AI-901 study plan: pass in 4 weeks
This plan assumes about 5–7 hours a week and some exposure to Azure. If you have never written a line of Python, add a week at the front for basic syntax — the exam assumes it.
The weighting drives the plan. Concepts are 40–45% of the score, implementation 55–60%. Most people invert this, spend three weeks reading about responsible AI, and then meet a question about the Foundry SDK they have never opened.
Week 1: concepts and responsible AI
- Learn the six responsible AI principles by name: fairness, reliability and safety, privacy and security, inclusiveness, transparency, accountability. Be able to match a scenario to the right one — that is exactly how they are tested.
- Understand at a working level how generative models produce output, and what a token, a prompt and a context window are.
- Learn to pick a model by capability: large versus small, text versus multimodal.
- Learn the vocabulary of the workloads: generative and agentic AI, text analysis, speech, computer vision, information extraction.
Week 2: into the Foundry portal
This is where the exam actually lives. Create a free or pay-as-you-go Azure account and do the work.
- Deploy a model in the Foundry portal and chat with it there.
- Write system prompts and user prompts, and see how changing the system prompt changes behaviour.
- Note what the deployment options and parameters do — temperature, and the settings around capacity.
- Create and test a single agent in the portal, and give it one tool.
Week 3: the four implementation areas
Build one small thing in each. Lightweight is the point — the exam says “lightweight application” repeatedly.
- Generative: a small chat client using the Foundry SDK, plus a client that talks to your agent.
- Text and speech: a text analysis call, and something that responds to a spoken prompt with a multimodal model or Azure Speech.
- Vision: pass an image into a prompt and reason about it; generate an image.
- Information extraction: run a document, an image, and an audio or video file through Content Understanding and look at the structured output.
Week 4: practice and repair
- Take a full practice exam under real conditions, no notes.
- Sort every wrong answer into one of the two skill areas.
- Rebuild whichever one is weaker — for most people it is implementation, and the fix is portal time, not more reading.
- Take a second practice exam. Book when you are consistently above 80%.
Where the hours should go
| Week | Focus | Reading | Hands-on |
|---|---|---|---|
| 1 | Concepts, responsible AI | 4–5 h | 1 h |
| 2 | Foundry portal | 1 h | 5 h |
| 3 | The four implementation areas | 1 h | 5–6 h |
| 4 | Practice and repair | 2 h | 4 h |
Note the inversion after week 1. That is deliberate and it mirrors the weighting: the implementation domain is worth more than the concepts domain, and it is the one you cannot absorb passively.
If you only have ten days
Compress, do not skip:
- Days 1–2. Responsible AI principles and the workload vocabulary. Make a one-page sheet.
- Days 3–6. Foundry portal every day: deploy a model, prompt it, build an agent, call it from a short script. Touch vision, speech and Content Understanding at least once each.
- Days 7–8. A full practice exam, then repair the weaker area.
- Days 9–10. Second practice exam, light revision of your sheet, book it.
Ten days is enough only if you already write some Python. If you do not, the honest answer is that four weeks is the realistic number, and the extra time goes into week 2, not week 1.
Using practice exams properly
A practice exam is a diagnostic, not a rehearsal. Three rules make them worth the time:
- Sit it under real conditions. No notes, no pausing, one sitting. A score you got with a browser tab open tells you nothing.
- Review every question you got right but guessed. A lucky guess is a gap that has not surfaced yet.
- Sort mistakes by skill area, not by topic. The exam reports against two areas; your revision should too.
If time runs out, cut from week 1, not week 3. Concepts can be revised from notes in an evening; the implementation questions reward people who have actually clicked through Foundry, and no amount of reading substitutes for that.