Free AIF-C01 practice test: 20 questions
Twenty questions spread across the five AIF-C01 domains in roughly the proportions of the real exam. This is a foundational exam: it tests whether you can match a business scenario to the right AI concept or service category.
Fundamentals of AI and ML
Question 1. A logistics company wants to estimate delivery time in minutes from historical trip data. Which type of problem is this?
- A. Regression
- B. Binary classification
- C. Clustering
- D. Dimensionality reduction
Show answer
Answer: A
Predicting a continuous numeric value is regression. Classification predicts a discrete category, and clustering groups unlabelled data.
Question 2. A model scores 99% on training data but 62% on unseen data. What is happening?
- A. Underfitting
- B. Data drift
- C. Overfitting
- D. Label leakage from production
Show answer
Answer: C
A large gap between training and unseen performance is overfitting: the model memorised the training set rather than learning generalisable patterns.
Question 3. Which scenario describes unsupervised learning?
- A. Predicting churn from labelled past churn outcomes
- B. Grouping customers into segments discovered from behaviour alone
- C. Classifying emails using a labelled spam corpus
- D. Forecasting revenue from labelled historical revenue
Show answer
Answer: B
Grouping customers without predefined labels is clustering, which is unsupervised. The other options all train against known labels.
Question 4. A team must hold back part of the data to measure how a model performs on examples it never saw. What is this set called?
- A. The training set
- B. The feature set
- C. The label set
- D. The test or holdout set
Show answer
Answer: D
The test or holdout set measures generalisation. Training data fits the model and validation data tunes it, so neither gives an unbiased final estimate.
Fundamentals of GenAI
Question 5. What is a foundation model?
- A. A rules engine encoding business logic
- B. A large model trained on broad data and adapted to many tasks
- C. A model trained only on one customer’s data
- D. A database of question and answer pairs
Show answer
Answer: B
A foundation model is trained on broad data at scale and adapted to many downstream tasks. Being narrow, rule-based or customer-specific describes other approaches.
Question 6. A generative model invents a product feature that does not exist. What is this called?
- A. Data drift
- B. Bias
- C. Hallucination
- D. Throttling
Show answer
Answer: C
Confident but unsupported output is a hallucination. Drift and throttling describe operational conditions, and bias describes systematic unfairness.
Question 7. A team wants a model to answer from internal documents it was never trained on, without changing the model. Which approach fits?
- A. Retrieval-augmented generation
- B. Continued pre-training
- C. Fine-tuning on the documents
- D. Increasing the temperature
Show answer
Answer: A
Retrieval-augmented generation supplies relevant documents at inference time. Fine-tuning and continued pre-training both modify the model, which the requirement excludes.
Question 8. What do embeddings represent?
- A. Compressed copies of the original text
- B. Encrypted versions of the input
- C. A file format for model weights
- D. Numeric vectors capturing semantic meaning, enabling similarity search
Show answer
Answer: D
Embeddings are numeric vectors capturing semantic meaning, which is what makes similarity search possible. They are not compressed text, encryption or a file format.
Question 9. Which prompt technique gives the model a handful of worked examples inside the request?
- A. Zero-shot prompting
- B. Few-shot prompting
- C. Fine-tuning
- D. Continued pre-training
Show answer
Answer: B
Few-shot prompting includes example input-output pairs in the prompt. Zero-shot gives none, and fine-tuning changes weights rather than the prompt.
Question 10. A company needs lower per-request cost and latency for a simple classification task currently handled by a very large model. What is the reasonable first step?
- A. Evaluate a smaller model sized to the task
- B. Increase the maximum token limit
- C. Raise the temperature
- D. Send more examples in every prompt
Show answer
Answer: A
Evaluating a smaller model for a narrow task often preserves accuracy at much lower cost and latency. Raising token limits or temperature does not reduce cost.
Applications of Foundation Models
Question 11. A support assistant must answer from a product manual that changes monthly. Which design minimises ongoing effort?
