist retrieves relevant passages from a knowledge base built on Amazon S3 Vectors and populated through automated ingestion of public website content, API documentation, help articles, product source code, and curated editorial content. The knowledge base uses Cohere Embed English v3 embeddings to index textual content, which is segmented semantically and prepended with context summaries to maintain precise topical scope during retrieval.

Role and organization awareness are central design constraints. Before generating a response, the system verifies that the user has proper authorization to access certain workflows or information, ensuring that answers correspond precisely to the user's dashboard permissions. This prevents incorrect guidance that might arise if the assistant suggested inaccessible administrative functions.

Strict boundaries prevent the assistant from providing tax advice; it only conveys product documentation and compliance workflow information. Data residency requirements are addressed with geography-aware inference that keeps processing within the US or EU and supports customer-managed encryption keys. Importantly, Taxbit Assist never accesses or reads customer tax data, limiting risk.

Quality assurance is continuous and multifaceted. Answer quality is monitored live via OpenTelemetry traces and Amazon CloudWatch, with automated checks for harmfulness, coherence, and relevance. Human experts review sampled interactions, submitting corrections that quickly update the knowledge base. A foundation model classifier detects any unintentional drift into tax advice territory, maintaining regulatory compliance.

Despite its capabilities, Taxbit Assist's responses are confined to documented product knowledge and do not extend into user-specific data queries or financial advice. The model cost is modest, averaging under 20 cents per conversation due to prompt caching efficiencies. Early customer feedback emphasizes the assistant’s role in accelerating integration workflows and reducing support requests.