Taxbit Assist is in production in the US and EU, AWS details its controls and evaluation

AWS’s November 6, 2026 case study describes Taxbit Assist, an AI assistant embedded in the Taxbit dashboard and built on Amazon Bedrock AgentCore. Taxbit says the service has been in production in the US and EU since May 2026, with EU traffic served using in-geography inference. According to Taxbit’s internal production analytics, more than 55% of users who ask a question return to ask another within 24 hours.

The assistant is built with the Strands Agents SDK, and its knowledge base draws on published content, code-derived product documentation, and curated material. For each query, the system searches 75 candidates, reranks them to 12, limits retrieval to two chunks per source document, and passes five chunks to the model. Only published sources can produce customer-facing citation links; the other materials can ground an answer but do not generate links. AWS documentation describes AgentCore Runtime as supporting multiple frameworks, including Strands Agents; the Taxbit example shows one enterprise-assistant use of that framework support.

The agent’s access and actions are bounded: in this version, it cannot read customer tax data, and its only write actions are on the caller’s own support tickets. It has four ticket tools—for creating, listing, updating, and closing those tickets. Taxbit also supports a bring-your-own-key model, encrypting conversation history and other stores holding client data with the customer’s AWS KMS key. Its guardrails block certain secrets and anonymize specified numeric identifiers.

For quality monitoring, AgentCore online evaluations sample 10% of live sessions and score them with six built-in evaluators. A foundation-model classifier also checks every response for tax-advice drift. The AWS post reports near-zero false positives for that classifier, compared with roughly 25% for a topic filter. Taxbit also reports average model cost below 20 cents per conversation, with about three-quarters of input tokens served from cache.