Home / Blog / AWS announces Amazon Textract Insights: Verifiable…
Tech News

AWS announces Amazon Textract Insights: Verifiable generative AI

AWS launched Amazon Textract Insights, adding natural language field extraction, packet splitting, and bounding box grounding for files up to 3,000 pages.

By Dillip Chowdary • Nov 20, 2026 • Source: AWS Machine Learning Blog

AWS announces Amazon Textract Insights: Verifiable generative AI
Image: AWS Machine Learning Blog

At a glance

Product
Amazon Textract Insights
Availability
General availability
Maximum file size
Up to 3,000 pages per file
Supported formats
PDF, PNG, JPEG, TIFF
Pricing model
Per-page pricing with no minimums
Core regions
18 AWS Regions for Textract Core

Amazon Web Services announced the general availability of Amazon Textract Insights, introducing a generative artificial intelligence capability accessible through a dedicated application programming interface within its managed document extraction service. As detailed in AWS Machine Learning Blog's report, the feature enables organizations to define target extraction fields using natural language prompts without training custom machine learning models. The managed system accepts single documents or mixed packets containing up to 3,000 pages directly from Amazon Simple Storage Service buckets, returning structured data alongside confidence scores and page-level grounding for every extracted value.

This breakdown examines the architecture of Amazon Textract Insights, its verification mechanisms, security controls, and operational prerequisites. The analysis is targeted at technical leads, compliance officers, and software engineers in financial services, insurance, and healthcare who build automated document pipelines and require auditable, deterministic output from unstructured paperwork.

01Mixed packets and custom pipeline trade-offs

Organizations in regulated sectors routinely process unstructured document bundles rather than uniform, single-page submissions. A standard mortgage submission or insurance onboarding packet arrives as an unsorted scan containing paystubs, W-2 forms, bank statements, and handwritten notes. Because these sub-documents follow disparate layouts and appear in arbitrary sequences, standard optical character recognition systems struggle to isolate target records without extensive pre-processing.

Engineering teams typically choose between building complex custom workflows or deploying general-purpose foundation models. Assembled pipelines force developers to stitch together separate classification, optical character recognition, splitting, prompting, and human review systems. General-purpose large language models produce fluent text but omit calibrated confidence scores and verifiable source coordinates, making it difficult to automate decisions safely.

Extraction ApproachOperational OverheadSource GroundingCalibrated Confidence
Custom PipelineHigh integration and maintenance burden across multiple specialized servicesManual mapping requiredVaries by custom model implementation
General LLMModerate prompting overhead without dedicated ingestion infrastructureUnreliable citationLacks calibrated numerical scores
Textract InsightsSingle asynchronous API call on managed AWS infrastructurePage number and bounding box coordinatesCalibrated numerical score per field

02Blueprints and automated packet splitting

System administrators specify extraction criteria through blueprints defined as JSON Schema specifications. Each field within a blueprint incorporates a natural language description detailing the target data and assigns an inference type. That inference configuration dictates whether the service extracts text verbatim from the scanned page or infers logical answers directly from surrounding document context.

Developers can deploy pre-built AWS public blueprints for standard government and financial forms, including Form 1040, Form 1098, W-2 statements, passports, and birth certificates. Teams can also author bespoke blueprints tailored to proprietary internal documents and combine multiple blueprints within a single project. This modular architecture allows a unified pipeline to parse diverse incoming records against matching structural definitions without retraining underlying models.

Mixed document bundles move through automated segmentation when developers activate the splitting parameter in their extraction requests. A single asynchronous call to the StartDocumentInsights endpoint initiates processing on multi-document files stored in Amazon S3. The service parses the file, divides pages into logical sub-documents, maps each segment to its matching blueprint, and writes structured output directly to designated S3 buckets.

Advertisement

Tech Pulse Daily

Get tomorrow's pulse first

Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.

