You Can Pass This Exam For Free
Choose Your Study Path
Limited Azure or AI development experience. You need to build foundational knowledge of cloud AI services, Python SDK usage, and generative AI concepts before diving into Foundry-specific topics.
Week 1Learn Azure fundamentals: resource groups, subscriptions, managed identity, RBAC, Key Vault, and networking basics. Set up an Azure free account and explore the portal
Week 2Study generative AI foundations: LLMs, SLMs, multimodal models, prompt engineering (zero-shot, few-shot, chain-of-thought), tokenization, temperature, top-p, and model parameters
Week 3Deep dive into Microsoft Foundry: project creation, model catalog, model deployment, Foundry Tools overview, and Foundry SDK basics in Python
Week 4Study RAG architecture end-to-end: Azure AI Search indexing, vector search, hybrid search, semantic ranking, chunking strategies, and embedding generation
Week 5Learn Foundry Agent Service: agent roles, goals, conversation memory, tool schemas, function calling, and multi-agent orchestration patterns
Week 6Study computer vision: image generation with DALL-E, inpainting, multimodal understanding, Content Understanding pipelines, alt-text generation, and video analysis
Week 7Cover text analysis and information extraction: entity extraction, sentiment analysis, speech-to-text, text-to-speech, Azure Translator, Document Intelligence, and OCR
Week 8Study responsible AI: safety filters, content moderation, guardrails, evaluators, trace logging, provenance metadata, and agent oversight modes
Week 9Practice questions across all domains. Take full mock exams and review explanations. Focus on generative AI and agentic solutions (33% of exam)
Week 10Final review: re-study weak areas, review all service comparisons, ensure you can map each Azure service to its correct use case
Exam Overview
Format
Approximately 50 questions (40-60 range), 120 minutes. Multiple choice, multiple select, drag-and-drop, case studies, and scenario-based questions.
Scoring
Scaled score 100-1000. Passing: 700. No penalty for wrong answers — always answer every question.
Domains & Weights
- Plan and Manage an Azure AI Solution28%
- Implement Generative AI and Agentic Solutions33%
- Implement Computer Vision Solutions13%
- Implement Text Analysis Solutions13%
- Implement Information Extraction Solutions13%
Registration
$165 USD. Available at Pearson VUE testing centers or online proctored from home. Exam fee is $165 USD. The exam launched in beta in early 2026 and is transitioning to general availability in June 2026.
Topic Priority Table
Not all topics are tested equally. Focus your study time on Tier 1 first, then Tier 2. Tier 3 topics rarely appear — just recognize what they do.
Domain 128% of exam
Plan and Manage an Azure AI Solution
This domain covers selecting the right Foundry services, setting up AI infrastructure, managing and monitoring deployments, and implementing responsible AI practices. It tests your ability to design Azure AI architectures, choose appropriate models, configure security, and apply responsible AI controls across generative AI and agentic systems.
Key Topics
Microsoft FoundryAzure OpenAI ServiceAzure AI SearchFoundry Agent ServiceFoundry ToolsManaged IdentityKey VaultContent Moderation
Must-Know Concepts
- How to choose the right model for each task: LLMs for broad reasoning, SLMs for cost-efficient domain tasks, multimodal models for vision and audio, and code models for development assistance
- Foundry service selection: Azure OpenAI for generation, Azure AI Search for retrieval and grounding, Foundry Agent Service for agent workflows, Content Understanding for multimodal processing
- Retrieval and indexing methods: vector search (embedding similarity), semantic search (AI reranking), hybrid search (keyword + vector), and how to configure each in Azure AI Search
- Agent solution architecture: memory services for conversation tracking, tool integration patterns, knowledge store connections, and function calling schemas
- Azure infrastructure design: resource groups, Foundry projects, deployment options, and CI/CD pipeline integration for AI apps and agents
- Quota management: scaling policies, rate limits, token quotas, and cost footprint management for model and agent workloads
- Monitoring: model performance drift, safety events, grounding quality, data ingestion health, search index relevance, and token analytics
- Security configuration: managed identity, private networking, keyless credentials, RBAC role policies, and Key Vault integration
- Responsible AI: safety filters, guardrails, risk detection, content moderation, evaluators, safety evaluations, trace logging, and provenance metadata
- Agent governance: oversight modes (autonomous vs semi-autonomous), behavioral constraints, tool-access controls, and approval workflows
Common Exam Traps
Managed identity and keyless credentials are the recommended security pattern. API keys are acceptable for development but the exam expects you to choose managed identity for production scenarios
Foundry projects have specific deployment options — know the difference between serverless deployment, managed compute, and provisioned throughput
Content moderation and safety filters are configured at the model deployment level within Azure OpenAI, not at the individual prompt level. A single resource can have multiple deployments each with different content filtering policies. Know where each control is applied
CI/CD integration with Foundry requires connecting to the project, not just the individual service. The exam tests whether you understand project-level orchestration
Monitoring grounding quality is different from monitoring model performance. Grounding quality measures whether retrieved data is relevant; model performance measures generation accuracy
Domain 233% of exam
Implement Generative AI and Agentic Solutions
The heaviest domain at 33% — expect roughly 16-17 questions. Covers building generative applications with Foundry, deploying and consuming models, implementing RAG, building agents with memory and tools, multi-agent orchestration, and optimizing generative AI systems. This domain tests hands-on implementation skills with Foundry SDKs and agent frameworks.
