AI-901 exam tips: how to gain marks on the day
These are tips for the exam itself rather than for studying — how to read AI-901’s questions, which distinctions decide marks, and how to handle the sitting. If you are still planning your preparation, the study plan is the better starting point.
Read for the constraint
Almost every implementation question is a scenario with a constraint buried in it: cost matters more than nuance, it must apply to every conversation, it must work on scanned pages, the output must be structured fields. Find that clause first, then eliminate every option that violates it. Two options usually disappear immediately.
Watch for “least effort”, “simplest” and “most appropriate”. When those appear, a technically workable but heavier answer is wrong on purpose.
The distinctions that decide marks
A handful of pairs account for a disproportionate share of the questions. Know each cold:
| If the question is about… | The answer hinges on |
|---|---|
| Reading an image vs making one | Multimodal model vs image-generation model |
| Always/never behaviour vs one message | System prompt vs user prompt |
| Reasoning over audio vs converting it | Multimodal model vs Azure Speech |
| Shorter prose vs named fields | Summarisation vs extraction |
| Tone vs things mentioned | Sentiment analysis vs entity detection |
| Unequal outcomes vs unexplained ones | Fairness vs transparency |
If you learn nothing else the night before, learn this table.
Responsible AI questions: find the primary harm
These questions nearly always describe a situation where several principles could apply, then ask which is primarily or most directly concerned. That qualifier is the question. Ask yourself what the described harm actually is:
- A group gets worse outcomes → fairness
- Nobody can explain the decision → transparency
- No owner, no audit trail → accountability
- Some users cannot access it at all → inclusiveness
- Data exposed or over-shared → privacy and security
- It behaves erratically or fails unsafely → reliability and safety
Handle the yes/no series carefully
Several statements share one scenario, and in most Microsoft exam formats you cannot return to them once answered. Read the scenario twice before answering the first statement. The statements are often near-identical with one word changed, and the differences are the point.
Time and flagging
The exam length is not published, but the general advice holds: do not stall. Answer, flag, move on. There is no negative marking, so an unanswered question is strictly worse than a guess. On a two-domain exam you will usually finish with time to spare, and a second pass over flagged items is worth more than a slow first pass.
Do not overthink the code
Code snippets on AI-901 are short and the answer is usually structural rather than syntactic — which call, which order, which parameter. You are not being asked to spot a missing colon. If you find yourself parsing character by character, step back and ask what the snippet is trying to do.
Preview features exist
Microsoft states that questions cover generally available features, and that commonly used preview features may also appear. Foundry moves quickly. If an option describes something you recognise as new but real, do not discard it purely for being new.
The practical bits
- Online proctored: clear the desk, run the system check the day before on the same machine and network, and handle drinks and breaks beforehand — you cannot leave the frame.
- Test centre: arrive early, bring the ID the booking specifies, and expect lockers for everything else.
- Either way: the exam sandbox lets you try the interface in advance for free. Doing that once removes a surprisingly large amount of exam-day friction.
After you submit
You get a preliminary result immediately and a score report broken down by the two skill areas. With only two bars, a weakness is obvious at a glance. If you did not pass, you can retake after 24 hours — but read the report first and fix the short bar rather than simply resitting.