# AWS launches Amazon Textract Insights for verifiable generative AI on documents

published: 2026-11-21T02:05:42.352Z
topic: infrastructure
source: https://aws.amazon.com/blogs/machine-learning/announcing-amazon-textract-insights-verifiable-generative-ai-for-documents/
url: https://aiagentnews.io/stories/aws-launches-amazon-textract-insights-for-verifiable-generative-ai-on-documents

Amazon Textract Insights is now generally available as a new API in Amazon Textract, letting developers describe the fields they need in natural language and get structured results with confidence scores and page-level grounding. It accepts single documents or mixed packets of up to 3,000 pages and splits them into sub-documents matched to blueprints.

Developers define the fields to extract in a blueprint, which is a JSON Schema where each field has a natural language description and an inference type that decides whether a value is read verbatim or inferred from the document. AWS provides public blueprints for common documents such as a W-2, Form 1040, Form 1098, a passport, and a birth certificate, and teams can write their own. A project can hold several blueprints for a mixed packet.

A file of up to 3,000 pages is submitted from an Amazon S3 bucket with splitting enabled, and a single StartDocumentInsights call starts the job. Each value comes back with a confidence score and the page and bounding box it was drawn from, so a reviewer can check it against the original document. Supported inputs are PDF, PNG, JPEG, and TIFF, covering English, French, German, Italian, Portuguese, and Spanish.

AWS also updated the underlying Amazon Textract Core engine, which now covers 40 languages for the DetectDocumentText and AnalyzeDocument APIs, and made it generally available in Asia Pacific (Tokyo) and Europe (Zurich). Amazon Textract Insights is not yet offered in those two Regions. Pricing is per page with no minimum commitment.

## Why it matters if you build agents

Teams processing mixed document packets can replace a custom OCR, splitting, classification and prompting pipeline with one managed asynchronous request, and set a confidence threshold to decide which values are accepted automatically and which go to review.

Because results carry page-level grounding, developers can route low-confidence fields to people while automating the rest. Confidence thresholds should be tuned on a representative sample of their own documents before relying on them.

Insights is not available in every Region, so teams with in-region processing needs should check the supported list first.