Key Topics
Azure OpenAI ServiceFoundry Agent ServiceAzure AI SearchFoundry SDKPrompt FlowRAGFunction CallingMulti-Agent Orchestration
Must-Know Concepts
- Deploying and consuming different model types: LLMs, SLMs, code models, and multimodal models through Foundry
- RAG implementation: chunking documents, generating embeddings, indexing in Azure AI Search, configuring retrieval, and integrating results into prompts
- Workflow design: tool-augmented flows, multistep reasoning pipelines, and chain-of-thought patterns
- Model and app evaluation: detecting fabrications (hallucinations), measuring relevance, quality, and safety scores using Foundry evaluators
- Foundry SDK integration: connecting applications to Foundry projects, using connectors, and calling model endpoints
- Agent definition: roles, goals, conversation-tracking approaches, and tool schemas for structured interaction
- Agent tool integration: APIs, knowledge stores, search, Content Understanding, and custom functions as agent tools
- Multi-agent orchestration: coordinating multiple specialized agents, routing between agents, and managing shared context
- Autonomous vs semi-autonomous workflows: implementing safeguards, approval flow controls, and human-in-the-loop patterns
- Agent monitoring: evaluating agent behavior, error analysis, and integrating monitoring into deployed agents
- Prompt engineering optimization: adjusting temperature, top-p, frequency penalty, presence penalty, and max tokens
- Advanced generation techniques: model reflection, chain-of-thought evaluations, self-critique loops, and hybrid LLM-rules engines
Common Exam Traps
Function calling is NOT the same as prompt injection. Function calling is a structured, authorized way for agents to invoke tools. The exam tests whether you understand the security implications of each
Multi-agent orchestration requires shared context management. Each agent has its own memory, but they need a coordination layer to pass information between them
Approval workflows are required for semi-autonomous agents. Fully autonomous agents are acceptable only when the risk level is low and appropriate safeguards are in place
Fabrication detection (hallucination checking) is an evaluation step, not a guardrail. Guardrails constrain outputs; evaluators measure quality after generation
Foundry SDK connectors are different from REST API calls. The SDK provides typed, authenticated access; REST APIs require manual authentication and response parsing
Chain-of-thought is a prompt engineering technique, not a model architecture. It works with any LLM but produces better results with more capable models
Domain 313% of exam
Implement Computer Vision Solutions
This domain covers image and video generation, multimodal understanding, and responsible AI for visual content. Expect questions on DALL-E image generation, inpainting, Content Understanding pipelines, visual question answering, accessibility-focused alt-text generation, and content moderation for visual output.
