Amazon AI Reads 3,000-Page Files And Shows Where Each Answer Came From

Amazon AI: illustration for Amazon Gives Exact Source For Answers Pulled From Scanned Files

Amazon is tackling one of the biggest reasons banks, insurers and hospitals still hesitate to let AI read their paperwork: it now shows exactly where on the page every answer came from, with a score for how sure it is, across files of up to 3,000 pages.

Key Takeaways

  • Amazon Web Services (AWS) made Amazon Textract Insights generally available as a new feature of its Textract document reader
  • The service lets users submit a document or mixed packet and receive extracted values with confidence scores and grounding
  • It can split a mixed bundle, such as a mortgage file, into separate documents and pull the requested details from each in one step
  • AWS said Textract Insights is priced per page, not per token, with no minimum commitment

Amazon announced the release on its AWS machine learning blog. The company said organizations in financial services, insurance, healthcare, and the public sector use document processing to turn scanned files and paper forms into structured data their systems can act on.

Textract Insights Handles Mixed Document Packets

AWS said many high-value document workloads involve mixed packets rather than clean, single-purpose files. Its examples include mortgage files, claims files, and onboarding packages that can combine forms, tables, receipts, letters, and handwritten notes in one file.

Amazon said users describe what to extract in a blueprint, a simple template where each field is explained in plain language. For mixed packets, users can add multiple blueprints to the same project and enable splitting.

How AWS Describes the Document Processing Problem

AWS said teams usually either build a custom pipeline with optical character recognition (OCR) software that turns images of text into text, document splitting, classification, a foundation model prompt, and a review step, or use a general-purpose large language model that may not reliably cite the source page or give a confidence score.

The company said that without a traceable source and a calibrated score, it is harder to know which values to accept automatically and which to review.

Also Read: Amazon Lets Labs Fold Private Proteins Without Leaving Their Own Cloud

Textract Insights Uses One Managed API

AWS said Textract Insights brings the workflow into a single asynchronous request on Amazon Textract. Rather than assembling and operating a custom pipeline, teams use one managed application programming interface (API), the connection software uses to talk to the service, and move directly to the data they need.

The company said users can submit a single file of up to 3,000 pages from Amazon’s S3 cloud storage. Textract Insights can split the packet into logical sub-documents, match each one to a blueprint, extract the requested fields, and write the results to Amazon S3.

Built on Amazon Textract

AWS said Textract Insights is a new generative AI feature of Amazon Textract. It also said it updated Amazon Textract Core, improving accuracy, strengthening handling of dense tables and handwriting, and expanding language coverage to 40 languages for the DetectDocumentText and AnalyzeDocument APIs.

AWS said Textract Insights supports PDF, PNG, JPEG, and TIFF inputs and processes documents in English, French, German, Italian, Portuguese, and Spanish. It said the service is priced per page rather than per token, with no minimum commitment.

Grounding and Confidence Scores

AWS said Textract Insights returns both explicit fields read directly from the page and inferred fields derived from document content. It said both types come with a confidence score and grounding that identifies the page and coordinates the value was drawn from.

The company said that because each value is traceable to its source, a reviewer or auditor can confirm the result against the original document.

What It DoesWhy It Matters
Reads files up to 3,000 pagesWhole mortgage or claims bundles go in as one file
Splits mixed bundlesPay slips, bank statements and forms are separated automatically
Shows page and position for each answerReviewers and auditors can check every value against the original
Gives a confidence scoreSure answers pass straight through, doubtful ones go to a person
Charges per page, no minimumCosts are easier to predict than per-token AI pricing
Source: Amazon Web Services announcement, Nov. 20, 2026.

Why This Matters

AI models are good at reading documents but bad at showing their work. In finance, insurance and healthcare, an answer you cannot trace back to the page is close to useless, because someone has to sign off on every decision and an auditor may check it years later. A general chatbot that sounds confident but cannot point to its source is exactly what these industries cannot use.

By attaching a page location and a confidence score to every value, Amazon lets companies automatically accept the answers the system is sure of and send only the doubtful ones to a human. That is how AI moves from pilots to processing millions of real forms, and it puts Amazon in a tighter race with Google and Microsoft, which also sell AI document tools to big companies. Pricing per page rather than per token, the unit most AI services charge by, also makes costs easier for these firms to predict.

What To Watch Next

The real test is accuracy on messy, real-world files, such as faded scans and handwritten notes, which Amazon says its updated text reader handles better. Watch for banks and insurers to report how much of their paperwork they can now process without a human checking every field.

Language support is another limit for now. Textract Insights handles English, French, German, Italian, Portuguese and Spanish, so expansion to more languages will decide how far it reaches outside Western markets.

Read Next: AI Startup Arms Software Bots With Live Data From Billions Of Pages

Similar Posts