DestrezaAI ← Azure AI Fundamentals Study Kit
AI-901 · Study guide · Checked November 2026

How to pass AI-901, step by step

The exam is full of look-alikes: two AI ideas or Azure options that sound almost the same, like RAG and fine-tuning, or fairness and inclusiveness. Each question asks which one fits, and you get 45 minutes for the whole exam. This guide shows how the exam thinks, what changed from AI-900, and how to split your days. Three interactive parts, zero walls of text, and nobody asks you to train a neural network by hand.

9 minute read3 interactive partsFacts checked against Microsoft
01

The AI-901 exam, on one ticket

Before any strategy, the facts, exactly as Microsoft publishes them. AI-901 is a beginner level exam, but it is not AI-900 with a new number: Microsoft writes it for someone at the start of a career in AI solution development, with basic Python and some familiarity with Azure resources.

Exam
AI-901
Microsoft Azure AI Fundamentals, replaced AI-900
Questions
40-60
Typical for Microsoft exams; no exact count published
Time
45 min
Plan 65 minutes in the seat, with instructions
Exam fee
By country
Microsoft sets it per country or region
Valid for
No expiry
Fundamentals certifications do not expire
Where
Anywhere
Pearson VUE test center or online; Certiport for students
700 / 1000
Scaled pass mark

Never leave a blank

A wrong answer simply earns nothing: no points are deducted. A guess can only help you.

Multi-part questions pay per part

On most multi-part questions you earn one point per correct part. Three statements right out of four still counts, so never skip a whole set.

Some questions do not count

Microsoft tests new questions on real candidates and does not tell you which ones. Treat every question as real, and do not panic at a strange one.

Two details people miss. The score runs from 1 to 1,000 and is scaled, so 700 is not the same as “70% correct”. And if you already passed AI-900, you keep Azure AI Fundamentals: it is the same certification, and AI-901 is now the only exam that leads to it.

02

Where the points live, and your plan

Microsoft publishes the weight of each section as a share of the scored content. Here is how the points split:

SECTION 1Identify AI concepts and capabilities
40-45%
SECTION 2Implement AI solutions by using Microsoft Foundry
55-60%

Read it like this. Section 2 alone is 55 to 60% of the exam, and its verbs are Create, Deploy and Build, not Describe. Section 1 gives you the vocabulary every Foundry question assumes, so start there, then spend most of your days on section 2.

How many days do you have?

Pick a number: the days are split in proportion to the points, and each one comes with its three tools, in this order. It assumes about an hour a day.

1. Understand20 min with the Concept Map of the day’s section: see how its terms connect and which ones look alike.
2. Remember15 min with the Flashcards of the same section, so the terms stick.
3. Prove it25 min of practice questions, to check you can use what you learned.
    03

    Spot the impostor

    This is the real skill the exam tests: telling two look-alikes apart under time pressure. Six pairs below. Each clue describes one of the two. Pick the right one before you see the rule, because on exam day nobody hands you the rule.

    1. A loan model approves similar applicants at different rates depending on their neighborhood.

      Answer: Fairness. Similar people get different results, fairness. Some people cannot use the system at all, inclusiveness.

    2. An assistant must answer from the company's price list, which changes every week.

      Answer: RAG. Private or changing facts, RAG. A consistent style or format, fine-tuning.

    3. A bank's prompts must be processed inside the EU, at pay-per-token prices.

      Answer: Data Zone Standard. No residency need, Global Standard. Processing must stay in the EU or US, Data Zone Standard.

    4. A news site wants every person, company and city named in each article, with its type.

      Answer: Entity detection. Main subjects or tags, keyword extraction. People, places and dates with a type, entity detection.

    5. A shop wants a photo of a shelf turned into a short answer: is the cereal box upside down?

      Answer: Multimodal model. Describe or answer about a picture, multimodal model. Make or edit a picture, image generation.

    6. An HR agent must book a meeting room in the calendar when an employee asks.

      Answer: Action tool. Look up a policy or a rate, knowledge tool. Book, create or send something, action tool.

    All 10 pairs on the free Concept Map (opens in a new tab) Drill them as Flashcards (opens in a new tab)
    04

    Zoom out: the deployment types

    Some look-alikes are a whole family. The free Section 1 Concept Map puts the deployment types side by side: where your prompts are processed, how you pay, and how fast the answer comes back. Tap a type to open its card, exactly as on a Concept Map.

    Global Standard
    Pay per token; requests may be processed in any Azure region. The default choice: newest models first, lowest price, highest quota. Global Standard = no residency need. Data Zone Standard = processing must stay in a data zone. “Pay only for what you use”, “no residency requirement”, “variable traffic”.
    Data Zone Standard
    Pay per token; processing stays inside a Microsoft data zone (US, EU or Asia Pacific), with more quota than the single-geography Standard type. Data Zone = residency in a zone. Standard (regional) = residency in one Azure geography, lower quotas. Data stored at rest stays in your geography for every type. “EU only”, “data residency”, “keep processing in the US”.
    Provisioned throughput
    Reserved capacity, bought in provisioned throughput units (PTU), for steady, high-volume, latency-sensitive traffic. Comes as Global, Data Zone and Regional Provisioned. Provisioned = reserved and predictable. Standard = pay per token, best effort. “Predictable latency”, “steady high volume”, “reserved capacity”.
    Global Batch
    Large jobs sent as a file and processed asynchronously, with a 24-hour target turnaround, at a 50% discount on the Standard price. Batch = the work can wait, so it costs less. Provisioned = fast answers needed all day. “Overnight”, “no rush”, “lower cost”, “asynchronous”.
    Tokens-per-minute (TPM) quota
    The rate limit given to a pay-per-token deployment from your subscription’s quota. Going over it returns rate-limit (429) errors. Rate-limit errors with spare quota: raise the TPM. Answers cut off mid-sentence: raise max output tokens. “429”, “rate limit”, “too many requests”.
    05

    Four answers that feel right and are not

    The best wrong answers are not random. They are the reflex answer, the one your brain reaches for first. Each card below is one of the free practice questions: the reflex, what the question actually asked, and the rule to keep.

