Enterprise
AWS Textract Insights Adds Verifiable GenAI for Mixed Documents
New managed service extracts structured data from multi-document packets using natural language queries, with confidence scores and page-level citations for every field.
Key takeaways
- Amazon Textract Insights extracts structured data from mixed-document packets using natural language field descriptions, eliminating the need for custom OCR pipelines.
- Every extracted value includes a confidence score and page-level citation, enabling automated processing with auditable grounding for regulated industries.
- The service automatically splits multi-document files up to 3,000 pages, matches each section to the appropriate blueprint, and processes them in a single API call.
- AWS updated the underlying Textract Core engine with improved accuracy for tables and handwriting, plus support for 40 languages across existing APIs.
- Pricing is per-page rather than per-token, with support for PDF, PNG, JPEG, and TIFF files in six languages including English, French, German, Italian, Portuguese, and Spanish.
Amazon Web Services has released Amazon Textract Insights, a generative AI capability designed to extract structured data from complex, multi-document files while maintaining the auditability enterprises require for regulated workflows.
The service addresses a persistent challenge in document processing: handling mixed packets where multiple document types arrive in a single file. A mortgage application might contain W-2 forms, payslips, bank statements, and handwritten notes bundled together in no predictable order. Insurance claims and customer onboarding files present similar complexity.
How the service works
Textract Insights operates through blueprints—JSON schemas where each field includes a natural language description of the data to extract. Users can start with AWS-provided blueprints for common documents like W-2s, Form 1040, passports, or birth certificates, or create custom blueprints for proprietary forms.
For mixed packets, users add multiple blueprints to a single project. The service automatically splits files of up to 3,000 pages into logical sub-documents, matches each to the appropriate blueprint, and extracts the requested fields. Every extracted value returns with a confidence score and grounding that identifies the source page and bounding box coordinates.
Fields come in two types: explicit values read directly from the page, and inferred values derived from document content. Both include traceable citations to their source location.
Why it matters
Document processing teams in financial services, healthcare, and insurance have faced a difficult trade-off. Custom pipelines combining OCR, classification, foundation models, and review steps deliver accuracy but require significant engineering overhead. General-purpose large language models offer simplicity but lack the confidence scores and source citations needed for automated decision-making in regulated environments.
Textract Insights collapses this multi-step workflow into a single managed API. The confidence scores and page-level grounding enable organizations to automate acceptance of high-confidence extractions while routing uncertain values to human review—a critical requirement for audit trails in financial and healthcare applications.
Vijay Agrawal, Vice President of Information Systems at Symbeo, a CorVel company, noted that medical bills arrive from diverse sources with varying layouts, making extraction rules difficult to transfer between sources. The natural language field descriptions and confidence scoring allow his team to focus review efforts only where the service reports lower confidence.
Technical updates and availability
AWS updated the underlying Amazon Textract Core engine with improved accuracy for dense tables and handwriting, plus expanded language support to 40 languages. These improvements apply automatically to existing DetectDocumentText and AnalyzeDocument API calls with no code changes required.
Textract Core is now available in 18 AWS Regions, including newly added Asia Pacific (Tokyo) and Europe (Zurich). Textract Insights supports PDF, PNG, JPEG, and TIFF inputs in English, French, German, Italian, Portuguese, and Spanish. Pricing is per-page rather than per-token.
The service uses cross-region inference within geographic boundaries to optimize compute resources while keeping data within its originating geography. All data transmission occurs over Amazon's encrypted network.
These details were first reported by AWS in a machine learning blog post announcing the general availability of Amazon Textract Insights.