AWS AIF-C01 AI Practitioner guide for AI/ML fundamentals, GenAI, foundation model applications, responsible AI, governance, Bedrock, SageMaker AI, RAG, and prompting.
This guide targets AWS Certified AI Practitioner (AIF-C01), AWS’s foundational certification for candidates who need to understand AI, ML, and generative AI on AWS without dropping into engineer-level implementation depth. The exam is mostly about business fit, service fit, risk fit, and control fit: what problem the business is trying to solve, which AI pattern fits, which AWS-managed path is appropriate, and which responsible-AI or governance control keeps the answer safe.
ML: Machine learning, which trains models from data so they can make predictions or classifications on new inputs.
| Exam fact | Current official value |
|---|---|
| Level | Foundational |
| Duration | 90 minutes |
| Format | 65 questions |
| Cost | 100 USD |
| Delivery | Pearson VUE testing center or online proctored |
| Scored / unscored mix | 50 scored + 15 unscored |
| Question types | Multiple choice, multiple response, ordering, matching |
| Passing score | 700 scaled |
| Validity | 3 years |
| Target candidate | Up to 6 months of exposure to AI/ML on AWS; uses AI/ML solutions, but does not necessarily build them |
As of May 17, 2026, AWS’s current certification page and current exam guide are aligned on the domain structure, candidate level, duration, and question count. The exam guide adds the scored vs unscored split, question-type mix, and passing-score details that AWS uses on the live exam.
AWS frames AIF-C01 as a certification for people who are familiar with AI and ML on AWS, but who do not necessarily build the solutions themselves. That keeps the best answer at the right altitude: use case, AI pattern, AWS service category, risk, cost, and governance boundary rather than algorithm development, feature engineering, hyperparameter tuning, or production pipeline implementation.
AWS’s current exam guide breaks AIF-C01 into five weighted domains, and this online guide now follows that structure directly:
| Official domain | Weight | What the exam is really checking |
|---|---|---|
| Fundamentals of AI and ML | 20% | Can you separate AI, ML, deep learning, supervised learning, unsupervised learning, model lifecycle, data quality, and business value without over-engineering the answer? |
| Fundamentals of GenAI | 24% | Can you explain foundation models, tokens, embeddings, prompts, hallucinations, limitations, value, and AWS GenAI service categories at practitioner depth? |
| Applications of Foundation Models | 28% | Can you choose between prompting, RAG, agents, model customization, evaluation, and managed AWS services for a business scenario? |
| Guidelines for Responsible AI | 14% | Can you recognize bias, fairness, explainability, transparency, human oversight, unsafe outputs, and responsible use patterns? |
| Security, Compliance, and Governance for AI Solutions | 14% | Can you keep IAM, encryption, data privacy, logging, monitoring, compliance, and governance in the same decision as model choice? |
AIF-C01 is not asking you to build a model from scratch. It is asking whether you can choose the right AI pattern and recognize when a more advanced-sounding answer is actually less appropriate for the stated business risk.
| Domain | Weight | What to study first |
|---|---|---|
| 1. AI/ML Basics | 20% | 1.1 AI/ML Data, 1.2 Use Cases, 1.3 ML Lifecycle |
| 2. GenAI Basics | 24% | 2.1 Tokens & Models, 2.2 Value & Limits, 2.3 AWS GenAI Services |
| 3. FM Apps | 28% | 3.1 RAG & Agents, 3.2 Prompting, 3.3 Fine-Tuning, 3.4 Evaluation |
| 4. Responsible AI | 14% | 4.1 Bias & Fairness, 4.2 Transparency |
| 5. Governance | 14% | 5.1 Securing AI, 5.2 Governance |
flowchart LR
Q["Business problem or AI use case"] --> C{"Which AI lane?"}
C -- "Prediction or classification" --> M["Traditional ML: data, training, evaluation, lifecycle"]
C -- "Generation or conversation" --> G["GenAI: model, prompt, grounding, output risk"]
C -- "Current or private facts" --> R["RAG or knowledge base before fine-tuning"]
C -- "Action or workflow" --> A["Agent or orchestration with tool boundaries"]
M --> S["Check service fit, risk, cost, and governance"]
G --> S
R --> S
A --> S
Use that loop on scenario questions before reading the answer choices. Many distractors sound plausible because they name real AWS services, but they solve the wrong class of AI problem.
| Scenario clue | First move |
|---|---|
| The business asks whether AI fits a problem | Classify the work as prediction, classification, search, generation, summarization, or automation before choosing a service. |
| The answer must use current or private facts | Think RAG, knowledge bases, grounding, citations, and data access controls before thinking fine-tuning. |
| The prompt includes sensitive customer data | Bring in privacy, IAM, encryption, logging behavior, and data governance, not just model accuracy. |
| The output could be harmful, biased, or misleading | Add responsible AI controls such as guardrails, human review, explainability, and evaluation. |
| The stem emphasizes cost or latency | Prefer a smaller fit-for-purpose model, prompt efficiency, managed service fit, and realistic usage controls. |
| The question compares Bedrock and SageMaker AI | Bedrock usually means managed foundation model access and GenAI application patterns; SageMaker AI usually means broader ML model development and lifecycle work. |
| Failure pattern | Better instinct |
|---|---|
| treating every AI question like a GenAI question | ask whether the problem is prediction, classification, search, generation, or governance first |
| picking the most advanced-sounding pattern | prefer the simplest AWS-native design that satisfies the stated business goal |
| answering like an ML engineer | stay in the practitioner lane unless the stem explicitly asks for deeper implementation detail |
| handling safety, privacy, and governance only after model choice | treat them as part of the answer, not as postscript controls |
| Trap | Why it is wrong |
|---|---|
| Assuming every AI question is a GenAI question | Many scenarios still test traditional ML concepts such as classification, regression, training data, evaluation, and lifecycle. |
| Choosing the largest or newest model by default | AIF-C01 rewards fit, cost, latency, risk, and manageability, not novelty. |
| Fine-tuning when the issue is fresh or private knowledge | RAG and knowledge bases usually fit current facts better than changing model weights. |
| Treating governance as paperwork after the solution | Security, compliance, privacy, auditability, and responsible AI are part of the correct design choice. |
| Picking implementation-heavy answers for business-fit questions | The target candidate uses AI/ML on AWS, but is not expected to code models, tune hyperparameters, or build pipelines. |
flowchart LR
A["Study plan"] --> B["1. AI and ML fundamentals"]
B --> C["2. GenAI fundamentals"]
C --> D["3. Foundation model applications"]
D --> E["4. Responsible AI"]
E --> F["5. Security, compliance, and governance"]
F --> G["Cheat sheet, glossary, and final review"]
| If you are coming from… | Bias your review toward… |
|---|---|
| business, product, or project roles | use cases, service categories, and governance language |
| cloud or IT support | service fit, security boundaries, and AWS terminology |
| future AI or ML engineering path | foundational concepts now, then move into MLA-C01 next |
For AIF-C01, the better answer is usually the one that: