AWS launches document AI that links extracted values to their source
Textract Insights handles files up to 3,000 pages through one managed request. Customers still choose which results to accept and which need review.
Loading page…
Textract Insights handles files up to 3,000 pages through one managed request. Customers still choose which results to accept and which need review.
Listen to this story
Amazon Textract Insights became generally available on Nov. 20, giving teams a managed way to extract schema-defined fields from documents and mixed packets without training a model. Results pair each value, including inferred answers, with confidence and page-level source coordinates, so teams can set cutoffs and route uncertain cases for review. The service is built around S3 workflows and charges by page; files can contain up to 3,000 pages, and customers need to account for regional availability and configure job encryption separately from project encryption.
Blueprints define fields with plain-language descriptions and specify whether Textract should read values directly or infer them; AWS also offers templates for forms such as W-2s and passports.
For mixed packets, teams can enable splitting and submit the expected document-type blueprints together in one project.
Textract Insights accepts PDF, PNG, JPEG and TIFF files in six languages; it was unavailable in the Tokyo and Zurich AWS Regions at launch.
AWS is pitching document AI that people can check against the original paperwork, not just answers they must take on trust. On November 20, it made Amazon Textract Insights generally available. The new API extracts requested information from documents and mixed packets, returning confidence scores and source locations alongside each value.
The service requires no model training. AWS says it combines work that teams would otherwise assemble from text recognition, document splitting, classification, a generative model and review. Customers submit one asynchronous request—meaning the job runs separately rather than returning an immediate answer—and receive structured results.
Setup starts with a blueprint: a JSON Schema, or structured definition of the fields to return. Each field includes a plain-language description of what to find and a setting for whether to read or infer the answer. Teams can write their own blueprints or use AWS templates for documents such as W-2s, passports and birth certificates.
For a mixed packet, customers add the expected document types’ blueprints to one project and enable splitting. A single file can contain up to 3,000 pages and must come from an Amazon S3 storage bucket. The service separates the packet into logical documents, matches them to blueprints, extracts the requested fields and writes the results back to S3.
Customers decide where automation stops. AWS recommends testing a confidence threshold on a representative sample of their own documents, accepting values above that threshold and routing the rest for review.
Symbeo, a CorVel company that evaluated the service, described the problem with fixed extraction rules. Its vice president of information systems, Vijay Agrawal, said medical bills and forms arrive in varied layouts, so rules built for one source do not always transfer to another. He said the company was excited to explore document splitting.
Textract Insights accepts PDF, PNG, JPEG and TIFF files. It processes documents in English, French, German, Italian, Portuguese and Spanish. At launch, the service is not yet available in the Tokyo and Zurich AWS Regions.
Processing can cross AWS Region boundaries within a geography. AWS automatically routes inference requests to available computing resources within the geographic area where they originated; requests originating in the United States stay within the United States. The company says data travels encrypted over Amazon’s network.
Projects and blueprints are encrypted at rest with AWS-owned keys by default, with customer-managed keys available. Intermediate job data is also encrypted and is deleted when the job finishes. Results use the output bucket’s default encryption. A project’s encryption key does not automatically apply to its jobs; customers must specify a customer-managed key in the job request if they want it used there.
Billing follows the paperwork: Textract Insights charges per page rather than per token, with no minimum commitment. AWS directs customers to its pricing page for current rates; the launch announcement does not list specific prices.
Loading discussion...
Join the conversation
Explain where useful interpretation becomes too much guesswork.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.