AWS Certified AI Practitioner (AIF-C01)
Optional foundational exam: AI, ML and generative AI concepts, foundation model applications, responsible AI and AI governance on AWS.
Verified against official sources on 2026-11-17.
AWS Certified AI Practitioner (AIF-C01) checks that you understand AI, ML and generative AI concepts, can recognise good and bad uses of them, and know how to use them responsibly and securely on AWS. It's optional in your roadmap. Take it if your bank is starting GenAI work and you want a shared vocabulary with risk and business teams.
At a glance
| Code | AIF-C01 |
| Level | Foundational |
| Intended audience | Up to 6 months' exposure to AI and ML on AWS. People who use AI and ML solutions but don't necessarily build them |
| Questions | 65 (50 scored + 15 unscored) |
| Time | 90 minutes |
| Passing score | 700 (scaled 100–1000) |
| Price | USD 100 |
| Question types | Multiple choice, multiple response, ordering, matching |
| Valid for | 3 years. Renewed by passing MLA (ML Engineer – Associate) or AIP-C01. There's no maintain option |
| Out of scope | Building models, data and feature engineering, hyperparameter tuning, building ML pipelines, statistical analysis, implementing AI security protocols or governance frameworks |
Official domains and weightings
| Domain | Weight |
|---|---|
| 1. Fundamentals of AI and ML | 20% |
| 2. Fundamentals of GenAI | 24% |
| 3. Applications of Foundation Models | 28% |
| 4. Guidelines for Responsible AI | 14% |
| 5. Security, Compliance, and Governance for AI Solutions | 14% |
Key AWS services tested
- Generative AI: Amazon Bedrock (models, Knowledge Bases, Agents, Guardrails, model evaluation), Amazon Q
- ML platform: Amazon SageMaker AI (including Clarify for bias and explainability, Model Monitor, model cards, JumpStart, Ground Truth)
- Managed AI services: Amazon Comprehend, Textract, Transcribe, Translate, Rekognition, Lex, Polly, Kendra, Personalize
- Security and governance: IAM, KMS, CloudTrail, Macie, PrivateLink, AWS Artifact, Audit Manager, Config
- ⚠ VERIFY the exact service list against the in-scope services page of the exam guide before you study, because AI services change quickly
Several services in this list are closed to new customers
AWS put these services into maintenance mode, and they stopped accepting new customers on 30 July 2026. Existing customers can still use them, but a new AWS account (like yours) can't. The exam guide may still list them, so learn what each one does for the exam, then build with the current alternative:
| Service in the exam guide | Status | Use instead in real projects |
|---|---|---|
| Amazon Bedrock Agents (now Agents Classic) | Closed to new customers | Amazon Bedrock AgentCore. Bedrock models, Knowledge Bases and Guardrails are not affected |
| Amazon Kendra | Closed to new customers | Amazon Bedrock Knowledge Bases |
| Amazon Q Business | Closed to new customers | Amazon Quick |
| SageMaker AI Clarify | Closed to new customers | The open-source SHAP library plus standard bias metrics in your own pipeline; Bedrock Evaluations for foundation models |
| SageMaker AI Model Monitor | Closed to new customers | AWS's open-source SageMaker AI monitoring solutions (MLflow and Evidently AI) with CloudWatch alarms |
| SageMaker AI Ground Truth, Amazon Augmented AI (A2I) | Closed to new customers | AWS names no direct replacement. Build human review with a Step Functions approval step; use Bedrock Evaluations human-evaluation jobs for FM output |
Sources: AWS service availability updates and AWS product lifecycle, checked on 2026-11-18.
What you already know vs what you must learn
| You already know | You must learn |
|---|---|
| Using Copilot or ChatGPT-style tools | Concepts and vocabulary: supervised, unsupervised and reinforcement learning, tokens, embeddings, context windows, temperature, hallucination |
| Calling REST APIs, handling JSON | Choosing the approach: prompt engineering compared with RAG, fine-tuning and pre-training (cost, effort, freshness) |
| Model risk and fairness concerns in banking (credit, fraud) | Responsible AI: bias, explainability, transparency, human oversight. SageMaker Clarify as exam vocabulary (closed to new customers; use SHAP and your own bias metrics in practice) |
| Data protection, PII handling | AI-specific risks: prompt injection, data leakage, model output governance. Where the shared responsibility line sits for AI services |
Preparation order and study plan (about 15–20 hours over 2–3 weeks)
| Week | Focus | Hours |
|---|---|---|
| 1 | Exam guide. Module 16 overview. Domains 1 and 2 (concepts) | 6–8 |
| 2 | Domain 3 (RAG, prompting, choosing models, evaluation). Domains 4 and 5 | 6–8 |
| 3 | Free Exam Prep course, the 20-question set, this page's questions, ordering and matching practice | 3–4 |
Mapped platform modules
- Module 16 — Generative AI on AWS for banking: the core of every domain
- Module 05 — Security: Domain 5 (IAM, KMS, logging, data classification)
- Module 01 — Fundamentals: the shared responsibility model and pricing basics
Practical skills to build
- Call a Bedrock model from C# and see how temperature and max tokens change the output
- Build a small Knowledge Base over fictional policy PDFs and compare answers with and without retrieval
- Add a guardrail that blocks a denied topic (for example, investment advice) and masks PII
- Write a one-page "AI use-case risk note" for a fictional bank use case: data, users, risks, human-in-the-loop, monitoring
Exam-prep checklist
- I can tell AI, ML, deep learning and generative AI apart, and give a banking example of each
- I know when to use prompt engineering, RAG, fine-tuning or continued pre-training
- I can explain inference parameters (temperature, top-p, max tokens) and their effects
- I know the evaluation approaches (human, automatic metrics, LLM-as-a-judge) at a high level
- I can name the responsible AI dimensions and the AWS tools for each (Clarify, Guardrails, model cards), and I know which of them are closed to new customers
- I know the AI security risks (prompt injection, data leakage) and the controls for them
- I've practised ordering and matching questions
Official links
Practice questions
These 5 questions are original, banking-flavoured scenarios. They aren't real exam questions.
