Amazon Textract can now split a 3,000-page packet and score every value
Amazon Textract Insights takes plain-language field descriptions and returns each value with a confidence score and the page and bounding box it came from, with no model training.

Amazon has announced the general availability of Amazon Textract Insights, a generative AI feature that pulls fields out of documents and returns each value with a confidence score and the page and bounding box it came from. It needs no model training.
One request for a whole mixed packet
Users describe what they want in a blueprint, which is a JSON Schema where each field has a natural-language description. Each field also carries an inference type, which decides whether the value is read verbatim from the page or inferred from the document’s content. Amazon supplies ready-made public blueprints for frequently processed paperwork, including US tax forms (the W-2, Form 1040 and Form 1098), passports and birth certificates, and customers can write their own.
For packets, several blueprints go into one project. A single file of up to 3,000 pages is submitted from the customer’s own Amazon S3 bucket with splitting enabled. The service divides the file into logical sub-documents, matches each to a blueprint, extracts the fields and writes results back to S3. The call is StartDocumentInsights, and the job is asynchronous.
Amazon’s worked example is a 26-page PDF mortgage application that bundles a W-2, payslips, bank statements and handwritten notes from a loan officer. The customer combines three public blueprints with one custom blueprint for the notes and submits the file once.
Output comes in two kinds: explicit fields read directly from the page, and inferred fields derived from content. Both carry a score and grounding. Amazon says this lets a reviewer or auditor check any value against the original, and lets teams accept values above a chosen confidence threshold while routing the rest to people. It advises tuning that threshold on a representative sample of one’s own documents.
Who can use it, and on what terms
Inputs can be PDF, PNG, JPEG or TIFF, and documents can be in English, French, German, Italian, Portuguese or Spanish. Pricing is per page rather than per token, with no minimum commitment. The announcement gives no rates and points to the Amazon Textract pricing page, so the cost per page cannot be stated from the source.
At rest, projects and blueprints are protected by keys that AWS owns by default, or by a customer managed AWS KMS key set in EncryptionConfiguration when calling CreateProject or CreateBlueprint. For jobs, intermediate data is encrypted at rest and, according to Amazon, deleted when the job completes. A key specified on a job protects that intermediate data, while the output bucket’s own default encryption applies to the results. A project’s key does not carry over to its jobs.
One detail matters for data-residency planning. Insights uses cross-Region inference: it picks a Region within the user’s geography to process requests, and Amazon says requests originating in the United States stay in the United States. The classic Textract APIs behave differently and keep processing in the Region the caller names. The announcement does not list the destination Regions, referring readers to the documentation, and says Insights is not yet available in the two Regions that just gained the updated reading engine, Tokyo and Zurich.
That engine, Amazon Textract Core, was also updated. Amazon says it improves accuracy, dense tables and handwriting, and now covers 40 languages for DetectDocumentText and AnalyzeDocument, with no API changes. Core now runs in 18 AWS Regions.
“we can describe the fields we need in plain language”
That is how Vijay Agrawal of Symbeo, a CorVel company, put it in a statement published by Amazon. He said the medical bills and forms his team receives vary widely in layout, and that the scores let reviewers concentrate on low-confidence fields. The comment comes from a customer that has evaluated the service, not from an independent test.
A bet on auditable automation priced by the page
The strategy angle is visible in what Amazon chose to sell. Rather than another model, it is selling the review process: scores and source coordinates that decide what a human sees. The post frames the two existing options as a self-built pipeline of OCR, splitting, classification, prompting and review, or a general-purpose language model that gives fluent answers without reliable citations. Insights is positioned between them as a managed service.
The billing unit fits that pitch. Amazon’s recent launches show a mix of metering styles. Amazon Connect Resolutions, which handles refunds and disputes with agentic steps, was priced at $0.40 a case, according to our earlier report, so the unit there is a finished piece of work. Textract Insights uses a page as the unit, which is predictable for a mortgage or claims team that already counts pages, whereas token pricing varies with how a prompt is written.
The cross-Region design also echoes a pattern Amazon is applying elsewhere. SageMaker recently gained the ability to split one GPU seven ways and route inference across Regions, as we reported. Taken together, the launches suggest Amazon is willing to let capacity, not the customer’s chosen Region, decide where inference runs, within a geography. For regulated buyers in finance, insurance and healthcare, which the post names as its audience, that trade-off is something to check against their own rules.
What the source does not establish is how reliable the scores are. Amazon publishes no accuracy figures, no benchmark and no comparison with a general-purpose model on the same documents. The only outside voice is a single customer who said it was excited about the potential and was still exploring the document splitting. Whether a value accepted above a threshold is right often enough to remove manual review is something customers will need to measure on their own files. Amazon is also promoting a re:Invent session on the service, which will be the next place to look for detail.
Sources
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.



