Free AI-901 practice test: 20 questions
Twenty questions across the two AI-901 skill areas, weighted roughly as the real exam is. This is a fundamentals exam, so the questions test whether you can recognise the right capability for a scenario rather than write code.
Identify AI concepts and capabilities
Question 1. A retailer wants to predict next month's sales volume per store from three years of history. Which type of machine learning is this?
- A. Classification
- B. Regression
- C. Clustering
- D. Anomaly detection
Show answer
Answer: B
Predicting a continuous numeric value from labelled historical data is regression. Classification predicts a category, and clustering finds groups without labels.
Question 2. A bank wants to flag transactions that differ markedly from a customer's normal behaviour, without a labelled list of fraud cases. Which workload fits?
- A. Regression
- B. Classification
- C. Translation
- D. Anomaly detection
Show answer
Answer: D
Anomaly detection identifies unusual patterns without needing labelled examples of each fraud type. Classification would require labelled fraud data.
Question 3. Marketing wants to group customers into segments that were not defined in advance. Which approach applies?
- A. Regression
- B. Classification
- C. Clustering
- D. Object detection
Show answer
Answer: C
Clustering discovers groupings in unlabelled data. Classification assigns items to categories that already exist, which contradicts ’not defined in advance'.
Question 4. A recruitment model was trained on ten years of past hiring decisions and now favours one group of applicants. Which responsible AI principle is most directly at stake?
- A. Fairness
- B. Reliability and safety
- C. Privacy and security
- D. Transparency
Show answer
Answer: A
Fairness concerns systems that treat groups differently without justification, which is what historical bias in training data produces. Reliability and transparency matter but are not the principle being violated here.
Question 5. Regulators ask a lender to explain why a specific loan application was declined by its model. Which principle does this requirement reflect?
- A. Fairness
- B. Inclusiveness
- C. Privacy and security
- D. Transparency
Show answer
Answer: D
Transparency covers being able to explain how a system reached a decision. Accountability concerns who answers for outcomes, which is related but broader than explaining an individual decision.
Question 6. Which scenario is an example of a computer vision workload rather than natural language processing?
- A. Summarising a legal document
- B. Counting vehicles in traffic camera images
- C. Detecting sentiment in product reviews
- D. Translating a web page
Show answer
Answer: B
Counting vehicles in camera images is image analysis. Summarising documents, detecting sentiment and translating text all operate on language.
Question 7. A generative model produces a fluent answer containing a statistic that does not exist in any source. What is this commonly called?
- A. Overfitting
- B. Model drift
- C. Hallucination
- D. Underfitting
Show answer
Answer: C
A hallucination is a fluent but unsupported or fabricated output. Overfitting and drift describe training and deployment problems, not individual fabricated outputs.
Question 8. What best describes a large language model's context window?
- A. The amount of text the model can consider in a single request
- B. The total size of the model’s training data
- C. Long-term memory of previous conversations
- D. The maximum number of requests per minute
Show answer
Answer: A
The context window is the amount of text, measured in tokens, the model can consider in one request, covering both input and generated output. It is not storage, training data size, or a rate limit.
Question 9. A team wants a model to adopt a specific writing style using a few examples inside the request itself, with no retraining. What is this called?
- A. Fine-tuning
- B. Few-shot prompting
- C. Clustering
- D. Feature engineering
Show answer
Answer: B
Providing examples within the prompt is few-shot prompting. Fine-tuning changes model weights and requires a training run.
Implement AI solutions by using Microsoft Foundry
Question 10. You need an application to answer questions from a company handbook that changes monthly, without retraining. Which pattern applies?
- A. Fine-tuning the base model monthly
- B. Increasing the temperature
- C. Retrieval-augmented generation over an index of the handbook
- D. Training a classification model
Show answer
Answer: C
Retrieval-augmented generation retrieves current content at query time, so answers track the handbook automatically. Fine-tuning would need repeating each month.
