AI-103 study resources

The official Microsoft Learn learning path for AI-103 is the only resource guaranteed to track the current skills outline, and it is free. Everything else is supplementary. The single highest-value activity, though, is not reading at all — it is building one retrieval-augmented generation application end to end.

Start with the official material

The skills outline on the exam page. Read it before anything else, and again the week you book. It lists the exact skills measured and their weights, and it is the document the exam is written from. Course material lags it by months.

Microsoft Learn. The learning path for this exam is free, modular and updated by the team that owns the services. Work through the generative AI and agent modules carefully — they carry 30–35% of the exam — and treat the vision, language and information extraction modules as lighter passes at 10–15% each.

The service documentation. For AI-103 specifically, the Azure AI Foundry documentation matters more than usual, because the exam assumes you know how the pieces fit rather than just what they are called.

The thing that actually moves the needle

Build a RAG application. Index a real document set, chunk it, retrieve, generate, and add citations.

Do this and a third of the exam becomes obvious. You will have made the chunking mistake the exam asks about, seen the difference between groundedness and relevance in your own output, and hit a quota limit at an inconvenient moment. Candidates who have done this consistently outperform candidates who have read twice as much.

An Azure free account plus a small amount of pay-as-you-go credit is enough. You do not need enterprise infrastructure.

Worth paying for

Practice tests, if they are weighted like the real exam and explain every option. The value is diagnostic: they tell you which domain to revisit. A practice test that only tells you the right answer is worth much less than one that explains why the other three are wrong.

The free 20-question practice test here is a reasonable first check before you spend anything. Under fourteen correct and more study will serve you better than more tests.

Instructor-led training, if your employer pays and you learn better with structure. Out of your own pocket it is hard to justify against free official material plus hands-on time.

What to skip

Braindumps. Sites offering “real exam questions” violate Microsoft’s terms, and using them can get your certification revoked. They also teach recognition rather than understanding, which is precisely what scenario-based questions defeat.

Courses written for AI-102. The exam changed substantially when AI-103 replaced it in June 2026 — much more weight on agents, RAG and production concerns. Material that predates that shift will send you to the wrong places. Check the publication date before you start.

Memorising service limits. The exam tests judgement about which service fits a requirement, not whether you can recite quotas.

A sensible order

1. Read the skills outline. Note the weights.
2. Work through Microsoft Learn’s generative AI and agent modules.
3. Build the RAG application. Break it. Fix it.
4. Cover the vision, language and extraction modules lightly.
5. Take the practice test. Go back to whichever domain scored worst.
6. Re-read the skills outline the week you book.

Before you book

Check the exam changes tracker. This exam already replaced AI-102 once, and objective weights move without the exam code changing.

The four-week study plan in this section turns the above into a schedule if you prefer one.