AWS AI Practitioner (AIF-C01): 20 Practice Questions by Domain

By Ravindra Bagale, trainer · LinkedIn · Last updated: 19 October 2026

Generative AI course · Chapter 49 · Estimated reading time: 14 minutes

1. Suppose two options both sound right

Suppose Asha, a support engineer, reads this question: "A company wants a chatbot that answers from its own policy documents and shows the source. The documents change monthly. What should it use?" Two options look good: fine-tune a model and use a knowledge base with RAG.

She remembers the chapter on fine-tuning vs RAG: changing facts and citations point to RAG. The key words were "own documents", "changes monthly" and "shows the source". Spotting such key words is most of the skill this exam tests.

2. The exam in one table (official AWS pages, 19 October 2026)

The 5 domains and their weights:

3. A 4-step way to read any scenario question

1. Find the goal: what does the company want? (answers from documents, lower cost, detect bias)
2. Underline the limits: "least effort", "lowest cost", "no coding", "must cite sources", "data changes often".
3. Cross out options that break a limit, even if they would work. Training a model from scratch "works", but never fits "least effort".
4. Pick the most managed, simplest option that meets every limit. Foundational exams usually reward the simplest correct choice.

4. Worked example 1: scenario words to AWS services

Copy this table by hand once. Writing it is part of learning it.

Drill it with AI. Paste the table into any AI chat with this prompt:

Using only the services in this table, quiz me with 5 one-line business scenarios.
Wait for my answer after each one, then tell me if I was right and which look-alike
service I might confuse it with. [paste the table]

Expected output (yours will be similar, not the same): one scenario at a time, such as "A hospital wants to pull patient names and dates from scanned admission forms", and after your answer, feedback like "Correct: Amazon Textract. Do not confuse it with Amazon Transcribe, which turns speech into text." Check any service you are unsure about on its AWS page; AI quiz feedback can also be wrong.

5. Worked example 2: 20 practice questions

Answer all 20 first, writing your letters in a line, for example BCAD.... The answers and explanations follow. These are our own practice questions in the exam's style, not real exam questions.

Domain 1: Fundamentals of AI and ML

1. A model learns from emails already marked "spam" or "not spam". What type of learning is this? A. Unsupervised learning B. Supervised learning C. Reinforcement learning D. Clustering
2. A trained model is used to predict whether a new transaction is fraud. What is this step called? A. Training B. Labelling C. Inference D. Fine-tuning
3. A shop wants to group customers into segments, with no labels in the data. Which approach fits? A. Clustering B. Classification C. Regression D. Translation
4. A company wants to predict next month's sales amount in rupees. What kind of problem is this? A. Classification B. Clustering C. Speech recognition D. Regression

Domain 2: Fundamentals of generative AI

5. Generative AI models read text, and are usually billed, in which units? A. Pixels B. Tokens C. Pages D. Sentences
6. A model gives a confident answer with a book title that does not exist. What is this called? A. Hallucination B. Overfitting C. Latency D. Drift
7. Which setting makes a model's answers more or less random? A. Maximum tokens B. Stop sequence C. Temperature D. Region
8. What is a large model trained on broad data that can be adapted to many tasks? A. A rule engine B. A decision tree C. A database D. A foundation model
9. What is the amount of text a model can consider at one time? A. A token B. The context window C. An embedding D. Latency

Domain 3: Applications of foundation models

10. A company wants answers from its own policy documents, which change monthly, with sources, and with the least effort on AWS. What fits best? A. Fine-tune a model every month B. Train a model from scratch C. Bedrock Knowledge Bases (RAG) D. Raise the temperature
11. A prompt includes 3 worked examples of the output format. What technique is this? A. Zero-shot prompting B. Fine-tuning C. RAG D. Few-shot prompting
12. A company has thousands of approved examples and needs a strict house style that prompts cannot achieve. What should it consider? A. Fine-tuning (model customisation) B. Higher temperature C. A larger context window only D. Shorter prompts
13. Which metric compares a generated summary with a reference summary? A. Accuracy B. ROUGE C. RMSE D. Latency
14. Where can a company store embeddings for semantic search on AWS? A. Amazon Polly B. AWS CloudTrail C. Amazon OpenSearch Service vector search D. Amazon Translate
15. Answers from a model are too long and unfocused. What is the lowest-effort first fix? A. Prompt engineering B. Pre-train a new model C. Fine-tune the model D. Move to another Region

Domain 4: Guidelines for responsible AI

16. A loan model approves fewer applicants from one group with similar finances. What is the main concern? A. Latency B. Bias C. Hallucination D. Cost
17. Which AWS tool helps detect bias in data and explain model predictions? A. Amazon Lex B. Amazon Polly C. Amazon Comprehend D. Amazon SageMaker Clarify
18. A team wants a standard document of each model's intended use, metrics and limits. What should it use? A. Amazon CloudWatch B. Amazon S3 alone C. Amazon SageMaker Model Cards D. Amazon Rekognition

