Azure AI Document Intelligence pricing,
explained for SMBs.
Per-page costs, tier breakpoints, the six hidden line items, and the three worked examples we use when scoping invoice, contract and claims pipelines for 30–300 person businesses. Written by the people who build these for a living — not by Microsoft’s procurement team.
Prices on this page are US East baseline as of May 2026 and are intended for scoping conversations, not as a substitute for Microsoft’s live pricing page. Always re-check the line items before signing a Microsoft commitment.
The six lines on the bill.
Every Document Intelligence workload bills against some combination of these six SKUs. Knowing which ones apply to your candidate workflow is the difference between a clean estimate and a 40% overrun.
Prebuilt models (Invoice, Receipt, ID, W-2, US Tax, Health Insurance Card)
Around $10 per 1,000 pages on standard tier. The first 500 pages per month are free for evaluation. This is the line you spend most of your money on for a working invoice pipeline.
Layout model (text + tables + structure)
Around $10 per 1,000 pages. The right answer when you want the document structure but don’t need a domain-specific prebuilt. Common substrate for downstream LLM reasoning over the extracted layout.
Read model (OCR-only)
Around $1.50 per 1,000 pages. The cheapest tier. Use it when you genuinely only need text, not structure, and the downstream LLM will do the parsing. Frequently the right call for contract triage.
Custom extraction models (your own template / neural)
Training: free up to 500 documents per model. Inference: around $50 per 1,000 pages. The most expensive tier per page, but the only one that fits non-standard document shapes (your specific PO, your specific claim form).
Custom classification model
Around $50 per 1,000 pages. Used when the first step is “what kind of document is this?” before routing to a model. Easy to forget when scoping; doubles the inference cost on any classified-then-extracted flow.
Add-on capabilities (query fields, key-value, barcode, formula)
Each add-on bumps the unit price ~20–30%. Cheap individually, sneaky in aggregate. We catalogue exactly which add-ons each candidate workflow needs at the design stage so the unit economics are fixed before the build starts.
Three worked examples.
What it actually costs to run an SMB-scale Document Intelligence pipeline at three volume points. Numbers are Azure consumption only — the build itself is a separate capex line.
Lean: 5,000 invoices/month
Pure AP automation, prebuilt Invoice model only
- 5,000 pages × $10 / 1,000 = $50 / month (Document Intelligence)
- Azure OpenAI reasoning on extracted fields (GPT-4o-mini, ~$15 / month at this volume)
- Logic App orchestration + storage (~$20 / month)
~$85 / month all-in. Azure consumption only. Excludes the integration build.
Mid: 25,000 mixed documents/month
Invoices + POs + contracts; routed by a classifier
- 25,000 pages × $50 / 1,000 (classifier) = $1,250 / month
- 20,000 pages × $10 / 1,000 (prebuilt + layout) = $200 / month
- 5,000 pages × $50 / 1,000 (custom contract model) = $250 / month
- Azure OpenAI reasoning ($200 / month)
- Container Apps + Service Bus + storage (~$150 / month)
~$2,050 / month all-in. The classifier line is the one most clients underestimate at scoping.
Heavy: 200,000 forms/month
Claims processing, custom neural extraction
- 200,000 pages × $50 / 1,000 (custom neural) = $10,000 / month
- Azure OpenAI reasoning at volume ($2,500 / month with caching)
- Container Apps + AKS + storage + observability ($1,500 / month)
~$14,000 / month Azure consumption. Still typically 50–70% cheaper than the SaaS alternative for this document type, but the run-rate is now a real budget conversation.
Six lines that ambush most estimates.
These are the items that turn a $500 / month estimate into a $1,200 / month bill. Catch them at scoping, not in the second month of production.
The classifier line.
If your pipeline first decides what kind of document it’s looking at, every page is billed at custom-classification rates ($50 / 1,000) before the extraction model runs. Most quick estimates miss this and come in 30–50% low.
Multi-page documents are billed per page.
“Per 1,000 documents” in your head is “per 1,000 pages” on the invoice. A 40-page contract is 40 billable pages, not one. Plan for it at the design stage by capping or summarising before extraction.
Add-ons compound, they don’t replace.
Turning on Query Fields, Key-Value pairs, and Formulas on the same call applies all the surcharges. Audit the add-ons you genuinely use; usually two of the four can come off without a quality hit.
