Announcing Amazon Textract Insights: Verifiable generative AI for documents

Organizations across financial services, insurance, healthcare, and the public sector depend on document processing. They turn scanned files and paper forms into structured data their systems can act on. As generative AI has matured, these organizations want more than accurate text extraction. They want to describe the information they need in their own words and get dependable, auditable answers from complex documents.

Today, we are announcing the general availability of Amazon Textract Insights, a new generative AI feature of Amazon Textract available through a new API within Amazon Textract. With Amazon Textract Insights, you describe the fields you want in natural language and submit a document or a mixed, multi-document packet. You receive structured results in which each value carries a confidence score and page-level grounding, with no model training required.

The challenge of extracting data from mixed packets

Many of the high-value document workloads involve diverse, mixed packets rather than clean, single-purpose files. A mortgage file, for example, typically arrives as a single scan holding a W-2, payslips, bank statements, and a page of handwritten notes. Each has its own layout and appears in no fixed order. Claims files and onboarding packages can look very similar: typed forms, dense tables, scanned receipts, and free-text letters interleaved in one file.

Teams usually take one of two approaches, and each has trade-offs. 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 all of it. The second is a general-purpose large language model used by default. It returns fluent answers but does not reliably cite the source page or give a confidence score. Without a traceable source and a calibrated score, it is hard to know which values to accept automatically and which to review. This uncertainty often limits how far teams can automate.

Amazon Textract Insights is built for reliable document processing, bringing this entire workflow into a single asynchronous request on Amazon Textract. Rather than assembling and operating a custom pipeline, your team works through one managed API and moves directly to the data it needs.

How Amazon Textract Insights works

With Amazon Textract Insights, you define what to extract in a blueprint. A blueprint is 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 content of the document. You can begin with an AWS public blueprint for a common document type, such as a W-2, Form 1040, Form 1098, a passport, or a birth certificate. You can also author your own blueprint for the documents specific to your business.

For a mixed packet, you add the blueprints for each document type you expect to the same project. You then submit a single file of up to 3,000 pages from your own Amazon Simple Storage Service (Amazon S3) bucket and enable splitting. Amazon Textract Insights splits the packet into logical sub-documents, matches each one to the blueprint that fits it, extracts the requested fields, and returns a confidence score and grounding for each value.

For example, suppose you receive a mortgage application as a single 26-page PDF that contains a W-2, payslips, bank statements, and a loan officer’s handwritten notes. You create an Amazon Textract Insights project with the AWS public W-2, payslip, and bank statement blueprints, add a custom blueprint for loan officer notes, and enable splitting. You then submit the file from your Amazon S3 bucket with a single StartDocumentInsights call. Amazon Textract Insights splits the file into its individual documents, matches each one to its blueprint, and writes the results to your S3 bucket. Each value includes a confidence score and the page and bounding box it was drawn from.

Confidence and grounding for auditable automation

Amazon Textract Insights returns two kinds of fields: The first are explicit fields read directly from the page. The second are inferred fields derived from the document’s content. Both return with a confidence score and with grounding that identifies the page and the coordinates the value was drawn from.

Because each value is traceable to its source, a reviewer or an auditor can confirm any result against the original document. You can extend automation to documents that would previously have gone to manual review, while keeping people in the process wherever judgment is required.

“We evaluated Amazon Textract Insights and are excited about its potential. Our medical bills and forms are received from a wide variety of sources and often have different layouts and formats. As a result, extraction rules developed for one source do not always transfer effectively to another.

With Amazon Textract Insights, we can describe the fields we need in plain language, and the service identifies and extracts the corresponding data. Each extracted value is returned with a confidence score, allowing our team to focus their review efforts only on the fields where the service has lower confidence. We are also excited to explore the document splitting capabilities, which could help automatically separate multi-document files into individual documents and further streamline our document processing workflow.”

— Vijay Agrawal, Vice President, Information Systems, Symbeo, A CorVel Company

Built on the Amazon Textract engine

We updated Amazon Textract’s reading engine, Amazon Textract Core, improving accuracy, strengthening the handling of dense tables and handwriting, and expanding language coverage to 40 languages for the DetectDocumentText and AnalyzeDocument APIs. All these benefits are available to Amazon Textract customers today with no changes. Your Amazon Textract APIs remain unchanged.