Question 11. A chatbot must give consistent, near-identical answers to the same factual question every time. Which setting should be adjusted?
- A. Lower the temperature
- B. Raise the temperature
- C. Increase the maximum response length
- D. Disable the system prompt
Show answer
Answer: A
Lowering temperature reduces randomness in token selection, producing more consistent output. Raising it increases variety, which is the opposite of the requirement.
Question 12. What is the purpose of a system message in a chat-based generative AI application?
- A. It records the conversation for auditing
- B. It sets the model’s role, scope and behaviour for the conversation
- C. It stores the user’s question
- D. It replaces the content safety filter
Show answer
Answer: B
The system message sets the model’s role, scope and behavioural rules for the conversation. It is not the user’s question, nor a log, nor a content filter by itself.
Question 13. Your application must block sexual, violent and self-harm content in both prompts and responses. What should you configure?
- A. A lower temperature
- B. A longer system prompt asking the model to behave
- C. A smaller model
- D. Content safety filters applied to prompts and completions
Show answer
Answer: D
Content safety filtering inspects input and output against those harm categories. Prompt wording alone cannot guarantee blocking, and temperature is unrelated to safety.
Question 14. A prototype must turn recorded customer calls into searchable text. Which capability do you need first?
- A. Speech to text
- B. Text to speech
- C. Translation
- D. Image captioning
Show answer
Answer: A
Speech to text transcribes audio into text, which can then be indexed. Text to speech does the reverse, and translation changes language rather than modality.
Question 15. You must extract invoice number, date and total from scanned invoices as structured fields. Which capability is designed for this?
- A. Plain optical character recognition only
- B. Image classification
- C. Document intelligence with field extraction
- D. Sentiment analysis
Show answer
Answer: C
Document intelligence combines OCR with layout and field extraction to return structured data. Plain OCR returns unstructured text, and image classification only labels the whole image.
Question 16. Which measure best protects an AI service key used by a production web application?
- A. Storing the key in the application’s source code
- B. Using a managed identity instead of a key
- C. Embedding the key in client-side JavaScript
- D. Emailing the key to the development team
Show answer
Answer: B
A managed identity removes the key entirely, letting the platform handle authentication and rotation. Storing keys in code, config files or client-side script all expose them.
Question 17. A model endpoint's answers gradually become less accurate over a year as customer behaviour shifts. What is this called?
- A. Hallucination
- B. Overfitting
- C. Model drift
- D. Throttling
Show answer
Answer: C
Model drift describes degrading performance as real-world data diverges from the training distribution. Hallucination and overfitting describe different failure modes.
Question 18. You must compare two prompt designs on the same set of test questions and pick the better one on evidence. What should you use?
- A. An evaluation run over a fixed test set with defined metrics
- B. Asking two colleagues which they prefer
- C. Counting user complaints after release
- D. Choosing the shorter prompt
Show answer
Answer: A
A systematic evaluation run scores both prompts against the same dataset and metrics, giving a comparable result. Informal impressions and user complaints are not controlled comparisons.
Question 19. A grounded chatbot must cite which document each answer came from. What does this primarily require?
- A. A higher temperature
- B. A larger model
- C. Shorter answers
- D. Retrieval that returns source references with each passage
Show answer
Answer: D
Returning source references alongside retrieved passages lets the application display citations. Raising temperature, enlarging the model or shortening answers do not produce provenance.
Question 20. Which statement about tokens is correct for cost estimation?
- A. Only the generated output is billed
- B. Both input and generated output are billed as tokens
- C. Only the input prompt is billed
- D. Tokens are billed per request regardless of length
Show answer
Answer: B
Both the input sent to the model and the output it generates are billed as tokens. Assuming only one side counts leads to serious underestimates, especially with long prompts.
How did you do?
Sixteen or more correct is a good sign for a fundamentals exam. Below fourteen, work through the domain guides in this section before booking — AI-901 rewards recognising the right capability far more than technical depth.