What Azure AI Document Intelligence does
Azure AI Document Intelligence provides OCR, layout analysis, prebuilt and custom extraction, classification, query fields, and batch document processing through APIs and containers.
Azure AI Document Intelligence is the service formerly known as Form Recognizer. It converts scans and native documents into machine-readable content without requiring a team to build OCR, reading-order, table, or key-value models from scratch. Prebuilt models cover common documents, while custom extraction and classification address organization-specific layouts. Query fields can request selected information, batch routes improve high-volume economics, and connected containers support some deployment needs. Output includes provenance that applications should retain: page numbers, polygons, spans, and confidence make it possible for a reviewer to compare a value with the source rather than trusting an isolated JSON string.
The Free tier processes up to 500 pages per month and excludes some premium capabilities. S0 is pay-as-you-go by 1,000 pages, with different meters for Read, prebuilt, custom extraction, custom generative extraction, classification, query fields, add-ons, training, and batch operations. Commitment tiers exchange an upfront monthly amount for included high-volume pages plus overage pricing. Microsoft's public page localizes rates and did not expose stable dollar figures in this review, so use the Azure calculator for the selected region and agreement. Add Blob Storage, networking, monitoring, review labor, containers, retries, and downstream compute to the estimate.
Confidence is a model signal, not the probability that a payment or identity decision is safe. Benchmark exact-match and tolerance-based accuracy per field, vendor, language, scan quality, handwriting, and document version. Enforce types, arithmetic, duplicate detection, purchase-order matching, and authoritative-record checks outside the model. Documents often contain PII, bank data, health information, signatures, and secrets: establish a lawful purpose, minimize collection, use managed identities and private networking where required, encrypt storage, restrict logs, retrieve then delete temporary results, and define custom-model deletion. Microsoft recommends human oversight; require source-image approval for low-confidence, novel, conflicting, high-value, regulated, or irreversible outcomes.
How Azure AI Document Intelligence works
A client sends a supported image or document to an Azure Document Intelligence resource and selects Read, Layout, a prebuilt model, a classifier, or a custom model version. OCR detects printed and handwritten text; layout analysis returns paragraphs, selection marks, tables, page geometry, and relationships; extraction models map content to named fields with values, normalized forms, coordinates, and confidence. Asynchronous jobs return an operation identifier that the client polls before retrieving JSON. Custom models learn from labeled examples in customer-controlled Blob Storage, and composed or classification routes can select among document types. Azure stores input and analysis results temporarily in the resource region so jobs can be retrieved; completed responses remain available for 24 hours unless deleted earlier through the API.
Define documents, fields, and privacy
Inventory document classes, lawful purpose, PII, regions, retention, encryption, identities, expected fields, source-of-truth rules, error costs, confidence thresholds, approvers, and deletion evidence.
Run the narrowest suitable model
Use Read, Layout, a prebuilt model, or a versioned custom model; preserve page coordinates and confidence, reject unsupported files, isolate keys, and delete temporary analyze results when no longer needed.
Compare output with source evidence
Apply types, totals, cross-field and master-data checks, then route low-confidence, conflicting, consequential, novel, or tampered documents to a reviewer who can see the original page.
Monitor quality and downstream effects
Track field accuracy by document slice, exceptions, drift, latency, pages, cost, access, retention, and corrections; reconcile writes and payments, pin model versions, and retain rollback.
How to set up Azure AI Document Intelligence
Define fields and decision risk
Inventory document types, PII, lawful purpose, expected fields, tolerances, downstream actions, error impact, review thresholds, regions, retention, deletion, and accountable owners.
Create a protected Azure resource
Select region and tier, prefer Entra authentication and managed identities, restrict keys and networks, configure storage, encryption, logging, budgets, and separate environments.
Select and benchmark models
Compare Read, Layout, prebuilt, and custom versions on representative holdouts including rotated, faint, handwritten, multilingual, blank, altered, and unseen layouts.
Build deterministic validation
Preserve polygons and confidence, enforce schema, totals, dates, identifiers, duplicates, master-data and cross-document matches, and route exceptions with the original page.
Operate with human approval
Review consequential fields before action, monitor accuracy, drift, latency, pages, cost, access, retention, and corrections, delete results promptly, pin versions, and reconcile writes.
Azure AI Document Intelligence FAQs
Is Azure AI Document Intelligence free?
A Free tier includes up to 500 pages per month with limitations. Paid meters vary by model, feature, volume, region, agreement, and deployment.
How long are analysis results stored?
Microsoft says completed analysis responses are stored for 24 hours for retrieval and can be permanently purged sooner using Delete Analyze Result.
Does a high confidence score guarantee accuracy?
No. Thresholds must be calibrated on representative documents and combined with deterministic checks and human review based on field impact.
Can it train custom extraction models?
Yes. Custom models use authorized labeled samples from customer Blob Storage; teams must manage dataset rights, quality, privacy, versioning, and deletion.
Should invoice output trigger automatic payment?
Not without controls. Match vendors, bank details, totals, tax, purchase orders, duplicates, approval limits, and source images, then require accountable payment authorization.
Listing reviewed 2026-07-15. Product details and pricing can change; verify important terms on the provider's website.
Related Developer Tools AI tools
Related AI guides
Reviews
Tell the community what you made, what worked, and what you wish you knew before starting.