# Amazon Textract Insights Adds Verifiable Generative AI for Document Extraction

> AWS launches a managed API that extracts fields described in natural language and returns confidence scores and page-level grounding for each value.

- Publication: Thoai Dispatch
- Canonical URL: https://vokythoai.dev/article/amazon-textract-insights-verifiable-generative-ai-documents
- Section: News
- Author: AWS Machine Learning Blog
- Published: 2026-11-20
- Tags: aws, amazon-textract, document-processing, generative-ai, ocr, api, confidence-scoring, s3

## What happened

AWS announced the general availability of [Amazon Textract Insights](https://aws.amazon.com/blogs/machine-learning/announcing-amazon-textract-insights-verifiable-generative-ai-for-documents/), a generative AI feature of Amazon Textract delivered through a new API. According to the AWS Machine Learning Blog, users describe the fields they want in natural language, submit a document or a mixed multi-document packet, and receive structured results in which each value carries a confidence score and page-level grounding. No model training is required.

The announcement targets document-heavy sectors such as financial services, insurance, healthcare, and the public sector, where scanned files and paper forms must be converted into structured data. AWS frames the feature as a response to organizations that want to specify information in their own words and receive dependable, auditable answers from complex documents.

## Why it matters

AWS describes two common approaches to these workloads and the trade-offs of each. The first is a custom pipeline that combines optical character recognition, document splitting, classification, a foundation model prompt, and a review step. The team must integrate, scale, and operate every component. The second is a general-purpose large language model used by default. AWS says such models return fluent answers but do not reliably cite the source page or provide a confidence score, which makes it hard to decide which values to accept automatically and which to review.

Amazon Textract Insights is positioned to collapse that workflow into a single asynchronous request on Amazon Textract. Rather than assembling and operating a custom pipeline, teams work through one managed API. The practical effect, per AWS, is that automation can extend to documents that would previously have gone to manual review, while people remain in the process wherever judgment is required.

## Technical details

Extraction is defined through a blueprint, which AWS describes as a JSON Schema in which each field carries a natural language description of what to find. Each field also has an inference type that determines whether the value is read verbatim from the page or inferred from the document's content. Users can start from AWS public blueprints for common document types, including W-2, Form 1040, Form 1098, a passport, or a birth certificate, or author custom blueprints for business-specific documents.

For mixed packets, users add the blueprints for each expected document type to the same project, submit a single file of up to 3,000 pages from an Amazon Simple Storage Service (Amazon S3) bucket, and enable splitting. The service splits the packet into logical sub-documents, matches each to the appropriate blueprint, extracts the requested fields, and writes results to S3. Each value includes a confidence score and the page and bounding box it was drawn from.

AWS illustrates this with a 26-page mortgage application PDF containing a W-2, payslips, bank statements, and a loan officer's handwritten notes. A project is created with the public W-2, payslip, and bank statement blueprints plus a custom blueprint for loan officer notes, splitting is enabled, and the file is submitted with a single `StartDocumentInsights` call.

The service returns two kinds of fields: explicit fields read directly from the page and inferred fields derived from the document's content. Both come with a confidence score and grounding that identifies the page and coordinates the value was drawn from, so a reviewer or auditor can confirm any result against the original document.

AWS also reported updates to Amazon Textract Core, the reading engine behind Amazon Textract. The update improves accuracy, strengthens handling of dense tables and handwriting, and expands language coverage to 40 languages for the `DetectDocumentText` and `AnalyzeDocument` APIs. Existing Amazon Textract APIs remain unchanged, and AWS says these benefits are available to customers today with no changes. Amazon Textract Core is now generally available in two additional AWS Regions, Asia Pacific (Tokyo) and Europe (Zurich).

Symbeo, a CorVel company, provided a statement in the announcement. Vijay Agrawal, Vice President, Information Systems, said medical bills and forms arrive from a wide variety of sources with different layouts, so extraction rules developed for one source do not always transfer to another. He said the team can describe needed fields in plain language and that confidence scores let reviewers focus on lower-confidence fields, and noted interest in exploring document splitting.

## What to watch

Teams evaluating the feature should note the operational constraints stated by AWS: submissions are asynchronous, files are read from and written to S3, and a single file can hold up to 3,000 pages. The blueprint model, with per-field natural language descriptions and an explicit choice between verbatim and inferred values, is the main design surface. Confidence scores and bounding-box grounding are the mechanisms AWS offers for deciding which values to automate and which to route to human review. Organizations with existing Amazon Textract integrations can adopt the Core improvements without API changes, and those operating in Tokyo or Zurich can now use Textract Core in-region.

## Sources

- [AWS Machine Learning Blog](https://aws.amazon.com/blogs/machine-learning/announcing-amazon-textract-insights-verifiable-generative-ai-for-documents/) (2026-11-20)
