Free Study Guide: Microsoft AI-103 — Developing AI Apps and Agents on Azure

Microsoft retires AI-102 (Designing and Implementing a Microsoft Azure AI Solution) on June 30, 2026, and replaces it with Exam AI-103: Developing AI Apps and Agents on Azure — the path to the new Microsoft Certified: Azure AI Apps and Agents Developer Associate credential. The rename is not cosmetic. The exam is rebuilt around Microsoft Foundry (formerly Azure AI Foundry, earlier Azure AI Studio), the Responses API agent runtime, and the way real teams ship AI in 2026.

I put together a set of dense, fact-checked study guides while working through the new blueprint, and I’m releasing them as a free, open-source resource. Here’s what’s in it, and why a certification still earns its keep in a world that re-prices every model and renames every service every quarter.

What the AI-103 Exam Covers

AI-103 is organized into five weighted domains. The guides map one-to-one to those domains, and are numbered in suggested study order — weight-first — not official exam order:

The generative and agentic domain (Guide 1) is the single largest slice at 30–35%. On AI-102, that material was split across two smaller domains — so the consolidation is a strong signal of where Microsoft wants engineers spending their time. One more trap worth flagging early: speech is not its own domain. It lives as a subsection inside the text-analysis domain (Guide 4), so don’t double-count it when you plan study time.

What’s in the Study Guide

The material is built for Azure AI engineers and developers who build, manage, and deploy agents and AI solutions with Python and Microsoft Foundry — and who want a service-selection-focused review rather than a click-through tutorial. Each guide ships with three things:

• A dense domain guide mapping every objective to the current Foundry service that satisfies it, with the decision criteria for choosing between overlapping options
• A companion slide deck (PowerPoint) for fast revision or teaching
• An audio deep-dive — a podcast-style episode you can listen to on a commute

Every factual claim is verifiable against current Microsoft Learn documentation, with key services linked inline and in a Further reading section per guide. The links target the current docs and were verified live in June 2026, against the skills outline dated April 16, 2026.

AI-103 Exam Topics & Key Study Areas

Generative AI and agentic solutions (Domain 1 — 30–35%)

The largest domain. Building generative apps on the Foundry project endpoint; the Responses API (Agents v2) runtime model — agents, conversations, and responses; agent tools, function calling, and multi-agent orchestration; plus tuning, evaluation, tracing, and observability. This is where the exam tests whether you can actually wire an agent together, not just describe one.

Plan and manage an Azure AI solution (Domain 2 — 25–30%)

Choosing models, Foundry services, and tools; designing retrieval, indexing, memory, and knowledge integration; standing up Foundry resources and deployments (Standard / Provisioned / Batch / Data Zone); CI/CD pipelines; operating, monitoring, and securing solutions with RBAC, Entra ID, and private networking; and responsible-AI governance with guardrails. This is the architect’s domain.

Information extraction solutions (Domain 3 — 10–15%)

Retrieval and grounding pipelines — pull vs. push ingestion, and full-text / vector / hybrid / semantic / agentic retrieval — plus document extraction. The high-value distinction here is Azure AI Document Intelligence (deterministic fields) versus Content Understanding (generative, schema-driven, RAG-ready output), and knowing which one a question is actually asking about.

Text analysis and speech (Domain 4 — 10–15%)

Entity, topic, and summary extraction with structured JSON outputs; sentiment, PII, and content safety; Azure Translator vs. LLM translation flows; and speech-to-text / text-to-speech, custom speech, the Realtime API, and multimodal audio reasoning.

Computer vision solutions (Domain 5 — 10–15%)

Image generation (gpt-image-*, not the retired DALL-E 3) and video generation (Sora 2); vision-enabled chat models; captions and accessibility alt text; Content Understanding analyzers; and multimodal safety, provenance (Content Credentials / C2PA), and moderation.

High-Leverage Facts That Are Easy to Get Wrong

Microsoft renames and re-scopes these services fast, and stale answers are how strong candidates lose easy points. A few that the guides hammer:

• The platform is Microsoft Foundry. Don’t reintroduce the old Azure AI Foundry / Azure AI Studio names except when contrasting legacy versus current.
• The exam is AI-103, not AI-102 — and AI-102 retires June 30, 2026.
• Keyless model inference in new Foundry uses the Foundry User role, not Cognitive Services User (that was the classic Azure OpenAI answer — watch the resource type in the question).
• The runtime is the Responses API (Agents v2) — conversations / items / responses / agent versions — not the legacy Assistants API (threads / messages / runs / assistants), which is reported to sunset August 26, 2026.