Key Topics
DALL-EAzure AI VisionContent UnderstandingMultimodal ModelsGPT-4 VisionVideo Analysis
Must-Know Concepts
- Image generation with DALL-E: text-to-image, inpainting, mask-based edits, and prompt-driven modifications
- Video generation from text prompts and reference media, and workflows for editing generated videos
- Generation and editing controls: resolution, style, quality settings, and platform-provided parameters
- Multimodal understanding: analyzing visual context using GPT-4 Vision and other multimodal models
- Caption generation: configuring concise vs detailed captions for single or multiple images
- Visual question answering: implementing solutions that answer questions grounded in visual evidence
- Accessibility: generating alt-text and extended image descriptions aligned to accessibility guidelines
- Content Understanding pipelines: single-task (standard) mode for images, video, audio, and documents; pro mode is documents-only with multi-step reasoning and multi-input support
- Video analysis: processing and interpreting video segments, identifying objects and regions within video
- Responsible AI for vision: filtering unsafe visual content, detecting prompt injection in embedded image text, enforcing watermarks and brand rules
Common Exam Traps
Inpainting and mask-based edits are different operations. Inpainting fills in missing or selected areas. Mask-based edits use a mask to specify exactly which regions to modify
Content Understanding single-task (standard) mode supports all content types (documents, images, audio, video) with lower cost and latency. Pro mode supports documents only but adds multi-step reasoning, multi-input document processing, and reference data integration. The exam objective uses the term 'single-task and pro-mode' — know both names
Alt-text generation must follow accessibility guidelines, not just describe the image. The exam tests whether you know the difference between a generic caption and accessibility-compliant alt-text
Embedded text in images can be used for indirect prompt injection. The exam tests whether you know to scan image text for malicious instructions
Video analysis in Foundry uses Content Understanding pipelines, not just Azure Video Indexer. Know which tool to use for which video processing task
Domain 413% of exam
Implement Text Analysis Solutions
This domain covers text analysis using both generative prompting and Foundry Tools, including entity extraction, sentiment analysis, translation, and speech services. Expect questions on when to use LLM-based analysis vs dedicated Foundry Tools, speech-to-text integration with agents, and multimodal reasoning from audio inputs.
Key Topics
Azure AI LanguageAzure AI SpeechAzure TranslatorAzure OpenAI ServiceFoundry ToolsCustom Speech Models
Must-Know Concepts
- Entity extraction, topic identification, and summarization using both generative prompting and Foundry Tools — know when to use each approach
- Structured JSON output extraction from text using LLMs with defined output schemas
- Sentiment and tone detection: configuring models to identify sentiment, tone, safety issues, and sensitive content
- Text translation: Azure Translator in Foundry Tools vs LLM-powered translation flows — know the tradeoffs
- Domain-specific customization: compliance summarization, domain extraction, and custom language model outputs
- Speech-to-text and text-to-speech: implementing workflows for agentic interactions where speech is an input/output modality
- Custom speech models: training and deploying speech models for domain-specific vocabulary and accents
- Multimodal reasoning from audio: extracting insights and making decisions based on audio input data
- Speech translation: converting speech into other languages using language models and Foundry Tools
- Integrating speech as an agent modality for hands-free or accessibility-focused interactions
Common Exam Traps
LLM-based text analysis (generative prompting) is more flexible but more expensive than Foundry Tools. Use Foundry Tools for high-volume, standardized tasks. Use LLMs for complex, nuanced analysis
Azure Translator provides deterministic, high-quality translation for supported languages. LLM-powered translation is more flexible but less consistent. The exam tests when to choose each
Custom speech models require training data. They are not just configuration changes — you need domain-specific audio and text pairs
Structured JSON output from LLMs requires explicit schema definition in the prompt or API call. The model does not automatically produce structured output without guidance
Speech-to-text for agents is bidirectional — you need both STT (input) and TTS (output) for a complete voice-enabled agent interaction
Domain 513% of exam
Implement Information Extraction Solutions
This domain covers building retrieval and grounding pipelines for RAG, extracting content from documents, and producing structured outputs for downstream AI reasoning. Expect questions on content ingestion, search index configuration, enrichment skills, Document Intelligence models, and Content Understanding analyzers.
