Taxbit Assist: a cited, permission-aware agent at under 20 cents a chat

Taxbit, a digital-asset tax and accounting platform, has published how it built Taxbit Assist, an AI assistant embedded in its customer dashboard. The write-up appeared on the AWS Machine Learning Blog on 6 November 2026, authored by AWS staff, and is vendor-published: AWS is describing a customer running on its Amazon Bedrock AgentCore platform. According to the post, Assist has been in production in the US and EU since May 2026 and its model cost averages under 20 cents per conversation.

What the assistant has to do

The post lists the constraints: every answer must be cited to a page the user can open; it must know the user's role and organisation and not send a support user to an admin-only workflow; it must not give tax advice; and EU traffic must stay in the EU. Taxbit says it is developed in line with ISO/IEC 42001 and that it does not train or fine-tune any model.

The agent is a Strands Agents SDK application on AgentCore runtime with a system prompt and eight tools: knowledge-base search, two access checks (organisation entitlement and dashboard role), four tools to create, list, update and close the caller's own support tickets, and a clock. In this version it has no read access to customer tax data.

Retrieval work that Taxbit says mattered most

Taxbit says scoping the corpus beat any prompt change. Its sitemap has roughly 400 URLs, of which about 70 belong in an answer corpus; more than 300 marketing pages under /blogs/ were competing with help-center articles as the cited source, and excluding them needed no model work.

The second fix was contextual chunking. A hook prepends each chunk with its document's H1 title and lead paragraph, usually the statement of where the content applies, capped at about 128 tokens. Taxbit says a production probe found about 28% of chunks sat in the 385 to 512 token band and would have had their tails silently truncated by the embedding model's 512-token limit, so chunk bodies are capped at 384.

At query time it fetches 75 candidates, reranks them, keeps the top 12, allows at most 2 chunks per document and passes 5 to the model. Citations are enforced in code: only the three published sources can be cited, because only those URLs open for a customer.

Controls

• Identity is bound in code: email, organisation ID and roles come from the calling service and are fixed into each tool before the model sees them. Taxbit says an early test build let the model pass identity as an argument and it moved this before launch.
• Amazon Bedrock Guardrails screens each user turn; passwords, card verification codes and cloud keys are blocked, and numeric identifiers such as Social Security and bank numbers are anonymised in place.
• A model-based classifier checks every response for tax-advice drift. Taxbit says it reaches near-zero false positives, against roughly 25% for a topic filter.
• Online evaluations sample 10% of live sessions with six built-in judges, plus custom faithfulness and relevance judges, and sampled answers go to expert reviewers whose corrections reach the knowledge base within minutes, per the post.

Cost and what the post leaves out

Taxbit says about three-quarters of input tokens come from cache, cutting model cost by more than half, and that cost allocation tags show each workload in AWS Cost Explorer. The only usage figure is that more than 55% of users who ask a question return with another within 24 hours, from Taxbit's internal analytics.

The post does not report answer accuracy, how many conversations that $0.20 average covers, which model is used, how often the tax-advice classifier fires, or how many tickets the assistant deflected. Treat the cost figure as a vendor-reported average, not a benchmark.

Taxbit's own advice is to scope the corpus before tuning prompts, measure guardrail strength against your own traffic (it set PIN and bank-routing detectors to detect-only so form numbers like 1099 pass), and keep the system prompt byte-stable so caching works across users.