A bank has 10 years of loan records, each labelled with whether the borrower defaulted. It wants to estimate the probability that a new applicant will default. Which approach fits?
- A. Generative AI text generation with a large language model
- B. Supervised learning (binary classification)
- C. Unsupervised clustering
- D. Reinforcement learning
Correct answer: B
Explanation: Labelled historical outcomes (defaulted or not) plus a prediction of that same label for new cases is the definition of supervised classification. Credit decisions are heavily regulated, so expect explainability, fairness testing and human oversight requirements, and a model-risk review.
Why the others are wrong:
- A: LLMs generate content. They aren't the right tool for a calibrated risk probability from tabular data.
- C: Clustering finds groups in unlabelled data, but here labels exist and a prediction is needed.
- D: Reinforcement learning learns from rewards through interaction over time. It doesn't fit a fixed labelled dataset.
An internal assistant must answer staff questions from the bank's policy documents, which are updated every week, and must show which document each answer came from. Which approach is most suitable and cost-effective?
- A. Fine-tune a foundation model every week on the latest documents
- B. Retrieval-Augmented Generation (RAG), for example with Amazon Bedrock Knowledge Bases, re-syncing the data source when documents change
- C. Pre-train a new foundation model on the documents
- D. Paste all policy documents into every prompt
Correct answer: B
Explanation: RAG retrieves the relevant passages at question time and passes them to the model. Updating the knowledge means re-indexing documents, not retraining, and the retrieved passages give you citations.
Why the others are wrong:
- A: Weekly fine-tuning is expensive, slow, and doesn't give reliable citations.
- C: Pre-training is the most expensive option by far and isn't needed here.
- D: Large prompts hit context-window limits and cost a lot for every question.
A compliance FAQ bot gives noticeably different wording and sometimes different facts when asked the same question several times. The team wants more consistent, deterministic responses. Which change helps most?
- A. Increase the temperature
- B. Lower the temperature, for example close to 0
- C. Increase the maximum output tokens
- D. Switch to a model with a larger context window
Correct answer: B
Explanation: Temperature controls how random the choice of the next token is. Lower values make the model pick the most likely tokens, which gives more consistent output. For factual compliance answers, combine a low temperature with RAG grounding and a guardrail to reduce hallucinations further.
Why the others are wrong:
- A: A higher temperature increases variety, which is the opposite of what's needed.
- C: Max tokens limits response length, not consistency.
- D: A bigger context window lets you include more input. It doesn't make outputs deterministic.
A model-risk review finds that a credit-scoring model approves applicants from one age group noticeably less often than others with similar financial profiles. Which AWS capability helps the team measure bias and explain the model's predictions?
- A. Amazon Comprehend
- B. Amazon SageMaker Clarify
- C. Amazon Bedrock Guardrails
- D. Amazon Macie
Correct answer: B
Explanation: SageMaker Clarify calculates bias metrics before and after training, and provides feature-attribution explanations (SHAP-based) for model predictions. That's exactly what a fairness investigation needs. Whether the result is lawful is for the bank's legal and model-risk teams to decide.
In practice: Clarify has been closed to new customers since 30 July 2026, but it can still appear on the exam. In a new account, compute the same bias metrics with pandas and scikit-learn, and explanations with the open-source SHAP library that Clarify itself uses.
Why the others are wrong:
- A: Comprehend does natural language processing (entities, sentiment, PII). It doesn't measure bias in a tabular model.
- C: Guardrails filter GenAI inputs and outputs (topics, harmful content, PII). They don't measure statistical bias.
- D: Macie discovers sensitive data in S3.
The bank allows only one approved Bedrock model for a pilot. Only pilot team members may call it, and every invocation must be auditable. Which two controls should be used?
- A. IAM policies that allow
bedrock:InvokeModelonly on the approved model's ARN, attached only to the pilot team's role - B. Turn on Bedrock model invocation logging to a protected destination, and use CloudTrail for the API audit trail
- C. Amazon Inspector
- D. AWS Shield Advanced
- E. S3 Transfer Acceleration
Correct answers: A and B
Explanation: IAM scoped to the specific model resource enforces least privilege. CloudTrail records who called the Bedrock APIs. Invocation logging captures request and response data for review, so protect that destination well, because it may contain sensitive prompts. An SCP can deny all other models across the organisation as an extra safeguard.
Why the others are wrong:
- C: Inspector scans for software vulnerabilities.
- D: Shield Advanced protects against DDoS attacks.
- E: Transfer Acceleration speeds up long-distance S3 uploads.
Independent learning resource — not affiliated with, endorsed or sponsored by Amazon Web Services. AWS and related marks are trademarks of Amazon.com, Inc. or its affiliates. All banking data in examples is fictional. Sample architectures are for learning and are not automatically compliant with GDPR, PCI DSS or any banking regulation.
AWS Certified Solutions Architect – Professional (SAP-C03)
The hardest exam in the roadmap. It covers complex, multi-account, hybrid, secure, resilient and AI-integrated architectures.
AWS Certified Generative AI Developer – Professional (AIP-C01)
Optional professional exam: building production GenAI applications on AWS with RAG, agents, guardrails, evaluation and cost control.