
Taxbit Assist combines cited answers, permissions and live evaluation
Taxbit Assist is an AI assistant embedded in the Taxbit dashboard. The company describes it as cited and permission-aware, with Amazon Bedrock AgentCore providing its foundation. More specifically, it is a Strands Agents SDK application running on the AgentCore runtime. The resulting system is not merely a model-facing chat interface: its design separates input controls, retrieval, memory, identity, tools and production evaluation.
Screening happens before the model is invoked
An incoming message is first screened with Amazon Bedrock Guardrails. Personally identifiable information is masked in place, while a blocked message does not invoke the model at all. The sequence matters because the safety decision is made before answer generation. A message rejected at this boundary does not proceed into the model stage.
Identity is also kept outside the model's control. A trusted calling service supplies the user's email address, organization ID and roles. Those values are bound into tools before the model sees them. The model therefore neither handles nor selects identity. Permission-sensitive tool behavior relies on context supplied by a trusted service rather than on an identity assertion produced during the conversation.
Retrieval supplies the basis for an answer
For an allowed request, the agent retrieves passages from an Amazon Bedrock Knowledge Bases knowledge base backed by Amazon S3 Vectors. It reranks the retrieved passages and then streams an answer with source links. AgentCore memory carries the conversation across turns, allowing later questions to continue from the existing exchange rather than being treated only as isolated prompts.
The knowledge base ingests published material, content derived from code and curated content. It uses Cohere Embed English v3 to produce 1,024-dimensional float32 vectors, which are compared with cosine distance. These details describe Taxbit's stated implementation. They do not establish that the same choices or results apply to unrelated systems.
Preserving document scope in each chunk
Taxbit used semantic chunking together with a post-chunking Lambda transformation. The transformation prepends every chunk with the source document's title and lead paragraph. Consequently, a chunk entering the vector index retains an indication of the document's broader scope instead of carrying only an isolated passage. This prepares each retrieved unit with contextual material drawn from its own document.
Access remains deliberately bounded
In the described version, the agent has no read access to customer tax data. Its only write operations concern the caller's own support tickets. This is an important boundary on the current capability: the architecture described here does not support a claim that Taxbit Assist presently queries customer reporting data.
Regional handling is addressed through geography-prefixed inference profiles. EU inference remains within the EU, and US inference remains within the US. For bring-your-own-key accounts, conversation history and other stores containing client data are encrypted with the customer's AWS KMS key. Regional routing, customer-controlled encryption and trusted identity binding operate as distinct controls within the overall design.
Live sessions are sampled for evaluation
AgentCore online evaluations sample 10 percent of live sessions. They score helpfulness, harmfulness, coherence, instruction following, tool-selection accuracy and tool-parameter accuracy. The set therefore examines both the generated response and how the agent chooses and configures its available tools.
Taxbit says Taxbit Assist has been in production in the US and EU since May 2026. According to its internal production analytics, more than 55 percent of users who ask the assistant a question return to ask another within 24 hours. Taxbit also reports an average model cost below 20 cents per conversation and says that about three-quarters of input tokens are served from cache. These are organization-reported figures for this deployment and should not be generalized into benchmarks for other agents or environments.
Taken together, the description presents a series of explicit control points. The message is screened before model use; identity and roles arrive from a trusted caller; relevant passages are retrieved and reranked; answers are streamed with source links; memory supports multiple turns; and a sample of live sessions is evaluated across response and tool-use criteria. The access limitations are part of that picture, not a footnote: cited output and permission awareness depend as much on what the agent cannot access as on what it can produce.