Study guide

5.3 Azure AI Document Intelligence

Azure AI Engineer Associate AI-102

Track your progress as you read

Finished sections get a check mark.

1 Introduction

About the Azure AI Engineer Associate (AI-102) Exam Study Strategies and Exam Tips for the AI-102

2 Domain 1: Plan and Manage an Azure AI Solution (20-25%)

1.1 Azure AI Services Overview and Selection 1.2 Resource Provisioning and Configuration 1.3 Security and Access Management 1.4 Responsible AI Principles and Implementation 1.5 Azure AI Solution Architecture Patterns

3 Content Safety and Moderation (within Plan and Manage, Domain 1)

2.1 Azure AI Content Safety Overview 2.2 Prompt Shields and Adversarial Attack Detection 2.3 Text and Image Moderation Implementation 2.4 Content Filtering in Azure OpenAI Service

4 Domain 4: Implement Computer Vision Solutions (10-15%)

3.1 Azure AI Vision — Image Analysis 4.0 3.2 Azure AI Face Service 3.3 Azure AI Custom Vision 3.4 Optical Character Recognition (OCR) 3.5 Video Analysis and the Retired Spatial Analysis

5 Domain 5: Implement Natural Language Processing Solutions (15-20%)

4.1 Azure AI Language — Text Analytics Features 4.2 Conversational Language Understanding (CLU) 4.3 Custom Text Classification and Custom NER 4.4 Azure AI Speech Services 4.5 Azure AI Translator Service 4.6 Question Answering and Conversational AI

6 Domain 6: Implement Knowledge Mining and Information Extraction Solutions (15-20%)

5.1 Azure AI Search — Fundamentals and Architecture 5.2 AI Enrichment with Skillsets 5.3 Azure AI Document Intelligence

7 Domain 2: Implement Generative AI Solutions (15-20%)

6.1 Azure OpenAI Service — Models and Deployment 6.2 Chat Completions API and Prompt Engineering 6.3 Embeddings and Vector Search for RAG 6.4 Fine-Tuning and Model Customization 6.5 Image Generation (gpt-image; DALL-E 3 Retired) 6.6 RAG After On Your Data, and AI Agents

8 Domain 3: Implement an Agentic Solution (5-10%)

3.1 Agentic Solutions: Foundry Agent Service and the Microsoft Agent Framework

9 Exam Review: Cross-Domain Topics and Advanced Practice

7.1 Service Comparison and Selection Guide 7.2 SDK and REST API Patterns 7.3 Responsible AI Across All Domains 7.4 End-to-End Solution Architectures 7.5 Exam Day Quick Reference

5.3 Azure AI Document Intelligence

Key Takeaways Explain This Topic AI

Azure AI Document Intelligence (formerly Form Recognizer) extracts text, key-value pairs, tables, selection marks, and typed fields; the v4.0 / 2024-11-30 API and Document Intelligence Studio are the current tooling.

Read and Layout require no training; Layout adds tables, selection marks, and structure, while prebuilt models (invoice, receipt, ID, W-2, 1099, health-insurance, contract) return typed fields with per-field confidence.

Custom models come in two build modes: template (consistent fixed layouts) and neural (varying layouts); both need a minimum of FIVE labeled training documents in a Blob container reached via SAS.

In v4.0 you build a custom classifier and combine it with extraction models so a single call classifies a document then routes it to the right model (the modern replacement for v3.0 composed-model auto-routing).

Every returned field carries a confidence score; production pipelines branch on a confidence threshold and route low-confidence extractions to human review.

Last updated: October 2026

Quick Answer: Document Intelligence (formerly Form Recognizer) extracts structured data from documents. Use Read for text, Layout for tables and structure, prebuilt models for common documents (invoices, receipts, IDs), and custom models for your own forms. Custom models are template (fixed layout) or neural (varying layout) and need at least five labeled samples. Combine a custom classifier with extraction models to auto-route mixed document types. Every field returns a confidence score.

Current Edition and Tooling

The service uses the v4.0 (2024-11-30 GA) REST API and SDKs, edited in Document Intelligence Studio (a no-code labeling portal). The Python client is DocumentIntelligenceClient , and analysis is asynchronous: begin_analyze_document returns a poller, and poller.result() blocks for the result. The old name Form Recognizer still appears in legacy docs and may show up as a distractor on the exam.

