Free AI-901 sample questions with answers
Try these before you look at the answers. Each maps to an official AI-901 skill area, and the mix reflects the real weighting: two on concepts, three on implementation.
Question 1. A bank uses an AI model to screen loan applications. Testing shows it approves applicants from one postal district at a noticeably lower rate than equally qualified applicants elsewhere. Which responsible AI principle does this primarily concern?
- A. Fairness
- B. Reliability and safety
- C. Transparency
- D. Accountability
Show answer
Answer: A
Fairness is about a system treating comparable people comparably and not disadvantaging a group. Reliability and safety concerns consistent, safe behaviour; transparency concerns people understanding how a decision was made; accountability concerns who answers for the system. All matter here, but the unequal outcome itself is a fairness problem.
Question 2. A support team wants to know, for each incoming ticket, whether the customer sounds angry, which product they mention, and a one-line version of the complaint. Which combination of text analysis techniques is needed?
- A. Keyword extraction only
- B. Translation and sentiment analysis
- C. Sentiment analysis, entity detection and summarisation
- D. Speech recognition and keyword extraction
Show answer
Answer: C
Three separate techniques map to the three requirements: sentiment analysis for tone, entity detection for the product named, and summarisation for the one-line version. Keyword extraction alone would not give tone or a summary, and translation is not part of the requirement.
Question 3. In the Foundry portal you want a deployed model to always answer in formal British English and refuse questions outside your company's product range. Where do you put those instructions?
- A. In each user prompt, repeated every time
- B. In the system prompt for the deployment
- C. In the model name when you deploy it
- D. In the deployment’s capacity settings
Show answer
Answer: B
The system prompt sets the model’s standing behaviour and constraints for every turn of the conversation. A user prompt applies only to one message, the model name does not carry behaviour, and the deployment settings control capacity and parameters rather than instructions.
Question 4. You have thousands of scanned supplier invoices as PDFs. You need the invoice number, date and total as structured fields an application can read. Which Foundry capability fits?
- A. An image-generation model
- B. Speech to text
- C. A chat completion over the raw text
- D. Azure Content Understanding in Foundry Tools
Show answer
Answer: D
Content Understanding is built to extract structured information from documents and forms, combining reading the page with returning named fields. Image generation creates pictures, speech to text handles audio, and a plain chat completion over raw text would not reliably produce consistent structured fields from scanned layouts.
Question 5. A prototype assistant must answer questions about photographs that users upload, such as whether a component looks damaged. Which approach is the simplest that meets the requirement?
- A. Pass the image into a prompt for a deployed multimodal model
- B. Use an image-generation model to recreate the photograph
- C. Convert the photograph to speech and analyse the audio
- D. Ask users to type a description instead of uploading a photo
Show answer
Answer: A
A deployed multimodal model accepts the image directly in the prompt and reasons about it, which is exactly the requirement. Generating an image solves the opposite problem, speech services handle audio, and asking users to describe the photo in text defeats the purpose of accepting images.
How did you do?
Five questions sample a direction, not a score. Notice that three of the five are implementation questions — that is the real weighting, and it is where most people lose marks. If questions 3 to 5 felt less certain than 1 and 2, spend your next session in the Foundry portal rather than in your notes.