On November 20, 2026, Amazon Web Services announced in its official blog that Textract Insights, a new feature of Amazon Textract, is generally available: describe the fields you want in plain language, with no model training, and pull structured data out of piles of documents — with every value traceable to its place in the source file and a confidence score attached.
It can swallow 3,000 pages at once
Textract Insights is built for mixed document packets, in three steps:
- Split: a mixed packet of up to 3,000 pages is automatically broken into separate documents and matched to templates.
- Define fields: describe the fields you want in natural language, or define them with a JSON Schema — no model training needed.
- Trace: every result carries a confidence score, a source page and a bounding box, so people can check and audit it.
At launch it supports six languages — English, French, German, Italian, Portuguese and Spanish — and is priced per page; existing Textract users need no interface changes and can call the new API directly.
The selling point is verifiability, not just reading text
Unlike traditional OCR, which only reads characters, Textract Insights aims to understand what a field means: the amount on an invoice, the date in a contract, the name on an expense form — and then prove which page and position each figure came from. That is exactly what regulated industries such as finance, insurance and legal need: results that can be audited are results that can go into a business workflow. The limits are worth stating: Chinese is not among the six launch languages; Insights has not yet reached the two new regions, Tokyo and Zurich — Textract Core has expanded to 18 regions and 40 languages, but the new feature still has to wait; and for poor-quality scans, key fields still need a human check.
Two things are worth watching next: when Chinese support arrives, and when the new regions open. Teams already using Textract for invoice reconciliation, contracts and expense forms can first trial it on non-Chinese files, checking whether the confidence scores and traceability meet their internal audit requirements before putting it into production.