Choosing the Right Model

  • Model Training? Returns
  • prebuilt-read No Text, language, handwriting detection
  • prebuilt-layout No Text + tables + selection marks + structure
  • prebuilt-invoice No Vendor, customer, totals, tax, line items
  • prebuilt-receipt No Merchant, date, items, subtotal, tax, tip, total
  • prebuilt-idDocument No Name, DOB, address, doc number, expiry
  • prebuilt-tax.us.w2 / 1099 / 1098 No Typed US tax-form fields
  • prebuilt-businessCard (deprecated in v4.0; v3.1 and earlier only) No Name, title, company, phone, email
  • prebuilt-healthInsuranceCard.us No Insurer, member ID, group, plan
  • Custom (template/neural) Yes Your labeled fields

Decision shortcuts the exam rewards: only need raw text -> Read ; need tables and checkboxes from any document -> Layout ; the document is a standard invoice/receipt/ID -> the matching prebuilt ; it is your company's unique form -> custom .

  • Read vs Layout vs General poller = client.begin_analyze_document(model_id="prebuilt-layout", body=f)
  • result = poller.result()

for table in result.tables:

print(table.row_count, "x", table.column_count)

for page in result.pages:

for mark in page.selection_marks:

print(mark.state) # "selected" or "unselected"

Layout is the only no-training model that returns tables and selection marks (checkboxes), so any scenario mentioning checkbox state or table cells points to Layout, not Read.

Prebuilt Models with Confidence

Prebuilt models return strongly-typed fields, each with a confidence score between 0 and 1. Production code branches on a threshold and sends low-confidence results to human review. poller = client.begin_analyze_document(model_id="prebuilt-invoice", body=f)

for doc in poller.result().documents:

total = doc.fields.get("InvoiceTotal")

if total and total.confidence >= 0.80:

post_to_erp(total.content)

else:

route_to_human_review(doc) # low confidence

Trap: the confidence value is a per-field reliability measure, not an accuracy guarantee. Exam scenarios that say "automatically post invoices over a reliability bar, otherwise queue for a person" are testing whether you branch on field.confidence against a threshold.

Custom Models: Template vs Neural

Both build modes need a minimum of five labeled training documents in a Blob container referenced by a Shared Access Signature (SAS) URL.

Build mode Layout assumption Best for Notes

template Consistent, fixed positions Standardized forms, applications Faster, cheaper, single language per model

neural Varying layouts Contracts, letters, multi-vendor forms Deep learning, better generalization, longer training poller = client.begin_build_document_model(

  • build_mode="neural", # or "template"
  • blob_container_url="https://acct.blob.core.windows.net/training?<SAS>",
  • model_id="po-neural-v1")
  • model = poller.result()

If documents arrive from many vendors with shifting layouts, the answer is neural ; if they are one rigid in-house form, template is cheaper and faster.

Classifiers and Routing Mixed Documents

When a single pipeline receives mixed document types, you build a custom classifier ( begin_build_classifier ) and then combine the classifier with extraction models so one begin_classify_document (or a classifier-driven analyze) classifies first, then routes each document to the correct extraction model. This v4.0 classifier approach is the modern successor to v3.0 composed models , which combined multiple custom extraction models behind one model ID and auto-selected among them. poller = client.begin_classify_document(

classifier_id="mailroom-classifier", body=document_bytes)

for doc in poller.result().documents:

print(doc.doc_type, doc.confidence) # e.g. "invoice" 0.97

  • Need Use
  • Sort/triage documents into categories only Custom classifier
  • Extract typed fields from one known type Prebuilt or custom extraction model
  • One endpoint that sorts THEN extracts mixed types Classifier + extraction models (formerly composed model)
  • Input Constraints and Security

Know the practical limits the exam likes to probe. Supported inputs include PDF, JPEG/JPG, PNG, BMP, TIFF, HEIF , and Office formats for some models. The free (F0) tier caps analysis at the first 2 pages of a document, while paid (S0) processes the whole file up to service limits. Training data must sit in Azure Blob Storage reached through a SAS URL (or a managed identity); the model itself never stores your source documents after training.

For security, prefer Microsoft Entra ID (managed identity) over the resource key so credentials are not embedded in code, and scope storage access with role-based access control . Use a customer-managed key when regulatory requirements demand control of encryption at rest.

On the Exam: A scenario describing a mailroom that receives invoices, receipts, and contracts and must process each correctly through one entry point is the classic classifier-plus-routing (composed-model) question — no manual pre-sorting step is needed. "Five documents" is the memorized minimum for custom training, and "varying layouts" always signals the neural build mode over template.