    Section 1 · Conversation history

    The forgetful chatbot

    A chat app answers the first question well, then answers the follow-up “What about in winter?” with something unrelated. The app sends only the latest message.

    Your reflexThe context window is too small. Forgetting sounds like a memory limit.
    What it askedEach call is independent, so the app must resend the earlier messages. The history was never sent at all.

    Carry this: models are stateless between calls. Conversation memory is something the app provides by resending the history.

    Section 2 · Deploy a model

    The name that is not found

    A developer deploys gpt-5-mini under the name support-chat. The code asks for gpt-5-mini and gets “deployment not found”.

    Your reflexUse the model's name. It is the obvious identifier.
    What it askedUse the deployment name, support-chat. The model name works only when the deployment kept the default name.

    Carry this: code calls a deployment by its deployment name, which can differ from the model's name.

    Section 2 · Agents

    The confident wrong limit

    An expenses agent with only a model and instructions gives a plausible taxi limit that is not the company's real policy.

    Your reflexPick a bigger model. Wrong answers sound like a model too weak.
    What it askedThe agent has no source with the policy, so it answers from training data. No model can know a private policy it was never given.

    Carry this: plausible but wrong on company facts means the agent is not grounded. Add the company's documents as knowledge.

    Section 1 · Deployment types

    The EU rule

    A European insurer must keep prompt processing in the EU, wants pay-per-token pricing and the highest default quota within that rule.

    Your reflexGlobal Standard with a resource in Europe. The resource location feels like it decides.
    What it askedData Zone Standard, using the EU data zone. A global deployment processes prompts anywhere, wherever the resource sits.

    Carry this: residency plus quota is Data Zone. Global processes anywhere, Standard stays in one geography with lower quotas, Provisioned is reserved capacity.

    Now try one for real

    On the real exam you only learn whether you passed. Here, every option is explained, the wrong ones too, because that is where the learning is. Pick an answer.

    Section 1 · Free practice question

    A support team wants a generative model to find common complaints in chat transcripts that contain customer names, emails and card numbers. What should happen before the text is sent?

    Rule to carry forward: minimize data before it reaches the model: remove personal details the task does not need.

    Try the 10 free questions (opens in a new tab) The full question bank
    06

    A guide is a map. You still need reps.

    Reading about look-alikes is not the same as spotting them. These three are free, with no sign-up, and they cover Section 1 and the first ten practice questions.

    Understand
    Free Concept Map Every Section 1 term on one page, with the 10 look-alike pairs the exam swaps and the clue words that give the answer away.
    Open the map →
    Remember
    Free Flashcards Each Section 1 term as a card that comes back just before you would forget it.
    Flip a card →
    Prove it
    10 free questions Scenario questions with an explanation for every option, including why the tempting one is wrong.
    Start practicing →
    See the full Study Kit → Back to the overview

    Practice improves your chances. Nobody can promise a pass. Section 1 is free; Sections 2 to 2 come with the full Study Kit.

    07

    What the AI-901 outline never asks you to do

    Microsoft publishes no out-of-scope list. The official skills outline gives a lot away anyway: section 1 says Describe and Identify, section 2 says Create, Deploy and Build, but always a lightweight app, a single agent, a simple client.

    Not asked by the outline

    • Training or fine-tuning a model from scratch, or the math behind it
    • Designing a production architecture for scale, cost and networking
    • Writing long programs: the code you meet is a few lines of Python
    • Building multi-agent systems or custom machine learning pipelines

    Our reading of the official skills outline. Microsoft does not publish this list.

    What you need instead

    Recognize each idea and Foundry feature, pick the one a scenario needs, and read a short Python client without panic.

    Lightweight is not a soft word. You will never ship a production app, but you must know which client, which deployment name and which tool makes the few lines work.

    08

    Five small rules for 45 minutes

    1. Do the clock math.45 minutes for 40 to 60 questions is under a minute each. Answer what you know on the first pass, mark the rest for review, and come back with the time you saved.
    2. Never leave a blank.No points are deducted for a wrong answer. In the last minute, fill every gap.
    3. Read the code, do not run it.When a question shows Python, look for the one line that matters: the client, the deployment name, the tool. The rest is scenery.
    4. Judge each statement on its own.Multi-part questions usually pay one point per part. A set of statements is not all or nothing, so do not let one hard line make you rush the others.
    5. Rehearse in the sandbox.Microsoft’s free exam sandbox shows the real screens and question types. Ten minutes there means no surprises on the day.
    Start with the free questions →
    Sources. Exam facts checked on November 10, 2026. Microsoft can change them, so confirm before you book.