Blog Post
AI-103 Study Guide: Microsoft Foundry, Agents, and RAG Explained (2026)
AI-103 covers 5 domains, but Generative AI and Agentic Solutions alone is 30-35% of the exam. Here's exactly where to spend your study time and why.
AI-103 Study Guide: Microsoft Foundry, Agents, and RAG Explained
Most AI-103 material online falls into one of two traps: a repackaged AI-102 guide with a new exam code, or a two-paragraph overview that tells you the exam exists without telling you what's on it.
AI-103 is not AI-102 renamed. It's built around Microsoft Foundry, agentic AI, and RAG — and nearly a third of it is content that had no AI-102 equivalent at all. Here's what's actually tested, in the proportions it's actually tested.
Quick facts: 40-60 questions · 120 minutes · 700/1000 to pass · $165 USD · annual free renewal
The Domain Weights (Read This First)
| Domain | Weight |
|---|---|
| Generative AI and Agentic Solutions | 30-35% |
| Plan and Manage Azure AI Solutions | 25-30% |
| Computer Vision Solutions | 10-15% |
| Text Analysis Solutions | 10-15% |
| Information Extraction Solutions | 10-15% |
One domain is worth more than the other three combined. If your study plan gives all five equal time, you're underinvesting in the section that decides whether you pass.
👉 Follow the full 6-week AI-103 study plan — it's already weighted to match this table.
Microsoft Foundry: The Platform Everything Sits On
Skip this and every later topic feels harder than it should.
- Hub — the top-level container. Holds connections to Azure AI Search, Storage, Key Vault, and shared security/networking config.
- Project — sits inside a hub. Where the real work happens: model deployments, agents, evaluations.
- Model deployment — pick from LLMs (complex reasoning), small language models (cost/latency-sensitive), multimodal models (image/video/audio + text), or code models (code tasks). The exam tests this as a decision, not a definition.
- Connections — how a project reaches Azure AI Search (retrieval), Storage (documents), and Key Vault (secrets). Every RAG or agent question assumes you understand these are configured, not just present.
Don't treat Foundry as "the new portal." Treat it as an architecture you can draw from memory.
Related: How hard is AI-103
Agents: The Domain That Decides Your Score
An agent is not a smarter chatbot. It runs a loop: reason → act → observe → repeat. Questions describe a requirement and ask what tools and instructions the agent needs — you're designing the loop, not recognizing it.
Know these cold:
- Instructions — defines the agent's role and boundaries
- Tools — function tools, Code Interpreter, Azure AI Search grounding, Bing grounding
- Threads — how conversation memory persists across turns
- Tool schemas — the structured contract for what a tool accepts and returns
Two orchestration patterns get tested:
- Orchestrated multi-agent — one orchestrator routes to specialist agents (e.g., billing agent vs. support agent)
- Autonomous/semiautonomous — agents act with safeguards and human approval checkpoints
Governance matters as much as capability. Agents need Azure managed identities scoped with RBAC — never inherited user credentials or shared service accounts. This is scored content, not a nice-to-have.
The exam guide references "Foundry SDKs and connectors" without naming a specific orchestration framework. Study the concepts — plugins, function-calling, planners — not any one SDK's syntax.
RAG: Shows Up Everywhere, Not Just One Domain
A model's knowledge is frozen at training time. RAG retrieves current, relevant data at query time and grounds the response in it. This isn't confined to Information Extraction — it's core to the agents domain too.
Azure AI Search, the primary retrieval engine, tests three modes:
- Vector search — similarity by meaning
- Hybrid search — vector + keyword (BM25); the default for most RAG scenarios
- Semantic ranking — reranks hybrid results for top relevance, at the cost of latency
The ingestion pipeline: Document Intelligence (OCR/layout extraction) → chunking (semantic chunking usually beats fixed-size) → enrichment → index. The exam tests the whole flow from document to agent, not retrieval in isolation.
Evaluation is where candidates lose easy points. The Azure AI Evaluation SDK checks:
- Groundedness — is the answer actually supported by retrieved content?
- Relevance — did retrieval pull the right content?
- Quality — accurate and well-structured?
- Safety — harmful or biased content?
A question describing a plausible-but-wrong RAG answer is testing whether you know which of these four to check first.
Content Understanding: The Service Most Study Guides Miss
Azure AI Content Understanding uses LLM-based extraction (not traditional OCR) to turn documents, images, audio, and video into structured output. It shows up in three places:
- Computer Vision — extracting visual characteristics from images/video
- Information Extraction — producing clean, grounded data for RAG and agents
- Agents — as a tool an agent can call directly
If your prep materials don't mention this service, they're incomplete for this exam.
Plan and Manage Azure AI Solutions (25-30%)
The governance layer underneath everything else. Four areas:
- Service selection — right model, right retrieval method, right agent tooling for the scenario
- Infrastructure and deployment — Azure AI infra design, deployment options, CI/CD integration
- Monitoring and security — quotas, cost, drift detection, managed identity, private networking
- Responsible AI — safety filters, guardrails, trace logging, agent oversight controls
Questions here regularly combine two of these at once — service selection plus security, or monitoring plus responsible AI. Don't study this domain as isolated bullet points.
The Three Smaller Domains, in Brief
- Computer Vision (10-15%) — image/video generation from prompts (not just analysis), multimodal understanding, unsafe content and prompt-injection detection
- Text Analysis (10-15%) — entity/sentiment/summarization via generative prompting, plus speech (STT/TTS) framed as an agent modality, not standalone transcription
- Information Extraction (10-15%) — document extraction, RAG ingestion with OCR, Content Understanding analyzers
Individually smaller, collectively a third of the exam. Don't skip them — just don't over-invest either.
The Mistake Most Candidates Make
Treating this like a vocabulary exam. Knowing that Azure AI Search does vector search, or that agents use tools, gets you through the easy questions. It won't get you through a scenario asking you to pick between three plausible architectures under a stated constraint (cost, latency, data sensitivity, safety).
Build these before your exam, not just read about them:
- One working RAG pipeline
- One multi-agent system with tool-access governance
Fast Answers
How is this different from AI-102?
Foundry-centric instead of separate service endpoints. Generative AI and agents (30-35%) barely existed on AI-102. Image/video generation, Content Understanding, and agent governance are all new. Expect 40-50% new content if you're coming from AI-102.
Do I need Python?
Yes. This is a developer exam — you'll read and reason about Foundry SDK code throughout.
How much hands-on practice do I need?
Significant. Build a real RAG pipeline and a real multi-agent system in Foundry. Configure Content Understanding analyzers. Set up safety filters and guardrails. Scenario questions assume you have done these things.
Is Semantic Kernel on the exam?
Not named explicitly, but it's the dominant orchestration framework in this ecosystem. Study the concepts (plugins, function-calling, planners), not SDK-specific syntax.
What's the passing score?**
700/1000, scaled scoring, so don't try to back-calculate how many questions you can miss.
Can I use reference materials during the exam?**
Yes. You can use the integrated Microsoft Learn platform. Practice searching so that you do not waste too much time.
Want to pressure-test your understanding with scenario-based practice across all five domains? ExamOS covers Foundry architecture, agentic solutions, and RAG design decisions the way the real exam actually tests them — daily practice that builds judgment, not just recall.
👉 Related Exam