Domain 5: Security, compliance and governance

19. An auditor wants a record of who called which Bedrock API, and when. Which service provides it? A. AWS CloudTrail B. Amazon Polly C. Amazon Translate D. Amazon Personalize
20. Under the shared responsibility model, who controls the IAM permissions for a company's Bedrock app? A. AWS B. The customer C. The model provider D. Nobody; they are fixed

Answers and explanations

1. B. Labelled examples (spam or not) mean supervised learning.
2. C. Using a trained model on new data is inference.
3. A. Grouping without labels is clustering, a kind of unsupervised learning.
4. D. Predicting a number is regression.
5. B. Models process and bill by tokens.
6. A. A confident, made-up answer is a hallucination.
7. C. Temperature controls randomness.
8. D. That is the definition of a foundation model.
9. B. The context window is how much the model can consider at once.
10. C. Own, changing documents with sources and least effort point to a managed RAG knowledge base.
11. D. Examples in the prompt are few-shot prompting.
12. A. A strict style with many examples, where prompts failed, is a case for customisation.
13. B. ROUGE compares generated text with reference text; it is common for summaries.
14. C. OpenSearch Service supports vector search (others include Aurora PostgreSQL with pgvector and S3 Vectors).
15. A. Always try a clearer prompt first; it is the cheapest change.
16. B. Different outcomes for similar people across groups is a bias concern.
17. D. SageMaker Clarify detects bias and explains predictions.
18. C. Model Cards record intended use, metrics and limitations.
19. A. CloudTrail records API calls; CloudWatch is metrics and logs.
20. B. The customer controls who has access to its resources.

Expected output (yours will be similar, not the same): your score out of 20 and the numbers of the questions you missed, for example Asha's "16/20, missed 8, 13, 17 and 20". Keep that list; example 3 turns it into a score for each domain.

6. Worked example 3: score yourself by domain

A total score hides where you are weak. This script scores the 20 answers by domain. The answers here belong to Asha, a dummy learner; put yours in the mine line.

KEY = "BCADB ACDBC DABCA BDCAB".replace(" ", "")          # answer key for the 20 questions on the page
DOMAIN = [1,1,1,1, 2,2,2,2,2, 3,3,3,3,3,3, 4,4,4, 5,5]     # domain of each question
NAMES = {1: "AI and ML basics", 2: "GenAI basics", 3: "Using foundation models",
         4: "Responsible AI", 5: "Security and governance"}
mine = "BCADB ACBBC DACCA BACAD".replace(" ", "")         # Asha's answers (dummy learner)

score = {d: [0, 0] for d in NAMES}
for k, m, d in zip(KEY, mine, DOMAIN):
    score[d][1] += 1; score[d][0] += (k == m)
total = sum(s[0] for s in score.values())
print(f"Total: {total}/20 ({total * 5}%)")
for d, (right, n) in score.items():
    pct = round(100 * right / n)
    print(f"  D{d} {NAMES[d]:<26} {right}/{n}  {pct:>3}%  {'REVISE' if pct < 70 else 'ok'}")
weak = min(score, key=lambda d: score[d][0] / score[d][1])
print(f"Start revision with: D{weak} {NAMES[weak]}")

Real output (19 October 2026):

Total: 16/20 (80%)
  D1 AI and ML basics           4/4  100%  ok
  D2 GenAI basics               4/5   80%  ok
  D3 Using foundation models    5/6   83%  ok
  D4 Responsible AI             2/3   67%  REVISE
  D5 Security and governance    1/2   50%  REVISE
Start revision with: D5 Security and governance

Asha's 80% looks safe, but domains 4 and 5 are weak. Her plan: re-read the privacy chapter and the security sections of the Bedrock chapters, add Clarify, Model Cards, CloudTrail and the shared responsibility model to her look-alike list, and take a fresh set of practice questions from AWS Skill Builder in a week. With only 20 questions a domain score is rough, so use it to choose what to revise, not to predict your exam result.

7. Traps that cost the most marks

• "Least effort" or "lowest cost" almost never means training or fine-tuning.
• Look-alike services: CloudTrail (who did what) vs CloudWatch (metrics and logs); Textract (documents) vs Transcribe (speech); Polly (speaks) vs Lex (chatbots).
• "Select TWO" questions: you must choose exactly 2, and both must be right.
• Absolute words like "always" and "guarantees" are often wrong. No method guarantees zero hallucinations.
• Time: 90 minutes for 65 questions is about 80 seconds each. Mark hard ones and come back.

8. Beginner mistakes and quick fixes

9. Try at home

Recap

Practise the way the exam asks: short scenarios, close options, one key word that decides. Score by domain, fix the weakest, and repeat.

Previous chapter: Amazon Q and Kiro: Write and Fix AWS Cloud Code Faster with AI

Next chapter: AI Certifications Compared: AWS vs Microsoft vs Google, India 2026

Ravindra Bagale, trainer: linkedin.com/in/ravindra-bagale. Asha and her answers are made up for learning, and the 20 practice questions are our own, not real exam questions. Exam facts were checked on the AWS Certified AI Practitioner page and the AIF-C01 exam guide on 19 October 2026. The Python output was produced on 19 October 2026.