--- title: "AI Document Intelligence" slug: "ai-document-intelligence-1" updated: 2026-07-08T09:34:59Z published: 2026-07-08T09:34:59Z canonical: "azure-cost-management-playbook.turbo360.com/ai-document-intelligence-1" --- > ## Documentation Index > Fetch the complete documentation index at: https://azure-cost-management-playbook.turbo360.com/llms.txt > Use this file to discover all available pages before exploring further. # AI Document Intelligence Note Document Intelligence used to be called Form Recognizer and still often shows up as this under the hood in the billing data. ## What it does Extracts structured data from semi-structured and unstructured documents: - **Pre-built models** — ready-made extractors for invoices, receipts, ID documents, tax forms, business cards, health insurance cards, and more. - **Layout model** — extracts all text, tables, selection marks, and document structure without domain-specific training. - **Custom models** — train on a small set of labelled examples to extract fields from proprietary document formats. - **Document classification** — automatically routes documents to the right extraction model. ## When you'd use it - You have high-volume document processing (invoices, contracts, claims, onboarding forms) that is currently manual or uses brittle template-based parsing. - You need to extract data from PDFs or scanned images, not just native digital text. - You want to combine extraction with downstream automation (e.g. ERP posting, approval workflows). ## Sample use case A logistics company receives hundreds of supplier invoices per day across multiple formats. Document Intelligence's pre-built invoice model extracts vendor name, line items, totals, and PO numbers. The output is validated against purchase orders in the ERP and, if matched, posted automatically — reducing manual data entry by over 80%. ## Example reference architecture ![](https://cdn.document360.io/4b481a5c-c963-4b45-80c8-bc19500830f4/Images/Documentation/image(69).png) ## Pricing & How Costs Work Azure AI Document Intelligence is priced **per page processed**. A page is defined as up to 1 A4-size page or 500 characters of extracted text — whichever is reached first for digital PDFs. For images, each image counts as one page. | **Feature / Model** | **Pricing unit** | **Notes** | | --- | --- | --- | | Pre-built models (invoice, receipt, ID, etc.) | Per page | Higher rate than the Layout model | | Layout model | Per page | Lower cost; extracts structure without domain logic | | Custom extraction model | Per page | Same per-page rate as pre-built models; training is separate | | Custom classification model | Per page | Charged per page submitted for classification | | Custom model training | Per hour (training compute) | One-off cost each time you train or retrain a model | | Add-on: query fields | Per query field per page | Charged on top of the base per-page cost when using ad-hoc field queries | **Key pricing mechanics:** - **Free tier (F0):** 500 pages/month at no cost across all model types. - **Standard tier (S0):** Per-page pricing; volume tiers reduce the unit cost as monthly page volumes increase (tiered pricing bands, e.g. first N pages at a higher rate, subsequent pages cheaper). - **Multi-page documents:** Every page in the submitted document is billed, even if only one page contains the data you care about. Splitting documents before submission can reduce costs. - **Training cost:** Custom model training is charged once per training run, in addition to ongoing per-page inference costs. - **Async vs sync:** Both real-time (synchronous) and batch (asynchronous) API calls are billed at the same per-page rate. ## What Drives Costs | **Cost driver** | **Why it matters** | | --- | --- | | **Page volume** | The dominant driver. Every page in every submitted document is billed. | | **Document page count** | Submitting a 100-page PDF to extract one field costs 100× more than a 1-page document. | | **Model type selected** | Pre-built and custom models cost more per page than the Layout model. | | **Query fields add-on** | Each additional query field per page adds cost on top of the base model charge. | | **Training frequency** | Retraining custom models incurs a training compute charge each time. | | **Document quality** | Poor-quality scans don't reduce cost but do reduce accuracy, often leading to re-processing. | **Cost optimisation levers:** - Pre-classify and route documents so only relevant pages are submitted (split multi-document batches). - Use the Layout model for scenarios where you only need structural extraction, not the pre-built domain logic. - Avoid re-submitting already-processed documents by caching extracted results. - Batch training runs rather than retraining on every small dataset update. ## Azure Cost Data — Meters & Meter Subcategories Azure AI Document Intelligence appears under the **Cognitive Services** service family in Cost Management. Resource type: `Microsoft.CognitiveServices/accounts` with kind `FormRecognizer`. | **Meter name** | **Meter subcategory** | **What it counts** | | --- | --- | --- | | Custom Form Pages | Standard S0 | Pages processed by a custom extraction model | | Pre-built Form Pages | Standard S0 | Pages processed by a pre-built model (invoice, receipt, ID, etc.) | | Layout Pages | Standard S0 | Pages processed by the Layout model only | | Training Documents | Standard S0 | Billed per training run (compute hours for custom model training) | | Query Fields | Standard S0 | Additional charge per query field applied per page | | Classification Pages | Standard S0 | Pages submitted to a custom classification model | **Meter subcategory notes:** - The `Quantity` field in billing exports represents the number of pages (or training hours) consumed in the billing period. - If you deploy Document Intelligence in multiple regions or use multiple resource instances (e.g. dev and prod), the `ResourceId` and `ResourceLocation` fields let you split costs by environment. - Volume tiering is applied **within** a billing period; billing exports show the blended effective rate across tiers for a given meter. ## Common Developer Mistakes That Drive Up Costs - **Submitting entire multi-page documents when only specific pages are needed.** Document Intelligence bills per page regardless of how many pages contain relevant data. Submitting a 50-page contract to extract a 2-page schedule means 48 pages are billed with no value returned. Pre-splitting or identifying relevant page ranges before submission is one of the highest-impact cost controls. - **Using the pre-built invoice or receipt model when the Layout model is sufficient.** The Layout model (which extracts structure without domain-specific intelligence) costs less per page than the pre-built models. Teams that only need text and table extraction — not the semantic field labelling of the invoice model — are overpaying if they use a pre-built model. - **Re-submitting previously processed documents without caching results.** Pipelines that re-process documents on every pipeline run (e.g. a daily job that re-processes all documents in a folder rather than tracking which have already been extracted) bill for every page on every run. - **Using query fields add-ons for fields already covered by the base model.** The query fields feature adds cost per field per page. Using it to extract fields that the selected pre-built model already identifies (e.g. `vendor_name` on the invoice model) duplicates cost without providing additional information. - **Not accounting for page count in cost estimates.** Teams building cost projections often estimate by document count rather than page count. A batch of 1,000 invoices averaging 3 pages each is 3,000 billable pages, not 1,000 — a 3× underestimate. --- ## What Could Make Your Bill Go Up or Down Next Month **Could go UP:** - Document volume increases as a new supplier onboards with high invoice frequency. - A new document type is added (e.g. multi-page contracts in addition to single-page invoices), significantly increasing average page count per submission. - A pipeline bug causes documents to be re-submitted without checking whether they were already processed. - Query fields are added to extract additional data points from existing documents, adding a per-field charge on top of the base model cost. - A model retraining decision triggers re-processing of historical documents through the updated custom model. **Could go DOWN:** - Document splitting is implemented so only the relevant pages are submitted, reducing per-submission page count. - Results are cached per document hash; previously processed documents are served from cache without re-calling the API. - The pre-built model is swapped for the Layout model on a use case where semantic field labelling was not actually used downstream. - Document volume drops following a seasonal processing peak (e.g. month-end invoice runs). --- ## Most Common Optimization Techniques | **Technique** | **Mechanism** | **Typical saving** | | --- | --- | --- | | **Pre-split multi-page documents** | Submit only the relevant page range rather than the full document | Proportional to the ratio of relevant pages to total pages — can be 60–90% on large documents | | **Cache extraction results by document hash** | Skip reprocessing for documents already extracted; store results in a database keyed on file hash | Eliminates 100% of cost for reprocessed documents | | **Use Layout model instead of pre-built where appropriate** | If you only need text/table extraction (not domain-specific field labelling), the Layout model costs less per page | Meaningful per-page rate difference on high volume | | **Avoid unnecessary query fields** | Only use query fields for information not already returned by the base model | Eliminates the per-field add-on charge | | **Track processed document state** | Maintain a processing log (blob metadata, database record) to prevent duplicate submission | Eliminates reprocessing cost in pipeline reruns | | **Estimate costs on page count, not document count** | Accurate forecasting prevents budget overruns when average page count is higher than expected | Planning rather than saving — prevents reactive surprises |