The short version
This sheet is for the last pass before the exam. It covers the two domains of the skills outline dated April 15, 2026, with term pairs people mix up and rules of thumb for scenario questions. It does not cover everything: Microsoft says related topics may appear, so use it for review, not as your only study source.
Exam facts at a glance
| Item | Fact |
|---|---|
| Exam | AI-901: Microsoft Azure AI Fundamentals |
| Passing score | 700 on a 1-1,000 scale (scaled, not 70%) |
| Time | 45 minutes listed for fundamentals exams; confirm when booking |
| Questions | Not published for this exam |
| Domain 1 | Identify AI concepts and capabilities, 40-45% |
| Domain 2 | Implement AI solutions by using Microsoft Foundry, 55-60% |
| Retake | 24 hours, then 14 days; five attempts in 12 months |
| Expiry | None |
| Replaced | AI-900, retired June 30, 2026 |
Domain 1: concepts and capabilities
Responsible AI: six principles
| Principle | Think of it as |
|---|---|
| Fairness | Similar people get similar treatment |
| Reliability and safety | It works as intended and fails safely |
| Privacy and security | Data is protected and used properly |
| Inclusiveness | It works for people with different abilities and backgrounds |
| Transparency | People can understand what it does and its limits |
| Accountability | Someone is answerable and governance exists |
Decision rules:
- A group is left out or cannot use the system: inclusiveness.
- A group gets worse outcomes for the same input: fairness.
- “Users should know they are talking to AI” or “explain why”: transparency.
- “Who is responsible when it goes wrong”: accountability.
How generative models work
- Token: a chunk of text the model reads and writes.
- Embedding: a list of numbers that captures meaning, so similar things sit close together.
- Temperature: lower is more predictable, higher is more varied.
- Token limit: caps how long the output can be. It does not change how creative it is.
- Grounding: giving the model your own data at query time. Fine-tuning: changing the model itself with training examples.
Decision rules:
- Content changes often: ground the model, do not fine-tune.
- Cheap, fast, simple task: a smaller model is a reasonable answer.
- “More consistent answers”: lower temperature.
AI workloads: pairs people mix up
| This | Versus this |
|---|---|
| Key phrases: main points, untyped | Entities: typed items such as person, place, organization |
| Sentiment: positive, negative, neutral feeling | Summarization: a shorter version of the content |
| Extractive summary: reuses existing sentences | Abstractive summary: writes new sentences |
| Speech recognition: audio to text | Speech synthesis: text to audio |
| Analyze an image: describe or read it | Generate an image: create a new one |
| Extract: pull out what is in the source | Generate: infer or write something new |
Domain 2: Foundry
Prompts, models, agents
- System prompt: persona, rules, and tone that apply to the whole conversation. User prompt: the request of the moment.
- Agent: a model plus instructions plus tools. If a question asks what to change to alter behavior, decide which of the three it is.
- Chat client: your code sends messages to a deployed model through the SDK and prints the reply. Know the parts: endpoint, credential, deployment name, messages.
Speech and text
- Stored recording and no hurry: batch-style transcription may fit.
- Live audio and instant results: real-time transcription.
- Spoken prompt to a multimodal model: the audio is sent to the model as input.
Vision
- Sending an image to a multimodal model means passing a link or encoded data, not a path on your own computer.
- Creating a new picture is a generation task, not an analysis task.
Information extraction
- Azure Content Understanding in Foundry Tools handles documents and forms, images, audio, and video.
- Values copied from the source are extractions. Values the model works out, such as a summary, are generated.
Common traps
- Picking the biggest model because it sounds powerful.
- Reading AI-900 material as if it were current: machine learning basics are not a separate domain now.
- Answering with a principle that sounds nice instead of the one the scenario describes.
- Leaving questions blank. Microsoft says wrong answers carry no penalty.
Last-day checklist
- Re-read the six principles and one example for each.
- Say aloud the term pairs above without looking.
- Skim code from one chat client and one agent and name every part.
- Check your booking, ID, and test environment (see the exam day guide).
- Sleep. Late cramming of new topics rarely pays off.
More guides for AI-901
- AI-901 Exam Format, Questions and Scoring
- AI-901 Study Plan: 4-Week and 8-Week
- Is AI-901 Hard? An Honest Difficulty Check
- AI-901 Prerequisites: What to Know First
- AI-901 Renewal and Retakes Explained
- Is AI-901 Worth It? An Honest Look
- AZ-900 vs AI-901: Which First?
- AB-900 vs AI-901: Which Exam First?
- AWS AIF-C01 vs AI-901: Which First?
Next step
For the full outline behind this sheet, open domains explained. When you are ready to check yourself, try the sample questions, then our AI-901 practice tests.