Verified, then logged.
Sun 22 Nov 2026 · 00:21 KST / 15:21 UTC한국어 | ENSearch /

Breaking02 / IT

Taxbit details tax AI design that keeps identity from models

The assistant for a regulated sector keeps permission scope outside the model, and Taxbit did not name the model that runs inference.

A tax specialist, seen from behind in an office, faces a monitor showing a blurred transaction dashboard and chat assistant panel, with a blank ID badge beside the keyboard. AI-generated image

Taxbit, a digital-asset tax platform, published the architecture of its AI assistant, Taxbit Assist, on the AWS blog on 7 Nov 2026. Dashboard users receive cited answers matched to their permissions, the company said.

Taxbit named "the model never handles identity" as the basis of its security design. The calling service binds each user's email and role information to the tools in advance, so the model cannot alter them. This version cannot read customer tax data. Its only write access is to the user's own support tickets.

The company ranked contextual chunking as its most effective retrieval change. Chunking is the splitting of documents into smaller passages for search. Taxbit attaches the document title and first paragraph to each passage, so the scope of the content is indexed too. Each question draws 75 candidates, which are reranked before only 5 go to the model.

Compatibility. Enabling a custom transformation, a user-defined step that processes documents during ingestion, replaces an existing data source and requires a full re-ingestion, according to Taxbit. Amazon Bedrock also requires the transformation's intermediate storage to sit in a bucket separate from the data source.

Runtime. Assist is deployed as an ARM64 container image on AgentCore runtime, according to Taxbit's post. AgentCore is Amazon Bedrock's platform for building and running AI agents. Inference profiles carry a regional prefix, so inference for EU deployments stays within the EU.

Knowledge about using the product was scattered across the website, API documentation and help centre, the company said. Its internal analytics show that more than 55% of users who ask a question ask another within 24 hours.

Taxbit did not name the foundation model used for inference in its description of the setup.