03Page coordinates and auditable confidence scores

The service extracts explicit values copied directly from the page alongside inferred fields derived through contextual reasoning. Both classifications return with a calibrated confidence score and coordinate-level grounding data detailing the exact source page and bounding box. This spatial attribution allows automated ingestion systems to route low-scoring results directly to verification queues while auto-approving values that exceed validation thresholds.

Downstream verification teams and audit specialists can cross-reference any generated field against the original high-resolution scan without searching through unindexed pages. Symbeo Vice President of Information Systems Vijay Agrawal noted that their teams receive medical bills and forms from diverse sources with varying layouts where static extraction rules fail. Agrawal explained that their team focuses review efforts exclusively on low-confidence fields while using automated splitting to separate multi-document packages.

04Core reading engine and regional footprint

Amazon also deployed foundational updates to Amazon Textract Core, the underlying extraction engine powering the wider document service ecosystem. The upgrade increases optical character recognition precision, strengthens table extraction on dense financial ledgers, and improves recognition across complex handwriting. Additionally, the update expands language coverage across DetectDocumentText and AnalyzeDocument endpoints to forty languages without requiring modifications to client applications.

The foundational reading engine expanded its physical infrastructure to two additional geographic destinations, launching general availability in Asia Pacific (Tokyo) and Europe (Zurich). This expansion brings the total global footprint for in-region Amazon Textract Core processing to eighteen AWS Regions worldwide. Existing workloads calling core APIs retain their current integration endpoints while inheriting accuracy improvements automatically.

Note: Amazon Textract Insights is not yet available in the Asia Pacific (Tokyo) and Europe (Zurich) AWS Regions, even though Amazon Textract Core now supports in-region processing across all eighteen regions.

05Cryptographic controls and cross-region inference

Security protections for Amazon Textract Insights operate at the project, blueprint, and job execution levels. Blueprints and project configurations are encrypted at rest using AWS owned keys or optional AWS Key Management Service customer managed keys specified during creation. Individual analysis jobs read input scans from Amazon S3 and write outputs using transient intermediate storage that is automatically purged upon completion.

  1. Define what you need Start from an AWS public blueprint or describe your own fields in natural language.
  2. Create a project Add your blueprints to it.
  3. Submit your documents Send a single document or a mixed packet from your Amazon S3 bucket with one API call.
  4. Review the results Accept values that meet your confidence threshold and route the rest for review.

Inference requests securely route across Amazon's internal network to compute clusters located within the geographic area of origin. Requests originating in the United States execute entirely on domestic infrastructure to fulfill regional data boundary requirements. Supported input formats include PDF, PNG, JPEG, and TIFF files across English, French, German, Italian, Portuguese, and Spanish, with billing structured per page rather than per token.

Frequently asked questions

What input file types and languages does Amazon Textract Insights support?

The service accepts PDF, PNG, JPEG, and TIFF files. It processes documents written in English, French, German, Italian, Portuguese, and Spanish.

How does Amazon Textract Insights handle mixed multi-document packets?

Users submit files of up to 3,000 pages from Amazon S3 with document splitting enabled. The system separates the packet into logical sub-documents and matches each to its designated blueprint.

How are extracted fields verified for audit compliance?

Every extracted field includes a calibrated confidence score and page-level grounding. The grounding metadata specifies the exact page number and bounding box coordinates from the original document.

How is pricing structured for Amazon Textract Insights?

Pricing is billed per processed page rather than per token, with no minimum usage commitments required.

Sources

Dillip Chowdary

Author

Dillip Chowdary

Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.

Related on Tech Bytes

Advertisement

5-min tech signal

Weekday briefing for engineers who skip the noise.

No spam · Unsubscribe anytime

Advertisement

✈️ CareerPilot

Your AI job-search copilot

Match your resume against live Ashby, Greenhouse & Lever openings — fit scores, job-specific resume optimization and email alerts.

Find matching jobs →

Free Tools

Browse all tools →