- A. Fine-tune monthly
- B. Paste the manual into every prompt
- C. Index the manual and retrieve relevant sections per query
- D. Summarise the manual once a year
Show answer
Answer: C
Indexing the manual and retrieving at query time keeps answers current with no retraining cycle. Monthly fine-tuning creates recurring work.
Question 12. Retrieval returns topically related but unhelpful passages. Chunks are very large. What should you try first?
- A. Increase temperature
- B. Reduce chunk size and add overlap
- C. Return 100 chunks per query
- D. Remove the system prompt
Show answer
Answer: B
Oversized chunks dilute embeddings so retrieval matches broad topic rather than specific content. Smaller overlapping chunks sharpen retrieval.
Question 13. Which metric tells you whether a generated answer is actually supported by the retrieved sources?
- A. Fluency
- B. Latency
- C. Coherence
- D. Groundedness
Show answer
Answer: D
Groundedness measures support by source content. Fluency and coherence assess language quality, and latency measures speed.
Question 14. An agent must call an internal API to place an order. What is the safest pattern for a first release?
- A. The agent proposes the call and a human confirms before it executes
- B. The agent calls the API directly with full permissions
- C. The agent emails the order to a shared mailbox
- D. The agent writes the order into the chat transcript
Show answer
Answer: A
Requiring confirmation before the write executes keeps a human in the loop while the agent’s reliability is unproven. Unrestricted write access is the riskiest option.
Question 15. A team needs the model to adopt a specialised internal writing style consistently across thousands of documents, and prompting has not achieved it. What should they consider?
- A. Raise the temperature
- B. Shorten the prompts
- C. Fine-tune on examples of the required style
- D. Switch to a smaller model
Show answer
Answer: C
Fine-tuning on examples of the desired style teaches it durably when prompting has failed at scale. Raising temperature increases variation, the opposite of consistency.
Question 16. What is the main risk of including retrieved third-party web content directly in a prompt?
- A. It increases latency slightly
- B. It can carry prompt injection instructions the model may obey
- C. It changes the model’s licence terms
- D. It disables the content filter
Show answer
Answer: B
Untrusted content can carry prompt injection instructions that the model may follow. Token cost and formatting are real but secondary concerns.
Guidelines for Responsible AI
Question 17. A hiring model trained on historical decisions disadvantages one group. Which responsible AI dimension is most directly implicated?
- A. Fairness and bias
- B. Robustness
- C. Cost efficiency
- D. Latency
Show answer
Answer: A
Fairness concerns systematically different treatment across groups, which historical bias reproduces. Explainability and robustness are separate dimensions.
Question 18. Which practice best supports transparency about an AI system's limitations for the people using it?
- A. Not disclosing that AI is involved
- B. Publishing the model weights only
- C. Reporting accuracy on the training set
- D. Documenting intended use, known limitations and evaluation results
Show answer
Answer: D
Documenting intended use, limitations and evaluation results communicates limits to users and reviewers. Hiding that the system is AI works against transparency.
Security, Compliance, and Governance for AI Solutions
Question 19. Prompts sent to a hosted model may contain customer personal data. Which control addresses this most directly?
- A. TLS between your service and the model endpoint
- B. Redact or tokenise personal data before sending the prompt
- C. Increase the model’s context window
- D. Log every prompt in full for audit
Show answer
Answer: B
Redacting or tokenising personal data before it leaves your boundary prevents it reaching the model at all. Encryption in transit does not stop the data being processed.
Question 20. Auditors ask which data a model was trained on and who approved its deployment. What should the organisation maintain?
- A. Endpoint latency metrics
- B. Monthly cost reports
- C. Model documentation covering data lineage, evaluation and deployment approval
- D. A list of users who called the endpoint
Show answer
Answer: C
Model documentation and an approval record capture data lineage and sign-off. Endpoint metrics and cost reports do not answer either question.
How did you do?
Sixteen or more correct is a good position for AIF-C01. Below fourteen, the domain guides in this section will close the gap faster than another practice run — this exam rewards recognising concepts, not memorising service names.