Retraining isn’t free at scale.
The first 500 documents per custom model are free to label and train. Above that you’re paying per-label, and most teams forget that improvement cycles (v1, v2, v3 of a custom model) each consume that allowance again.
Region matters for residency and price.
A few regions price 10–15% above the US dollar baseline. EU customers needing residency in specific countries occasionally pay the premium without realising it. Pick region at design, not at deploy.
The free tier (F0) won’t carry production.
500 pages / month free is for evaluation, not production. The throttling on F0 will break a pilot that runs over a weekend batch. Move to S0 before you run a real-world test.
Three times Document Intelligence is the wrong answer.
We’d rather tell you to spend less than build the wrong thing. The line of business decides; we just call the economics honestly.
Volumes under 1,000 pages / month.
At small volumes the marginal cost is dominated by integration cost, not extraction cost. Often a $50 / month SaaS tool is the right answer until the volume justifies an Azure build.
Documents that vary by every customer.
If every document type is bespoke and the volume per type is in the dozens, the custom-model training overhead dwarfs the per-page cost. A GPT-4o pass over the raw scan is often cheaper than Document Intelligence + custom model.
Pure text with no structure.
If you genuinely don’t care about tables, key-value pairs, or layout, the Read model is the floor and a direct LLM call can sometimes undercut it. Only worth the engineering tradeoff at meaningful volume.
Before you commit to a Document Intelligence build
Find out whether your document workflow is a Document Intelligence job — or a simpler LLM-only pass.
Score your readiness, see the pattern we’d actually build for your document shape and volume, and get a fixed-scope price range. Free and instant.
- 5 minutes
- Industry-benchmarked
- No signup until report
Frequently asked.
How much does Azure AI Document Intelligence cost per page?+
Read model is around $1.50 per 1,000 pages. Layout and prebuilt models are around $10 per 1,000 pages. Custom extraction and classification models are around $50 per 1,000 pages. Add-on capabilities (query fields, key-value, barcode, formula) layer 20–30% surcharges on top. All figures are current US East baseline; check the live Azure pricing page for your region before signing a commitment.
Is there a free tier?+
Yes. The F0 free tier gives you 500 pages per month for evaluation, plus the first 500 documents of custom-model training labels. Production workloads must use S0 (standard). The free tier exists to remove the activation barrier to a proof-of-concept; it will throttle and rate-limit anything resembling production.
What’s the difference between Document Intelligence and Form Recognizer?+
They are the same product. Microsoft renamed Form Recognizer to Azure AI Document Intelligence in mid-2023. Older Azure documentation, SDKs, and SaaS comparison sites still use both names interchangeably. The pricing and capability sets are continuous across the rename.
Does Document Intelligence replace SharePoint Premium / Syntex?+
Sometimes. Syntex (now SharePoint Premium) is the right answer when your documents already live in SharePoint, the templates are standardisable, and your governance model expects SharePoint-native auth and lifecycle. Document Intelligence is the right answer when documents arrive via email, SFTP, an API, or a queue, and the downstream system isn’t SharePoint. Both can coexist; we map the decision per-workflow during discovery.
How do we size custom-model costs before training?+
Two numbers: monthly inference volume × $50 / 1,000 pages = inference run-rate; training documents above the free 500 × per-label rate = training capex. We always quote both ranges in the design phase so the unit economics are signed off before a single label is collected. If either number kills the business case, we recommend a different substrate at that point, not after the build.
What does a full Document Intelligence build cost end-to-end?+
For a 30–300 person business, a first production pipeline (intake, classification, extraction, LLM reasoning, write-back to system of record) lands at $60–120K all-in for the build, plus Azure consumption matching one of the three worked examples above. The build cost is dominated by the integration and the eval suite, not by Document Intelligence itself.
Related reading
- Azure AI Consulting — the full capability surface & engagement shape · parent pillar
- Microsoft Copilot Consulting · when the SharePoint-native answer is cheaper
- AI ROI Calculator · model the savings against the run-rate above
- The Copilot vs custom Azure ROI write-up
Want a worked estimate for your document workflow?
Tell us the document type, monthly volume, and where the extracted data needs to land. We’ll come back with a fixed-scope price range, an Azure consumption estimate, and an honest call on whether Document Intelligence is even the right substrate.