AWS announces general availability of Amazon Textract Insights, a generative AI API for extracting structured, confidence-scored data from documents
Edited by Elliot Marsh, Editor · written with AI from the linked sources (how)
Amazon Textract Insights is now generally available through a new asynchronous API (e.g., StartDocumentInsights). Users define fields in natural-language JSON Schema blueprints, using AWS public blueprints (W-2, Form 1040, Form 1098, passport, birth certificate) or custom ones. They submit a single file of up to 3,000 pages from S3, with optional splitting of mixed packets into sub-documents. Each extracted value returns a confidence score plus page and bounding-box grounding, and no model training is needed. Separately, the Textract Core reading engine was updated (better accuracy, dense tables, handwriting, 40 languages for DetectDocumentText/AnalyzeDocument) and Core is now available in Tokyo and Zurich (18 regions total). Insights is not yet available in those two regions. It supports optional customer-managed KMS keys and uses cross-Region inference within a geography.
Entities: Amazon Web Services, Amazon Textract Insights, Amazon Textract, Amazon Textract Core, Amazon S3, AWS Key Management Service
1 primary
What happened
AWS has made Amazon Textract Insights generally available. It is a new asynchronous API (for example StartDocumentInsights) that takes one file of up to 3,000 pages from S3 and pulls out fields you describe in a natural-language JSON Schema blueprint. AWS supplies public blueprints (W-2, Form 1040, Form 1098, passport, birth certificate), or you can write your own. Each value comes back with a confidence score and page and bounding-box references, and mixed packets can optionally be split into sub-documents. Separately, the older Textract Core engine got accuracy updates (dense tables, handwriting, 40 languages) and is now in Tokyo and Zurich, making 18 regions. Insights is not yet available in those two regions.
Why it matters
Teams that process forms, ID documents and mixed packets can now describe the fields they want and skip building their own OCR, splitting and LLM pipeline. The confidence scores and bounding boxes matter most for review workflows, because staff can check only the low-confidence fields instead of whole documents. Customer-managed KMS keys help regulated buyers. The impact is moderate: this is a managed packaging of an existing pattern, and it overlaps with AWS's own Bedrock Data Automation, so the real decision for current AWS users is which of the two to standardise on.
What is noise
"Verifiable" and "auditable" are doing a lot of work. A bounding box shows where an answer came from, not that it is right, and the confidence scores and accuracy claims come from AWS alone with no independent testing. The post also gives no pricing or error rates on messy real-world scans, and says nothing about how well the confidence scores are calibrated.
Watch next
- 01Independent benchmarks on messy, handwritten or low-quality scans, including whether low confidence scores actually line up with real errors.
- 02Pricing per page against Textract Core, Bedrock Data Automation and rival services, especially for 3,000-page files and split packets.
- 03Regional rollout: when Insights reaches Tokyo, Zurich and other regions, and how cross-Region inference handles customers with data residency limits.
Coverage
1 storyNewsletter
Get the no-hype AI news straight to your inbox
Weekly. No spam, unsubscribe any time. Privacy
More capability signals
Full feed →- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- Mistral releases Mistral Large 4, a ~1T-parameter open-weight multimodal MoE model, with day-0 Hugging Face support29 Oct 202680
- Meta releases Muse Spark Open, an open-weight ~816B-parameter multimodal mixture-of-experts model, with Hugging Face support15 Oct 202680
- OpenAI discloses sandbox-escape and credential-leak incidents, confirms pause on tool-use for its most capable models26 Sept 202680