AWS AI Practitioner (AIF-C01) Exam Guide 2026 — What's Actually on the Test

The AWS Certified AI Practitioner (AIF-C01) launched in late 2024 and has quickly become the most-attempted AWS AI certification. If you are choosing between a generic “AI for business” course and a vendor cert, AIF-C01 wins on credibility because AWS validates the specific services you would actually use at work — Amazon Bedrock, SageMaker, Comprehend, Rekognition — instead of theoretical machine learning concepts.

This guide is built from 500 questions in Certsly's AIF-C01 bank, mapped to the official AWS exam blueprint. It covers what each domain actually tests, the trap-question patterns we see most often during writing, a domain-weighted 4-week study plan, and how to know you are ready before you book the exam.

Why AIF-C01 matters in 2026

Generative AI moved from research lab to production fast. Every mid-sized engineering team in 2026 has at least one project that touches a foundation model — document summarisation, semantic search, support-ticket triage, internal copilots, marketing copy generation. The teams that ship those projects know which AWS service does what, when to use Bedrock versus SageMaker, when to fine-tune versus just prompt better, and how to keep prompt injection from leaking customer data.

AIF-C01 is the credential that signals you understand that landscape at an operational level. It is not the deepest AWS AI certification (Machine Learning Specialty and Machine Learning Engineer Associate go further), but it is the most accessible: no prerequisites, $100 fee, and a syllabus that maps directly to the questions a junior or mid-level engineer faces when shipping AI features.

For hiring managers in 2026, AIF-C01 on a resume signals two things. First, the candidate has been deliberate about learning AWS AI services rather than picking up random ChatGPT tutorials. Second, they understand responsible-AI guardrails — a non-trivial concern when shipping LLM features to production. Both signals matter when the rest of the resume is light on AI experience.

What AIF-C01 actually tests

The exam has five domains. AWS publishes domain weights as a percentage of total scored questions:

The biggest chunk (28 percent) is foundation models — choosing the right model for a use case, customising them (prompt engineering, retrieval-augmented generation, fine-tuning), and evaluating outputs. Do not underweight Responsible AI either: it is only 14 percent but AWS phrases the questions like FAQs (“Which AWS service helps detect bias in training data?”) and they are easy points once you have memorised the service map.

Domain 1 — Fundamentals of AI and ML (20%)

This is the vocabulary domain. AWS wants to know you can tell supervised from unsupervised, classification from regression, and that you understand what an embedding is without confusing it with a vector database.

What gets tested:

• ML lifecycle stages (data collection → preparation → training → evaluation → deployment → monitoring)
• Supervised vs unsupervised vs reinforcement learning with use-case examples
• Common ML metrics: accuracy, precision, recall, F1, AUC, RMSE, MAE
• Inference patterns: batch vs real-time vs streaming
• AWS data pipeline services for ML: SageMaker, Glue, Athena, Lake Formation, S3 (as the canonical data lake)
• Vector embeddings and vector databases (OpenSearch, Aurora pgvector, Kendra)

Trap pattern we see most: AWS phrases questions so that two AWS services could technically work but only one is the “best practice” choice. Example: “A company needs real-time fraud detection on streaming transactions.” Both Kinesis Data Streams + SageMaker AND Bedrock could technically work, but the expected answer is SageMaker (purpose-built for ML inference on streaming data, not Bedrock which is for generative tasks). Always ask yourself which service is purpose-built for the workload described.

Domain 2 — Fundamentals of Generative AI (24%)

The largest non-foundation-model domain. Expect questions on prompt engineering basics, in-context learning, and the difference between traditional ML models and large language models.

What gets tested:

• LLM fundamentals: tokens, context window, temperature, top-p, top-k, frequency penalty
• Prompt engineering techniques: zero-shot, few-shot, chain-of-thought, ReAct
• Foundation model categories: text generation, image generation, multimodal, embeddings, classification
• Use cases: summarisation, classification, translation, code generation, semantic search, RAG
• Limitations and risks: hallucination, knowledge cutoff, prompt injection, data leakage
• Generative AI lifecycle: identify use case → model selection → customisation → evaluation → deployment → monitoring

Trap pattern: “Lowest cost” and “no fine-tuning” questions. If the prompt says the company wants to avoid fine-tuning and minimise cost, the answer is almost always prompt engineering or RAG — not fine-tuning, not custom model training. The cost order from cheapest to most expensive: prompt engineering < RAG < fine-tuning < continued pre-training < training from scratch. Memorise that order and 80 percent of cost questions become trivial.

Domain 3 — Applications of Foundation Models (28%)

The largest domain. This is where AWS-specific service knowledge matters. You need to know what Bedrock does vs what SageMaker JumpStart does vs what Amazon Q does — and when to use each.

What gets tested:

• Amazon Bedrock: model catalog, agents, knowledge bases (RAG), guardrails, model evaluation
• SageMaker JumpStart vs SageMaker Studio vs SageMaker AI (formerly SageMaker Unified Studio)
• Amazon Q (Q Business, Q Developer) — packaged generative-AI assistants
• Specialised services: Comprehend, Translate, Polly, Transcribe, Rekognition, Textract, Forecast, Personalize, Lex
• Customisation techniques in order of cost and effort: prompt engineering → RAG → fine-tuning → continued pre-training
• Bedrock Agents: how they orchestrate tools, action groups, and knowledge bases

Trap pattern: Bedrock vs SageMaker confusion. Bedrock is managed access to foundation models via API — no infrastructure to manage, you do not see the model weights, billing is per token. SageMaker is infrastructure for ML workflows — you bring your own model or train one, you pay for the underlying compute. If the scenario says “no ML expertise required” or “fastest time to integration,” it is Bedrock. If it says “custom model,” “full training control,” or “deploy our own fine-tuned model,” it is SageMaker.

