---
type: "post"
id: "taxbit-assist-a-grounded-role-aware-ai-agent-fe2ca8"
slug: "taxbit-assist-a-grounded-role-aware-ai-agent-fe2ca8"
title: "Taxbit Assist: A grounded, role-aware AI agent on Amazon Bedrock AgentCore"
description: "Taxbit says its embedded Assist agent has been in production in the US and EU since May 2026, answering dashboard questions with cited product knowledge and role-aware workflow checks."
retrieval_nugget: "Taxbit says its embedded Assist agent has been in production in the US and EU since May 2026, answering dashboard questions with cited product knowledge and role-aware workflow checks. The case matters to teams building AI support in regulated settings because it shows how retrieval scope, trusted identity, data residency, and live evaluation can shape a deployable system."
published_at: "2026-11-07"
updated_at: "2026-11-07"
record_date: "2026-11-07"
date_kind: "published_at"
topics: ["enterprise-agents","agents"]
entities: ["Taxbit","Taxbit Assist","Amazon Bedrock AgentCore","Strands Agents SDK","Amazon Bedrock Guardrails","Amazon S3 Vectors","Cohere Rerank 3.5","Amazon CloudWatch"]
related_patterns: ["ai-native-orgs-move-to-review-and-orchestration"]
novelty_label: "new_fact"
what_changed: "Taxbit presents Taxbit Assist, a grounded, role-aware AI agent built with Amazon Bedrock AgentCore."
prior_refs: ["https://newruntime.com/posts/building-an-ai-powered-contract-intelligence-platform-with-7ab3d6/","https://newruntime.com/posts/agentic-engineering-intelligence-with-amazon-nova-multimodal-embeddings-fe5804/"]
source_url: "https://aws.amazon.com/blogs/machine-learning/taxbit-assist-a-grounded-role-aware-ai-agent-on-amazon-bedrock-agentcore"
source_urls: ["https://aws.amazon.com/blogs/machine-learning/taxbit-assist-a-grounded-role-aware-ai-agent-on-amazon-bedrock-agentcore"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:taxbit-assist-a-grounded-role-aware-ai-agent-fe2ca8"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/taxbit-assist-a-grounded-role-aware-ai-agent-fe2ca8/","markdown":"https://newruntime.com/posts/taxbit-assist-a-grounded-role-aware-ai-agent-fe2ca8.md","json":"https://newruntime.com/posts/taxbit-assist-a-grounded-role-aware-ai-agent-fe2ca8.json"}
markdown_url: "https://newruntime.com/posts/taxbit-assist-a-grounded-role-aware-ai-agent-fe2ca8.md"
---

# Taxbit Assist: A grounded, role-aware AI agent on Amazon Bedrock AgentCore

## Retrieval answer

Taxbit says its embedded Assist agent has been in production in the US and EU since May 2026, answering dashboard questions with cited product knowledge and role-aware workflow checks. The case matters to teams building AI support in regulated settings because it shows how retrieval scope, trusted identity, data residency, and live evaluation can shape a deployable system.

## Summary
Taxbit says its embedded Assist agent has been in production in the US and EU since May 2026, answering dashboard questions with cited product knowledge and role-aware workflow checks. The case matters to teams building AI support in regulated settings because it shows how retrieval scope, trusted identity, data residency, and live evaluation can shape a deployable system.

## What happened
AWS's announcement describes Taxbit Assist, an assistant embedded in the Taxbit dashboard and built as a Strands Agents SDK application on Amazon Bedrock AgentCore Runtime. Taxbit is a digital-asset tax and accounting compliance platform. The company says customers previously had to search its website, API documentation, and help center—or file support tickets—to resolve product questions. In its internal production analytics, more than 55% of users who ask Assist a question return with another question within 24 hours.

A dashboard request reaches the runtime with the user's verified email, organization ID, and dashboard roles. Amazon Bedrock Guardrails screens the incoming turn and masks personally identifiable information; a blocked message does not invoke the model. The agent searches an Amazon Bedrock Knowledge Base backed by Amazon S3 Vectors, reranks retrieved passages, and asks a Bedrock model to reason over them. It checks whether the authenticated user has access to the relevant feature before recommending a workflow, then streams an answer with source links. AgentCore memory carries conversation context across turns, and OpenTelemetry traces go to Amazon CloudWatch.