Amazon Textract Core is now generally available in two additional AWS Regions, Asia Pacific (Tokyo) and Europe (Zurich), bringing in-region processing to a total of 18 AWS Regions. Amazon Textract Insights isn’t yet available in these AWS Regions.

Security and data handling

Projects and blueprints

Amazon Textract Insights encrypts projects and blueprints at rest with AWS owned keys. To use your own AWS Key Management Service (AWS KMS) customer managed key, specify it in EncryptionConfiguration when you call CreateProject or CreateBlueprint.

Insight analysis jobs

Amazon Textract Insights reads documents from, and writes results to, the Amazon S3 buckets specified in your StartDocumentInsights request. Intermediate job data is encrypted at rest with AWS owned keys. If you specify a customer managed key in the request’s EncryptionConfiguration, Amazon Textract Insights uses it to encrypt the job’s intermediate data. All intermediate data is transient and is deleted when the job completes. Your output bucket’s default encryption applies to the results written to your bucket when the job completes. A project’s key does not apply to its jobs.

If you disable your key or remove Amazon Textract’s access to it, Amazon Textract Insights can no longer use data encrypted under it.

Cross-Region inference

Amazon Textract Insights automatically selects the optimal AWS Region within your geography to process your inference requests. This maximizes available compute resources and model availability, and is designed to deliver a consistent customer experience. All data is transmitted encrypted across Amazon’s secure network. Amazon Textract Insights securely routes your inference requests to available compute resources within the geographic area where the request originated. For example, inference requests originating in the United States are processed within the United States. Amazon Textract Core APIs are unchanged and continue to process requests in the AWS Region you call. For the list of destination AWS Regions, see the Amazon Textract documentation guide.

Getting started

Getting started with Amazon Textract Insights takes four steps.

Define what you need: Start from an AWS public blueprint or describe your own fields in natural language.

Create a project: Add your blueprints to it.

Submit your documents: Send a single document or a mixed packet from your Amazon S3 bucket with one API call.

Review the results: Accept values that meet your confidence threshold and route the rest for review. Tune your threshold on a representative sample of your own documents.

Amazon Textract Insights supports PDF, PNG, JPEG, and TIFF inputs and processes documents in English, French, German, Italian, Portuguese, and Spanish. It is priced per page rather than per token, with no minimum commitment. For current rates across Amazon Textract, see the Amazon Textract pricing page.

Conclusion

Intelligent document processing has long forced a choice between how far you can automate and how well you can audit the result. Amazon Textract Insights narrows that trade-off so your team can decide where to automate based on evidence rather than guesswork. Learn more about getting started with Amazon Textract Insights.

Visit the Amazon Textract product page.

Follow the Amazon Textract Developer Guide to author a blueprint or start with an AWS blueprint.

Register for Accelerate intelligent document processing with Amazon Textract session at re:Invent.


About the authors

Keith Mascarenhas

Keith Mascarenhas

Keith leads worldwide go-to-market strategy for generative AI at AWS, focused on intelligent document processing and computer vision across AWS AI services and Amazon Bedrock. Since joining AWS, he has worked across product management, solutions architecture and go-to-market, giving him a builder’s view of how AI delivers business value.

Sneha Choudhury

Sneha Choudhury

Sneha is a Principal Product Manager driving product strategy and roadmap for AWS’s Documents & Visual AI initiatives. She shapes how AI-powered products are built and delivered to AWS customers.

Niklas Breuer

Niklas Breuer

Niklas is a Sr Product Manager on the AGI team, where he defines and delivers AI-driven information products for AWS. He works closely with customers, engineering, and science teams to shape how AWS AI services are built and brought to market.

Wrick Talukdar

Wrick Talukdar

Wrick is a Tech Lead focusing on generative AI at Amazon Web Services, driving innovation through multimodal AI, agentic solutions, and foundational models. He is also the author of the bestselling book “Building Agentic AI Systems” and “Generative AI Ethics, Privacy, and Security” for Large Language Models. He is a keynote speaker and often presents his innovations and solutions at leading global forums, including AWS re:Invent, ICCE, Global Consumer Technology conference, and major industry events such as World Technology Summit, CERAWeek and ADIPEC. In his free time, he enjoys writing and birding photography.

Malena Lopez-Sotelo

Malena Lopez-Sotelo

Malena is a Technical Product Marketing Manager, Specialized AI Services and Models. Malena helps customers understand and adopt AI services. She partners across teams to turn complex technology into clear stories and successful launches.