Domain 4 — Guidelines for Responsible AI (14%)

Smaller domain but easy points. AWS wants to know you understand bias, fairness, transparency, and the AWS services that address each.

What gets tested:

• Sources of bias: data bias, algorithm bias, sampling bias, label bias
• SageMaker Clarify: bias detection during training, post-training bias metrics, model explainability via SHAP values
• Bedrock Guardrails: content filters, denied topics, sensitive-info filters, contextual grounding checks
• Model cards and data cards: documentation artefacts for transparency
• Human-in-the-loop workflows (Amazon Augmented AI / A2I) for reviewing low-confidence predictions
• Fairness, accountability, transparency, ethics — the FATE principles

Trap pattern: Confusing Bedrock Guardrails with SageMaker Clarify. Guardrails operate at runtime — they filter inputs and outputs of a model at inference time. Clarify operates at development time — you run it on training data and model predictions to surface bias before you deploy. The exam will pair these in near-identical scenarios to test whether you know the distinction.

Domain 5 — Security, Compliance, and Governance (14%)

Mostly about IAM, encryption, and VPC connectivity for AI workloads.

What gets tested:

• IAM roles and least-privilege access for AI services (Bedrock, SageMaker, Comprehend)
• Encryption at rest (KMS) and in transit (TLS) for Bedrock and SageMaker
• VPC endpoints (PrivateLink) for private API access to Bedrock without traversing the public internet
• AWS service-specific compliance: Bedrock data isolation, model invocation logging via CloudTrail
• Audit and observability: CloudTrail for API calls, CloudWatch for metrics, Bedrock invocation logs
• Data residency: which AWS regions support which Bedrock models, prompt logging settings

Trap pattern: “Does AWS use my Bedrock prompts to train base models?” The answer is no — in standard Bedrock usage, model providers do not see your prompts or outputs. AWS does not use your inputs to train its or third-party base models. Memorise this exactly; the exam will phrase it three different ways and the answer is always the same.

Sample question walkthrough

Here is a representative AIF-C01 question shape, with the kind of analysis the exam rewards:

Notice the question gives you two explicit constraints (“fastest path” and “no ML engineers”). Those constraints disqualify A and B immediately. Then the remaining choice between C and D is about which service category fits an open-ended Q&A use case — that is RAG, which is Bedrock's knowledge-base feature.

AIF-C01 vs Cloud Practitioner (CLF-C02) — which first?

Both are foundational AWS certifications, but they serve different purposes:

• CLF-C02 (Cloud Practitioner) tests general AWS literacy: compute, storage, networking, billing, support plans, the shared responsibility model. It is the cert most non-technical AWS users start with.
• AIF-C01 (AI Practitioner) is narrower and deeper on AI services specifically. It assumes you already know what S3 is and that EC2 exists, but it does not require deep AWS infrastructure expertise.

If you have zero AWS exposure: take CLF-C02 first. The 20 percent of AIF-C01 that touches general AWS services will feel less abstract once you have the CLF-C02 vocabulary.

If you already work alongside AWS engineers and your role is AI-adjacent: skip CLF-C02 and go directly to AIF-C01. You will pick up the general AWS context from the AIF-C01 study materials.

4-week study plan

Total commitment: ~34 hours over 4 weeks, plus the 500 Certsly questions or any equivalent practice bank.

How to know you are ready

Score consistently above 80 percent on full-length Certsly mock exams (65 questions, 90-minute timer). We calibrate questions slightly above exam difficulty, so 80 percent on Certsly predicts a comfortable pass on the live AWS exam. If you score 70 to 79 percent, do another 100 questions and rerun the mock. Below 70 percent means revisit Domain 3 — that is where most failures cluster.

A second signal: can you read a 4-sentence Bedrock vs SageMaker scenario and pick the right service in under 60 seconds without hovering between two answers? If yes, you are ready. If you find yourself rationalising both options on every question, do more Domain 3 practice.

Common mistakes to avoid

1. Memorising services without understanding when to pick which. AIF-C01 questions are scenario-based. Knowing “Bedrock exists” is worthless on the exam; knowing why you would pick Bedrock over SageMaker JumpStart for a specific scenario is everything.
2. Skipping Responsible AI. It is only 14 percent but the questions are predictable. Memorise Guardrails vs Clarify and you bank 8 to 10 easy points.
3. Ignoring the IAM and security domain. Same as above — small weight, easy points if you have worked with AWS IAM before.
4. Taking the exam without a full timed mock. The 90-minute time pressure surprises people. Do at least two full timed mocks before booking.
5. Studying outdated Bedrock content. Bedrock moved fast through 2025 (knowledge bases, agents, guardrails, evaluations). Anything published before mid-2025 is partial. Stick to AWS's own docs and 2026-dated study materials.
6. Relying on exam dumps. AWS bans them and they will not help — the exam pool rotates faster than dumps can keep up. Original practice questions mapped to the blueprint are the only reliable path.

Frequently asked questions