AI-901
Deep Dive
"Implement AI solutions for information extraction by using Foundry" is the exam skill covered by this page. Content Understanding, within Foundry Tools, ingests documents, images, audio, and video, and extracts key, reliable information to power generative AI solutions and agents.
Contents
Content Understanding is first-class in the Microsoft Foundry portal — it is not a separate service you need to go find. It ingests the 4 content modalities and returns structured data ready for an agent or a generative app, using prebuilt analyzers for common scenarios.
📄
Documents
Fields, tables, and text from forms and contracts — the piece covered by Document Intelligence, detailed below.
🖼️
Images
Objects, embedded text, and visual metadata.
🎧
Audio
Transcription, topics, and mentioned entities.
🎬
Video
Key moments, transcription, and detected objects.
Key difference
Content Understanding is optimized for structured extraction — not for conversing about the content. If the scenario needs a model to "interpret and respond" about an image or an audio clip, the answer is a multimodal model, not Content Understanding.
Azure AI Document Intelligence (formerly Form Recognizer) is a service that uses ML models to extract text, tables, and key-value pairs from documents in a structured way. It doesn't just convert images to text (OCR) — it understands the semantics of the document and extracts specific fields with their values.
🔍
Advanced OCR
Extracts text from PDFs, images, scanned documents — even handwritten and low-quality documents.
🧠
Semantic understanding
Identifies fields by meaning, not just position. "Total" can be in different places depending on the vendor.
📊
Table extraction
Detects and extracts tables within documents, returning structured cells, rows, and columns.
Prebuilt models are trained by Microsoft on millions of real documents. They can be used immediately without your own training data.
Invoice
Extracts fields from invoices: vendor, date, invoice number, line items, total, taxes, bank.
Vendor invoices, corporate purchases.
Receipt
Extracts fields from purchase receipts: merchant, date, time, items, subtotal, tax, total, payment method.
Employee expenses, reimbursements, audits.
Identity Document (ID)
Extracts data from identity documents: ID card, passport, driver's license. Name, surname, DOB, document number, country.
Customer onboarding, KYC verification.
Business Card
Extracts contacts from business cards: name, company, email, phone, address, website.
CRM, digitizing business contacts.
W-2 (US)
Extracts data from the US W-2 tax form: employee, employer, income, withholdings.
Tax processing, payroll.
Contract
Extracts contract metadata: parties, start/end dates, renewal clauses, contract type.
Legal contract management, compliance.
Health Insurance Card
Extracts data from health insurance cards: insured party, policy number, group, plan, dates.
Medical processes, hospital admissions.
General Document
Extracts text, tables, figures, and key-value pairs from generic documents without a specific format.
Any document not covered by specific models.
Exam trap: service name
The service was called Form Recognizer until 2023. It's now called Azure AI Document Intelligence. The exam may use either name. They are the same service.
When no prebuilt model covers your document type (e.g., internal purchase orders, proprietary technical reports), you can train a custom model with examples of your labeled documents.
Custom Template
For documents with a fixed, predictable structure. Fields are always in the same relative positions.
Internal forms, documents with a fixed template.
Custom Neural
For documents with variable layouts. Uses deep learning to understand semantic context, not just position.
Contracts from different vendors, documents with format variations.
Composed Model
Combines multiple custom models. The service automatically detects which model applies to each document.
Processing multiple document types with a single endpoint.
Gather 5+ sample documents (recommended: 50+ for better accuracy).
Upload to Azure Blob Storage.
Label fields in Document Intelligence Studio (web labeling interface).
Launch training — minutes in the cloud.
Evaluate accuracy per field and re-label if needed.
Call the API with the trained model's model ID.
Minimum requirement
5 labeled documents is the absolute minimum for training. The model improves noticeably with 50+ examples. For Custom Neural, 500+ are recommended for high accuracy.
| Capability | Document Intelligence | Basic OCR (Read API) |
|---|---|---|
| Extracts text | ✅ Yes, advanced | ✅ Yes |
| Understands fields (e.g., "Total = $150") | ✅ Yes | ❌ No |
| Extracts structured tables | ✅ Yes (rows and columns) | ⚠️ Partial (text only) |
| Returns structured JSON | ✅ Yes, typed fields | ❌ Plain text only |
| Document-type-specific models | ✅ Invoice, Receipt, ID... | ❌ Generic |
| Custom models | ✅ Yes | ❌ No |
| Price | Higher (per page) | Lower (per call) |
Rule for the exam
If the scenario only needs to read text from an image or PDF → OCR / Read API.
If it needs to extract specific fields (date, total, vendor name) → Document Intelligence.
Key phrase: "structured data extraction from documents" → Document Intelligence.
Example: Invoice model response
// POST /analyze with invoice image
{
"documents": [{
"docType": "invoice"
"fields": {
"VendorName": {"value": "Contoso Ltd.", "confidence": 0.99}
"InvoiceDate": {"value": "2024-03-15", "confidence": 0.97}
"InvoiceId": {"value": "INV-2024-001", "confidence": 0.99}
"SubTotal": {"value": {"amount": 1200.00, "currencySymbol": "€"}}
"TotalTax": {"value": {"amount": 252.00}}
"InvoiceTotal": {"value": {"amount": 1452.00}}
}
}]
}
Each field includes the extracted value, data type, and confidence score. It's not plain text — it's structured JSON ready to consume.
❓ Can Document Intelligence process handwritten documents?
Yes. The underlying Read API model handles both printed and handwritten text. Prebuilt models can also process hand-scanned documents, although accuracy depends on legibility.
❓ Do you need prior training to process invoices?
No. Document Intelligence's prebuilt "Invoice" model is already trained on millions of invoices from multiple vendors and countries. You don't need your own training data for most standard invoice formats.
❓ Do Document Intelligence and Azure AI Vision (OCR) do the same thing?
Not entirely. OCR (Azure AI Vision's Read API) extracts plain text from images. Document Intelligence understands the document's semantics and extracts structured fields with types and values. For "read text" → OCR. For "extract the total from an invoice" → Document Intelligence.
Did you get this topic?
Test what you just learned
An insurance company receives hundreds of claim forms in PDF every day. They need to automatically extract: policy number, date of loss, insured party's name, and claimed amount to enter them into their system. Which service is most appropriate?
Put what you learned to the test with original AI-901 exam questions.
Start AI-901 mock examUnlimited mock exams with the Pro plan
Log in to track your progress.