Amazon Web Services announced the general availability of Amazon Textract Insights on November 20, 2026, adding a generative-AI document-processing API to Textract. The service splits, classifies and extracts structured data from mixed document packets in one asynchronous request. It launched in US East (N. Virginia) and US West (Oregon). (AWS announcement)
The useful detail is what comes back with each extracted value: a confidence score and grounding that identifies its source page and bounding box. That gives developers the pieces to build a review screen where a person can check a result against the relevant area of the original scan. AWS says the workflow requires no model training. (AWS launch post)
For regulated document automation, I’d put that traceability ahead of the generative-AI label. Structured output is useful; structured output that a reviewer can inspect against its source is much easier to incorporate into a controlled workflow. The important qualification is that traceability and correctness are separate properties.
Blueprints Define What to Extract
Textract Insights uses blueprints, which are JSON Schemas containing natural-language descriptions of the fields an application needs. Each field also has an inference type that determines whether the service reads a value directly from the page or derives it from the document’s content. (AWS launch post)
That makes a blueprint both an output structure and a set of extraction instructions. A business can describe the information it wants without training a separate model for every document layout.
The distinction between field types matters:
| Field type | What the service does | What a reviewer needs to check |
|---|---|---|
| Explicit | Reads a value directly from the page | Whether the value was read correctly and associated with the right field |
| Inferred | Derives a value from document content | Whether the supporting content justifies the derived answer |
AWS provides public blueprints for common documents, including W-2s, invoices and bank statements. Teams can also create custom blueprints for their own documents. For mixed packets, multiple blueprints can be added to the same project. (AWS GA notice)
A practical schema-writing concern is ambiguity. For example, a field described only as “date” leaves more room for confusion than one specifying the invoice issue date, excluding payment and delivery dates. That is an implementation recommendation based on how the documented mechanism works, not a published AWS accuracy result.
“No model training required” removes one setup task. Teams still need to define the fields, test their descriptions and decide what counts as an acceptable answer.
One Asynchronous Request Handles a Mixed Packet
Textract Insights accepts a file of up to 3,000 pages from the customer’s Amazon S3 bucket. With splitting enabled, it separates the packet into logical sub-documents, matches each to a suitable blueprint and extracts the requested fields. Results are written to the customer’s S3 bucket. (AWS launch post)
AWS illustrates the workflow with a 26-page mortgage application containing a W-2, payslips, bank statements and a loan officer’s handwritten notes. The project combines public blueprints with a custom blueprint for the notes, then submits the packet through a StartDocumentInsights call.
The managed request covers several jobs that a document application might otherwise coordinate separately:
- Reading text and document content.
- Finding boundaries between documents.
- Classifying each resulting document.
- Extracting fields into the requested structure.
- Returning confidence and source-location information.
The asynchronous design is worth noting. This is a job-submission workflow, not a promise that a large packet returns immediately in the initial API response. AWS supports access through its SDKs and the AWS CLI, while its developer guide covers running and monitoring jobs. (AWS GA notice, Developer Guide)
The operational benefit is fewer processing stages for an application team to connect and maintain. The application still needs its own business rules, exception handling and decisions about which outputs can proceed automatically.
Confidence and Grounding Serve Different Purposes
AWS says both explicit and inferred fields return with confidence scores and grounding that identifies the page and coordinates supporting the value. Those two outputs address different parts of the review process. (AWS launch post)
Confidence supports routing. An application can use it to decide which values to send for review.
Grounding supports inspection. A reviewer can open the source page and examine the relevant region, instead of searching through an entire packet.
Together, they provide a useful foundation for auditable extraction. A review interface could show the requested field, returned value, confidence score and highlighted source area alongside the scan. The output makes that design possible; AWS’s announcement is not evidence that every customer application implements it.

Inferred fields deserve particular attention. A highlighted passage can show the evidence behind a derived answer, but the reviewer must still assess the reasoning connecting that evidence to the result. An accurate citation can accompany a mistaken interpretation.
AWS’s announcement provides no independent accuracy benchmark. Confidence thresholds therefore need validation on representative documents before they control automatic acceptance. A score should be checked against observed errors, including differences between document types, field types and scan quality.
I’d also make the review threshold sensitive to the consequence of an error. A noncritical description and a bank-account identifier should not automatically share the same acceptance policy simply because both have confidence scores.
For an audit-friendly implementation, retain the original document, returned extraction, blueprint used and subsequent review decision. That recommendation follows from the workflow: source coordinates help explain an answer only while the corresponding source and processing context remain available.
Inputs, Availability and Per-Page Pricing
AWS lists PDF, PNG, JPEG and TIFF as supported inputs and says Textract Insights supports six languages. Its launch availability is limited to the two named US Regions. (AWS launch post, AWS GA notice)
Frequently Asked Questions
3 questions
1Does Textract Insights require model training?
No. AWS says customers define fields through natural-language descriptions in JSON Schema blueprints, using public templates or custom blueprints.