The agent is a small Python application with eight deliberately narrow tools: knowledge-base search, two entitlement and role checks, four operations on the caller's own support tickets, and a current-time tool. In this version it cannot read customer tax data; its only write actions concern the caller's own tickets. Taxbit says identity values are bound into tools by the trusted calling service before the model sees the request, so the model can choose whether to check access but cannot choose whose identity to use. Session memory is partitioned by memory, session, and actor ID. For regional deployments, geography-prefixed inference profiles keep EU inference in the EU and US inference in the US; BYOK accounts encrypt conversation history and other client-data stores with the customer's AWS KMS key.

The retrieval corpus combines published materials, documentation generated from product source code, and curated articles or approved corrections. Scheduled Lambda scrapers update content in S3; Taxbit says its sitemap has more than 400 URLs, but only about 70 belong in an answer-generating corpus. It excluded marketing pages, then added a document's H1 title and lead paragraph to each semantic chunk so the passage retains its scope. The excerpt reports that about 28% of chunks would otherwise fall in a range where tails could be truncated. At query time, vector search produces 75 candidates, Cohere Rerank 3.5 reranks them, a two-chunk-per-document cap is applied, and the top five reach the model. Only published sources can be cited to customers.

Taxbit reports continuous evaluation: AgentCore evaluators score a 10% sample of live sessions, custom judges measure faithfulness and retrieval relevance, and a separate classifier checks for tax-advice drift. Experts review sampled, scrubbed interactions; approved corrections enter the curated corpus and trigger incremental ingestion. The company says the production system has operated in the US and EU since May 2026. Average model cost is under $0.20 per conversation, and prompt caching serves about three-quarters of input tokens from cache, cutting model cost by more than half.

## What the archive adds
New Runtime's earlier coverage includes an [Amazon Quick and Bedrock AgentCore contract-intelligence case](https://newruntime.com/posts/building-an-ai-powered-contract-intelligence-platform-with-7ab3d6/) (September 30, 2026) and a [supply-chain intelligence case](https://newruntime.com/posts/from-data-to-decisions-conversational-supply-chain-intelligence-b8bdf2/) (October 22, 2026). This is a new customer application on a familiar platform story, with distinct details on role-bound support answers, corpus curation, and continuous evaluation.

## Trend signal
This case supports [AI-native organizations moving toward review and orchestration](https://newruntime.com/patterns/ai-native-orgs-move-to-review-and-orchestration/): the system routes product questions through constrained tools, evaluates responses, and lets subject-matter experts correct the knowledge base. It also illustrates that shipping the agent entails an operational review loop, not just model access.

## Why now
Taxbit says the assistant has been in production in the US and EU since May 2026. The timing is therefore a report on a deployed system and its observed usage and cost, rather than a launch announcement or a forecast.

## Why it matters
For teams serving customers through AI, this is a reminder that “grounded” is an end-to-end property. Taxbit narrowed its corpus before tuning prompts, added document scope to retrieved chunks, and separated faithfulness from retrieval relevance so teams can distinguish a bad answer from a bad source selection. Its reported 75-candidate retrieval pool also recognizes that useful authoritative material may rank well below the first results before reranking.

The access design is equally reusable: bind tenant identity and roles in trusted application code, keep tools narrow, and let the model operate only within that context. Continuous sampling and human correction turn customer interactions into an improvement loop, while a deployment-specific inference profile addresses residency and cost attribution. These are implementation choices other teams can test; the reported repeat usage and cost are company metrics, not proof that the same design or economics will transfer unchanged.

## Limits and uncertainty
The usage, cost, caching, and retrieval-impact figures are Taxbit's own reports in an AWS-published customer case; they are not independently audited here. The announcement does not establish a controlled comparison against a previous support workflow, nor does it quantify answer accuracy, resolution time, avoided tickets, or the share of questions escalated to humans. The architecture describes the current product-knowledge version and explicitly says it has no access to customer tax data; the performance and safety implications of a future broader assistant are not demonstrated by these results.

## Skill to build now
Take one recurring support question and trace it from user identity to final answer. Bind the user's organization and role in application code, retrieve only sources they can open, and add each document's scope to the retrieved chunks. Then log the retrieved sources separately from answer quality so a reviewer can tell whether a failure came from retrieval or generation.

## Reader takeaway
A production support agent is not made trustworthy by a careful prompt alone. Taxbit's case puts the work in the corpus, identity boundary, permission checks, and correction loop—and reports a sub-$0.20 conversation cost once those pieces are running.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Enterprise Agents](https://newruntime.com/topics/enterprise-agents.md): Explore the enterprise-agents topic hub.
- topic hub: [Agents](https://newruntime.com/topics/agents.md): Explore the agents topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)