Frequently Asked Questions
What is Microsoft Foundry?
Microsoft Foundry is the unified platform for building AI applications on Azure. It brings together model deployment, agent development, retrieval pipelines, monitoring, and responsible AI tools into a single environment. Think of it as the layer that unifies what were previously separate Azure AI services (Azure OpenAI, Azure AI Search, Azure AI Vision, etc.) into a cohesive development platform with shared projects, SDKs, and governance.
How is AI-103 different from AI-102?
AI-102 (Azure AI Engineer Associate) focused on traditional Azure AI services: Computer Vision, Language, Speech, Document Intelligence, and Azure OpenAI as separate services. AI-103 is fundamentally different:
- Foundry-centric: Everything is accessed through Microsoft Foundry, not separate service endpoints.
- Generative AI and agents: 30-35% of the exam covers generative applications, RAG, agents, and multi-agent orchestration. AI-102 had minimal generative AI content.
- Agent-focused: Building, monitoring, and governing agents is a core skill. AI-102 had no agent content.
- Image/video generation: AI-103 tests image and video generation from prompts, not just analysis.
- Content Understanding: A new Foundry Tool tested across vision and document domains.
- Responsible AI for generative systems: Safety filters, guardrails, and agent governance are integrated throughout, not limited to traditional content moderation.
AI-102 officially retired on June 30, 2026. If you are coming from AI-102 experience, expect approximately 40-50% new content focused on Foundry, generative AI, and agents.
Do I need Python experience?
Yes. This is a developer exam. The audience profile explicitly states you should have "experience developing apps by using Python." You will use the Foundry SDK, implement RAG pipelines, build agents with function-calling, and integrate retrieval workflows, all in Python. If you are not comfortable writing Python, the implementation questions will be difficult.
How many practice questions should I aim for?
At least 600 unique questions across all domains. Focus on questions that explain why each answer is right and why the others are wrong. For AI-103, the explanations matter more than the questions because the exam tests integration patterns, not isolated service features.
When should I book the exam?
Once you are consistently hitting 80% or above on Legend mode across two or more sessions, and you feel confident implementing Foundry-based solutions end to end, you are ready. Trust your preparation.
How long is the exam?
120 minutes for 40-60 questions. That averages to roughly 2-3 minutes per question. Some questions may take longer if they involve reading code snippets or scenario descriptions. Budget your time for the full 120 minutes.
What is the passing score?
700 out of 1000. Microsoft uses scaled scoring across all certification exams. The raw number of correct answers needed varies slightly between exam versions. Do not try to calculate how many questions you can afford to miss.
I am a non-native English speaker. Can I get extra time?
Yes. If the exam is not available in your preferred language, you can request an additional 30 minutes to complete the exam. Request this accommodation through the exam scheduling process. If the exam is available in your preferred language, you can still take it in English with the extra time if English is not your primary language.
Can I use reference materials during the exam?
No. The AI-103 exam is closed-book. You cannot access documentation, websites, or any external resources during the test. All questions are designed to be answerable from your knowledge of Microsoft Foundry, Azure AI services, and generative AI patterns.
Does the exam cover Preview features?
Possibly. Microsoft states that "most questions cover features that are general availability (GA). The exam may contain questions on Preview features if those features are commonly used." Focus on GA features but be aware of widely adopted Preview capabilities.
What is the most important domain to study?
Domain 2 (Implement generative AI and agentic solutions, 30-35%) is the largest and most heavily weighted domain. But Domain 1 (Plan and manage, 25-30%) is the second largest and provides the governance and infrastructure foundation that Domain 2 questions assume. Together they represent 55-65% of the exam. If you are strong in both, you have a significant margin even if you lose marks on the specialized domains.
What are the most common ways people fail?
- Treating AI-103 like a harder AI-102. If you study for AI-103 by reviewing the same Azure AI services you used for AI-102, you will miss the Foundry, generative AI, and agent content that makes up 30-35% of the exam.
- Not understanding RAG at implementation depth. The exam tests whether you can implement a RAG pipeline, not just explain what RAG is. You need to know about embedding models, vector search configuration, chunking strategies, and how retrieval connects to generation.
- Ignoring agent governance. Agents are tested heavily, and so is their governance: oversight modes, tool-access controls, approval workflows, and monitoring. Building an agent without governance is not enough.
- Confusing Foundry Tools with standalone Azure services. Azure Translator, Azure Content Understanding, and Azure AI Speech are now accessed as Foundry Tools. The exam tests this integration, not just the standalone service capabilities.
