Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableAWS Machine Learning Blog→ original

Amazon Textract Insights parses bundles of up to 3000 pages and shows the source of every field

Amazon has launched Textract Insights, a service that extracts the fields you need from scans based on plain-language descriptions, with no model training. It can split a 26-page mortgage application containing a W-2, statements and handwritten notes into separate documents in a single request (files can run up to 3000 pages). Every value comes with a confidence score, a page number and a bounding box on that page, so reliable fields flow into your system automatically while doubtful ones go to a human.

AI-processed from AWS Machine Learning Blog; edited by Hamidun News
Amazon Textract Insights parses bundles of up to 3000 pages and shows the source of every field
Source: AWS Machine Learning Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceAWS Machine Learning Blog ↗
Editor-in-Chief
Supported formatsPDF, PNG, JPEG and TIFF; Insights understands 6 languages, including English
Key metric #2Pricing is per page, not per token, with no minimum commitments
Key metric #3Textract Core grew to 40 languages and 18 regions; Insights is not yet available in the new ones

Banks, insurers and government agencies have been scanning paper for years, but distrust keeps them from handing the parsing over to a neural network entirely: it is unclear where a number came from and whether it can go into a database without review. Amazon has made Textract Insights generally available, a new feature of Amazon Textract that addresses exactly this question. Picture a mortgage application: a single 26-page PDF containing a W-2, pay stubs, bank statements and a loan officer's handwritten notes, all mixed together.

You describe the fields you need in plain words, send the file (up to 3000 pages) in one request, and the service splits it into documents on its own. For every value it finds, it returns a confidence score, plus the page and the bounding-box coordinates the value was taken from. The score lets you decide what to pass through automatically and what to send to a human, while the page and box let you quickly check the result against the original.

How it works, using a mortgage application as an example

Until now, teams had to choose between two options. The first was to build their own pipeline of text recognition, document splitting, a classifier, a language-model query and a review stage, and then maintain all of it. The second was to send the file to a general-purpose LLM. It answers fluently, but it does not always cite the source page and does not provide a calibrated confidence score.

In Textract Insights, you describe the fields you need in a so-called blueprint, which is a JSON Schema where each field has a natural-language text description. Each field is given a type: the value is either read verbatim from the page or inferred from the contents of the document. Ready-made public templates exist for common paperwork: W-2, Form 1040, Form 1098, passport and birth certificate. For your own documents, you write the template yourself.

Next, you add templates for all the expected document types to a project, turn on splitting, and send a file from S3 with a single StartDocumentInsights call. The service divides the bundle into logical documents, matches each one to its template and writes the result back to your bucket. The job runs asynchronously, and there is no model to train.

What this gives you in practice

The main value is controlled automation. An auditor can check any field against the original, and the team sets its own confidence threshold: values above it move on without human involvement, and values below it go to manual review. Symbeo (part of CorVel), which processes medical bills from many sources in varying formats, points to exactly this: manual review can be focused only on low-confidence fields. One caveat: Symbeo has so far only evaluated the service, and it describes bundle splitting as something it wants to explore, not as a solution it already runs.

Amazon also updated the underlying Textract Core engine. Accuracy is improved, dense tables and handwritten text are handled better, and the number of languages for the DetectDocumentText and AnalyzeDocument APIs has grown to 40. These APIs have not changed, so existing customers get the improvements without rewriting code. Textract Core has also arrived in the Tokyo and Zurich regions, bringing the total to 18.

Limitations and what to check before adopting it

  • Insights itself supports only six languages: English, French, German, Italian, Portuguese and Spanish.
  • Insights is not yet available in Tokyo and Zurich; those regions apply to Core only.
  • Inference requests may be processed in a different region within your geography. For example, a request from the US stays in the US. If you have strict data-residency requirements, check this against the documentation. Core, by contrast, runs in the region you call.
  • Intermediate job data is encrypted with AWS keys and deleted when the job finishes. You can specify your own KMS key through EncryptionConfiguration; the project key does not apply to jobs.
  • Pricing is per page, not per token, with no minimum commitments. The announcement does not give specific rates.

The takeaway for developers and product managers is simple: before moving a workflow to automatic processing, take a representative sample of your own documents and tune the confidence threshold. Amazon itself recommends this step. Calibrating on real data will show what share of fields actually passes without a human, and only then can you work out the project's economics. If your team currently maintains a homegrown pipeline of OCR, a classifier and an LLM, this is a reason to compare its total cost of ownership with a single managed API.

Story

Key Questions Answered

How it works, using a mortgage application as an example?

Until now, teams had to choose between two options. The first was to build their own pipeline of text recognition, document splitting, a classifier, a language-model query and a review stage, and then maintain all of it. The second was to send the file to a general-purpose LLM. It answers fluently, but it does not always cite the source page and does not provide a calibrated confidence score.

What this gives you in practice?

The main value is controlled automation. An auditor can check any field against the original, and the team sets its own confidence threshold: values above it move on without human involvement, and values below it go to manual review. Symbeo (part of CorVel), which processes medical bills from many sources in varying formats, points to exactly this: manual review can be focused only on low-confidence fields.

What are the key details on Limitations and what to check before adopting it?

Insights itself supports only six languages: English, French, German, Italian, Portuguese and Spanish. Insights is not yet available in Tokyo and Zurich; those regions apply to Core only. Inference requests may be processed in a different region within your geography. For example, a request from the US stays in the US. If you have strict data-residency requirements, check this against the documentation. Core, by contrast, runs in the region you call.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…