Why the Certification Still Matters in a Fast-Moving World

It’s a fair question. Models ship monthly. Services get renamed mid-quarter. By the time you pass an exam, isn’t half of it stale? Here’s why I still think AI-103 is worth the time:

• It forces breadth you’d otherwise skip. Most engineers go deep on the two or three services their current project needs and never touch the rest. A blueprint marches you through retrieval design, governance, vision, speech, and deployment models you’d never volunteer to study — and that breadth is exactly what makes you useful when the next project looks nothing like the last one.
• It validates judgment, not trivia. The hardest AI-103 questions aren’t “what does this API return.” They’re “given these constraints, which service do you choose?” Document Intelligence or Content Understanding? Translator or an LLM? Standard or Provisioned deployment? That service-selection muscle is the durable skill — it survives every rename, because the trade-offs underneath don’t change as fast as the marketing names.
• The fundamentals outlive the labels. Grounding, RBAC, observability, responsible-AI guardrails, retrieval architecture — these are the same whether the portal says Foundry, Studio, or whatever comes next. Learn the shape of the platform once and you re-skin it cheaply each time Microsoft moves the furniture.
• It’s a credible signal in a noisy market. Anyone can say they “do AI” now. A current, vendor-backed associate credential tells employers and clients you can build on Azure’s AI stack as it exists today — not as it existed two years ago. In a market flooded with hype, a recent cert is a cheap, legible proof of currency.
• The clock is a feature, not a bug. Recertification keeps you honest. A fast-moving field is exactly the field where a periodic forced refresh has the most value — it drags you back to the docs before your mental model quietly goes stale.

The half-life of any single fact is short. The half-life of knowing how the platform is organized and how to choose within it is long. That’s what a good study guide — and a current certification — actually buys you.

Who This Study Guide Is For

• AI engineers building generative and agentic apps on Azure with Python and Foundry
• Solution architects designing retrieval, deployment, security, and governance for AI workloads
• Developers preparing specifically for the AI-103 exam — or migrating off the retiring AI-102
• Anyone who wants a current, no-fluff map of Microsoft Foundry and the services around it

AI-103 Exam FAQ

Is the AI-103 certification worth it?

If you build AI on Azure, yes. It validates the service-selection judgment and architectural breadth that survive renames and model churn — exactly the skills a fast-moving field rewards. It’s also the current, credible signal that you can ship on the platform as it exists today.

I’m certified on AI-102 — what should I do?

AI-102 retires June 30, 2026. Plan your transition to AI-103. The conceptual overlap is real, but the exam is rebuilt around Microsoft Foundry and the Responses API, so don’t assume your AI-102 knowledge maps one-to-one — the guides call out exactly where it doesn’t.

What’s the hardest part of the exam?

The generative and agentic domain at 30–35%, and the service-selection questions throughout. The exam rewards knowing which service fits a constraint, not just what each one does.

How current is the study guide?

The links and facts were verified live in June 2026 against the April 16, 2026 skills outline. Some guides open with a dated accuracy note recording when the check happened. Microsoft moves fast, so always confirm specifics against the official AI-103 study guide before exam day.

Where do I register?

Through Microsoft’s official certification pages. Start with the exam study guide and the certification overview for current pricing, format, and prerequisites. The official training course is AI-103T00-A: Develop AI apps and agents on Azure.

Get the Study Guide

The entire AI-103 study guide — all five domain guides, slide decks, and audio deep-dives — is open source on GitHub:

github.com/kkaminsk/AI-103-Study-Guide

Clone it, fork it, study from it. If you find an error or a service that’s been renamed again, PRs and issues are welcome — keeping it current is a team sport in this field.

Need help building on Azure AI? Big Hat Group designs and ships production AI solutions on Microsoft Foundry — agentic workflows, retrieval pipelines, and governed enterprise AI platforms. Explore our consulting services or get in touch.