Step one: your exam date
This is the only setting on the site, and it drives everything else — the countdown, how many study days you get, what lands on each of them, and when each practice exam is scheduled.
AI-901 replaced AI-900 on June 30, 2026, and it is not a rename. The old exam had five skill areas and asked you to describe things. The new one has two, and more than half of it asks whether you can actually build something on Microsoft Foundry.
Two consequences. First, Microsoft now expects familiarity with Python syntax, REST APIs, SDKs, and CLIs — that is in the official audience profile. Second, and this is where people lose points: classic machine learning is gone. Regression, classification, clustering, confusion matrices, accuracy and recall, Azure Machine Learning studio, automated ML, compute clusters — none of it is on this exam. If you are studying from AI-900 material, you are learning a syllabus that no longer exists.
What moved, in one table
| Topic | On AI-900 | On AI-901 |
|---|---|---|
| Responsible AI | Part of skill area 1 | Still here Objective 1.1, largely unchanged |
| Classic machine learning | A whole skill area, 15–20% | Removed entirely |
| Computer vision | A whole skill area, mostly Azure AI Vision and Custom Vision | Reframed Now multimodal models and image generation |
| NLP | A whole skill area, service-by-service | Reframed Folded into text and speech on Foundry |
| Generative AI | 20–25%, conceptual | Expanded Now the backbone of the whole exam |
| Agents | Barely mentioned | New and major Building and grounding agents |
| Content Understanding | Did not exist | New A full objective of its own |
| Hands-on building | Not assessed | 55–60% of the exam |
What the exam actually looks like
The facts that shape how you study
| Exam code | AI-901 |
| Certification | Microsoft Certified: Azure AI Fundamentals |
| Replaced | AI-900, retired June 30, 2026 |
| Objectives version | April 15, 2026 |
| Questions | Typically 40–60 |
| Exam time | About 45 minutes, in a 60-minute appointment |
| Score scale | 1–1000 |
| Passing score | 700 |
| Wrong-answer penalty | None — never leave a question blank |
| Prerequisites | None formally, but Python and Azure familiarity is assumed |
What 700 really means
A 700 is not 70% of the questions in a simple sense — Microsoft scales scores and weights harder items more heavily, and it has never published the conversion. But 70% correct is the working assumption everyone uses, and it is what this site is calibrated to.
On a 50-question practice exam that is 35 correct. Miss sixteen and you are out.
Skill area 2 is up to 60% of the exam and it is the part you cannot bluff. Get into the Foundry portal early, deploy a model, and build one agent. An hour of clicking is worth a day of reading here.
The two skill areas
These are the official weightings from the April 15, 2026 skills-measured document. The third column is how this site distributes its 50-question practice exams.
| # | Skill area | Official weight | Questions here |
|---|
The seven objectives underneath them
Two skill areas is too coarse to study from, so every question, drill, and flashcard on this site is tagged to one of these instead.
| Obj | Objective | Questions here | Guide |
|---|
Objective 2.1 — generative AI apps and agents — is the single biggest block on the exam. Together with the rest of skill area 2 it is well over half your score. If you are short on time, read the Foundry guide and the Agents guide first, then Content Understanding, because it is brand new and candidates coming from AI-900 have never seen it.
How the plan is built
There is no fixed number of days. The generator takes the days between today and your exam and fits a real plan into them.
Baseline before you study
Day one is a cold practice exam. Not to feel good — to find the two objectives that are costing you the most, so the rest of the plan can aim at them.
Three exams, three repairs
Practice exams are spaced across whatever window you have, and each one is followed by a repair day. The repair day is where the score actually moves.
Hands-on is built in
Because most of the exam is implementation, the plan schedules short lab tasks in the Foundry portal alongside the reading. They are not optional.
Taking a practice exam measures your score. It does not raise it. The gain comes from the hour afterward, when you work out why each miss happened — a fact you never learned, a fact you know but read past, or a misread question. Sort every miss into one of those three buckets and the fix becomes obvious, because the three have three different fixes.
Your progress
Loading…
Everything on this site
Skill area 1 — AI concepts and capabilities
40–45% of the exam. Vocabulary, principles, and judgment.Responsible AI
The six principles and the giveaway word for each, plus content safety, prompt shields, and the four-stage process.
Models & Deployment
How generative models work, choosing a model, deployment options, and every setting you can turn.
AI Workloads
Naming the workload from a business sentence, generative versus agentic, and the text analysis techniques.
Skill area 2 — Building on Microsoft Foundry
55–60% of the exam. This is the half you have to actually do.The Foundry Platform
Projects, deployments, playgrounds, prompt engineering, and the Foundry SDK in a dozen lines of Python.
Agents new
What an agent is, prompt versus hosted, the tool catalog, grounding, and building one yourself.
Text & Speech
Azure Language and Azure Speech in Foundry Tools, and when a generative model beats a purpose-built feature.
Vision & Image Generation
Images inside prompts, generating pictures and video, and which classic vision services are on the way out.
Information Extraction new
Content Understanding: analyzers, field schemas, four modalities, and when to use Document Intelligence instead.
Practice and tools
Where the score actually moves.Your Study Plan
Day-by-day tasks with checkboxes that save automatically, sized to your exam date.
Topic Drills
Seventeen short timed sets on one skill at a time, with an explanation for every question.
Practice Exams 1–3
Three full-length 50-question exams weighted like the real thing, with scaled scoring and review mode.
Study Tools
Score estimator, exam timer, 180 flashcards, a searchable Foundry reference, and the error log.
Exam day logistics
Whichever way you are testing
- Two forms of ID, one of them government issued with a photo. The names must match your booking exactly.
- Arrive or log in thirty minutes early. Late is the same as absent.
- No notes, no phone, no watch, no second monitor.
- If you are testing online, run the system check on the machine you will actually use, days ahead.
- Online testing also means a clear desk and a clear room. The proctor will make you pan the camera.
- Eat something with protein first. Forty-five minutes is short, but a blood-sugar crash is still real.
Once the clock starts
- Read the last line first. BEST, FIRST, MOST, and NOT each change which answer is correct.
- Find the verb in the scenario. Extract, classify, transcribe, summarize, generate, and orchestrate each point somewhere different.
- Watch for “Select two.” Missing one costs the whole question.
- If a question passes about 75 seconds, answer it, mark it, and move on.
- Answer every question. There is no penalty, so a blank is strictly worse than a guess.
- Leave five minutes to sweep your marked questions. Do not rewrite answers you were confident about.