Key Topics
Azure AI SearchDocument IntelligenceContent UnderstandingOCRRAG PipelinesEnrichment Skills
Must-Know Concepts
- Content ingestion: indexing documents, images, audio, and video for grounding and retrieval
- Search configuration: semantic search, hybrid search, and vector search setup and when to use each
- Enrichment pipelines: using built-in and custom skills for text, image, and layout enrichment during indexing
- RAG ingestion flow: document chunking, OCR processing, embedding generation, and index population
- Connecting retrieval pipelines to agent tools and generative workflows as knowledge sources
- Document Intelligence: extracting information using multimodal pipelines combining OCR, layout analysis, and field extraction
- Content Understanding analyzers: generating structured or markdown outputs for downstream reasoning
- Producing clean, grounded representations for agents and RAG using Content Understanding
- Prebuilt vs custom models in Document Intelligence: invoices, receipts, IDs, business cards, and custom trained models
- Single-task (standard) vs pro mode Content Understanding pipelines: single-task covers all content types with lower cost; pro mode is documents-only but adds multi-step reasoning, multi-input document support, and reference data integration
Common Exam Traps
Enrichment skills run DURING indexing, not at query time. If you need real-time processing, you need a different approach
Vector search requires pre-computed embeddings. You cannot do vector search on raw text — the text must be converted to embeddings first during ingestion
Document Intelligence prebuilt models cover common document types (invoices, receipts). Custom models are needed for proprietary document formats
Content Understanding analyzers can generate both structured JSON and markdown outputs — you configure the output format when defining the analyzer schema. The default behavior depends on analyzer configuration, not a fixed default
RAG ingestion flow must include OCR for scanned documents. Without OCR, the indexer cannot extract text from images or scanned PDFs
Azure AI Services You Must Not Confuse
These pairs appear on nearly every exam. Learn the difference and you'll avoid the most common traps.
Top Mistakes to Avoid
Confusing Azure OpenAI (generation) with Azure AI Search (retrieval) — OpenAI generates responses, Search retrieves data to ground them. RAG uses both together
Thinking RAG and fine-tuning are interchangeable — RAG retrieves data at query time without changing the model, while fine-tuning modifies model weights and requires retraining
Mixing up Content Understanding (broad multimodal processing) with Document Intelligence (specialized document extraction) — Content Understanding is broader but Document Intelligence is more precise for structured documents
Using API keys in production instead of managed identity — the exam strongly favors keyless credentials with RBAC for production deployments
Confusing safety filters (input side) with guardrails (output side) — safety filters block harmful prompts before they reach the model, guardrails constrain model outputs after generation
Not knowing the difference between vector search, semantic search, and hybrid search — vector uses embeddings, semantic reranks keyword results, hybrid combines keyword and vector approaches
Treating Foundry Agent Service and Prompt Flow as the same thing — agents are autonomous and decide which tools to use, Prompt Flow pipelines are deterministic with developer-defined sequences
Assuming Content Understanding single-task (standard) and pro modes are the same — pro mode adds multi-step reasoning and multi-input support but is documents-only, while single-task mode covers all content types (documents, images, audio, video) with lower cost and latency
Forgetting that enrichment skills run during indexing, not at query time — this is a common trap in questions about real-time vs batch processing
Thinking speech-to-text alone creates a voice-enabled agent — you need both STT for input and TTS for output to enable full voice interactions
Exam-Ready Checklist
Can explain all 5 exam domains and their relative weights (28%, 33%, 13%, 13%, 13%)
Know how to select the right Foundry service for each scenario: OpenAI for generation, AI Search for retrieval, Agent Service for agents, Content Understanding for extraction
Can implement a complete RAG pipeline: document ingestion, chunking, embedding, indexing, retrieval, and prompt augmentation
Understand agent architecture: roles, goals, memory, tool schemas, function calling, multi-agent orchestration, and approval workflows
Can distinguish between vector search, semantic search, and hybrid search and know when to use each
Know managed identity vs API key authentication and why managed identity is preferred for production
Can configure responsible AI controls: safety filters (input), guardrails (output), evaluators, trace logging, and provenance metadata
Understand Content Understanding single-task (standard) vs pro modes: single-task covers all content types, pro is documents-only with multi-step reasoning
Can implement speech-to-text and text-to-speech for voice-enabled agent interactions
Know Document Intelligence prebuilt models vs custom models and when each is appropriate
Can evaluate AI applications for fabrication, relevance, quality, and safety using Foundry evaluators
Understand image generation with DALL-E: text-to-image, inpainting, mask-based editing, and responsible AI filters for visual content
Scored 70%+ on at least two full mock exams (700/1000 passing score)
Completed at least one hands-on lab building a RAG application and one deploying an agent with function calling
Recommended Resources
Free & Official Resources
Paid Courses & Practice Exams
These are recommended if you prefer a structured learning path. They can save time but are not required to pass.