- Skipping responsible AI for generative systems. Responsible AI is not just content moderation. It includes safety evaluations, trace logging, provenance metadata, and agent behavior governance. These appear in questions across Domains 1, 2, and 3.
- Not practicing with Python. Reading about the Foundry SDK is not the same as using it. If you have not imported the SDK, configured a client, deployed a model, or built an agent in code, you will struggle with implementation questions.
What are the three biggest preparation mistakes?
- Studying services in isolation. AI-103 tests how services work together in Foundry: a model deployment connected to a retrieval pipeline connected to an agent connected to monitoring. If you study each service separately, you will not be prepared for integration scenarios.
- Not building agents. Agents are the defining feature of AI-103. If you have not built at least one agent with function-calling, conversation memory, and tool integration, you are unprepared for 30-35% of the exam.
- Treating responsible AI as a separate topic. Responsible AI is integrated into every domain. Safety filters, guardrails, content moderation, and agent governance appear throughout the exam, not in a single section. Study responsible AI as a thread that runs through every solution you design.
Is there a renewal option?
Yes. Microsoft associate, expert, and specialty certifications expire annually. You can renew by passing a free online assessment on Microsoft Learn, typically 25-30 questions, available starting 6 months before your expiration date. No exam fee is required for renewal. If you don't pass on the first attempt, you can retake after 24 hours; a second and third retake each require a 14-day wait. Retakes are capped at 5 attempts per rolling 12-month period, and each attempt requires a new (free) voucher. Check the Microsoft Learn certification renewal page for current details.
Quick-Reference: What Is on the Exam (Skills measured as of April 16, 2026)
For the full, always-current list, visit the official AI-103 exam page.
| Domain |
Key Skills |
| Plan and manage an Azure AI solution (25-30%) |
Model selection (LLMs, small language models, multimodal models, Foundry Tools), Foundry service selection for generative tasks, grounding, vector search, agent workflows, multimodal processing, retrieval and indexing methods, agent memory and knowledge integration, Azure infrastructure design, deployment options, CI/CD integration, quotas and cost management, model performance monitoring, drift detection, safety monitoring, managed identity, private networking, keyless credentials, RBAC, safety filters, guardrails, content moderation, responsible AI evaluators, trace logging, provenance metadata, approval workflows, agent governance (oversight modes, constraints, tool-access controls) |
| Implement generative AI and agentic solutions (30-35%) |
Deploy and consume LLMs, small models, code models, multimodal models, RAG implementation, retrieval pipeline design, tool-augmented flows, multistep reasoning, model evaluation (fabrications, relevance, quality, safety), Foundry SDKs and connectors, Foundry project configuration, agent roles and goals, conversation tracking, tool schemas, function-calling, agent memory, agent tool integration (APIs, knowledge stores, search, Content Understanding), multi-agent orchestration, autonomous workflows with safeguards, agent monitoring and error analysis, prompt engineering, model parameter tuning, model reflection, chain-of-thought, self-critique loops, observability (tracing, token analytics, safety signals, latency), multi-model orchestration, hybrid LLM and rules engines |
| Implement computer vision solutions (10-15%) |
Image generation from text prompts and reference media, video generation from text prompts, image editing workflows (inpainting, mask-based edits, prompt-driven modifications), video editing workflows, generation and editing controls, multimodal understanding (visual context analysis, image captions, visual QA), alt-text and accessibility descriptions, Content Understanding for visual characteristics, video analysis workflows, single-task and pro-mode Content Understanding pipelines, object and region identification, unsafe visual content classification, indirect prompt injection detection in images, visual policy enforcement (watermarks, brand usage, prohibited symbols) |
| Implement text analysis solutions (10-15%) |
Entity extraction, topic extraction, summarization, structured JSON output using generative prompting and Foundry Tools, sentiment and tone detection, safety and sensitive content detection, text translation (Azure Translator, LLM-powered translation), domain-specific language model customization, speech-to-text and text-to-speech for agentic interactions, speech as agent modality, custom speech models, multimodal reasoning from audio inputs, speech translation |
| Implement information extraction solutions (10-15%) |
Content ingestion and indexing (documents, images, audio, video), semantic search, hybrid search, vector search for grounding, enrichment using custom or built-in skills, RAG ingestion flow with OCR, retrieval pipeline integration with workflows and agent tools, multimodal document extraction (OCR, layout analysis, field extraction), Content Understanding for clean grounded representations, analyzers for structured or markdown outputs for downstream reasoning |