How many questions are scored, and how long do I get?

65 questions in 90 minutes, but only 50 count toward your score. The other 15 are unscored questions mixed in and not marked, so treat every question as if it counts.

What score do I need to pass?

700 on a 100–1,000 scaled-score range. It is a single pass/fail result using compensatory scoring, so a weak domain can be offset by strong ones — you only need to clear 700 overall.

Are there hands-on labs or performance-based questions?

No. AIF-C01 uses only multiple choice, multiple response, ordering, and matching questions. There are no console-based labs.

Do I need experience or prerequisites to take it?

There are no required prerequisites. AWS recommends up to 6 months of exposure to AI/ML on AWS and familiarity with core services like Amazon Bedrock and Amazon SageMaker AI, but anyone can register.

How much does it cost and how long is it valid?

USD $100 per attempt. The certification is valid for 3 years; you recertify by passing the latest version of the exam or by earning AWS Certified Machine Learning Engineer – Associate.

What changed recently — is the exam different now?

Yes. As of exam guide version 1.1 (published April 30, 2026, live on exams from roughly late May 2026), the exam adds significant agentic AI content (AI agents, MCP, multi-agent patterns, plus in-scope services like Amazon Bedrock AgentCore and Strands Agents), context engineering, token-based pricing, model distillation, LLM-as-a-judge evaluation, and hallucination-detection/grounding topics. The five domains and their weights are unchanged.

How to Pass AWS Certified AI Practitioner (AIF-C01)

A practical, no-fluff guide to passing AWS Certified AI Practitioner (AIF-C01) — the current exam format and domain weights, scoring and cost, a realistic study plan, and the highest-leverage strategy to pass.

Last reviewed June 7, 2026. Exam logistics change — always confirm current details on the official certification site before you book.

Take a timed mock exam →

The exam at a glance

The AWS Certified AI Practitioner (AIF-C01) is a Foundational-level certification that proves you understand AI, machine learning, and generative AI concepts and how they map to AWS services. You do not need to build models to pass it — you need to know what the tools are and when to use them.

Questions: 65 total, but only 50 are scored. The other 15 are unscored questions mixed in and not marked — treat every question as if it counts.
Time: 90 minutes.
Item types: multiple choice, multiple response, ordering, and matching. There are no hands-on labs or performance-based questions.
Passing score: 700 on a 100–1,000 scale.
Fee: USD $100.
Validity: 3 years.
Delivery: Pearson VUE test center or online proctored, in 12 languages.

How it is scored

Your result is a single pass/fail with a scaled score from 100 to 1,000, and you need 700 to pass. Scaling lets AWS compare scores fairly across exam forms of slightly different difficulty, so don’t try to reverse-engineer “how many questions can I miss” — roughly 70% is a safe mental target.

The scoring is compensatory: you do not have to pass each domain individually, only the overall exam. A weaker area can be carried by a stronger one. Two more rules matter:

Unanswered questions are scored as incorrect, and there is no penalty for guessing. So answer every single question, even if you flag it and come back.
Section-level feedback on your score report is informational only — use it for future study, not as a per-domain pass bar.

Are you eligible — and what does it cost?

There are no formal prerequisites — anyone can register and pay the $100 fee. AWS recommends up to 6 months of exposure to AI/ML on AWS as the target experience level, and assumes you are familiar with (not necessarily building) AI solutions.

Helpful background before you start:

Core AWS services and their use cases (Amazon EC2, S3, Lambda, Amazon Bedrock, Amazon SageMaker AI).
The AWS shared responsibility model.
AWS IAM basics and AWS pricing models.

If AWS itself is brand new to you, do AWS Cloud Practitioner Essentials or AWS Technical Essentials first — it makes the security and governance domains far easier.

Build a realistic study plan

Most candidates with some cloud exposure pass in 3–5 weeks studying part-time. Here is a concrete 4-week plan.

Week 1 — Foundations (Domain 1, 20%). Lock in the vocabulary: AI vs. ML vs. deep learning vs. generative AI vs. agentic AI, supervised/unsupervised/reinforcement learning, inferencing types (batch, real-time, asynchronous, serverless), and the AI/ML pipeline stages. Map services to stages (SageMaker AI, Amazon Bedrock, Amazon Q).
Week 2 — Generative AI and Foundation Models (Domains 2 & 3, 52% combined). This is over half the exam — spend the most time here. Learn foundation models, Amazon Bedrock, prompt engineering, RAG, fine-tuning vs. pre-training vs. in-context learning, token-based pricing, context engineering, agents, and Model Context Protocol (MCP).
Week 3 — Responsible AI + Security/Governance (Domains 4 & 5, 28%). Bias and fairness, transparency/explainability (SageMaker Model Cards, SageMaker Clarify), Amazon Bedrock Guardrails, data privacy, IAM, encryption, hallucination detection and grounding, and the shared responsibility model for AI.
Week 4 — Practice and polish. Take the official AWS Skill Builder practice exam and at least one full timed mock. Review every wrong answer until you know why it’s wrong.

Free official resources: the AWS Skill Builder exam-prep plan and the Official Practice Question Set.

The exam mindset / highest-leverage strategy

The single biggest lever: the two GenAI/foundation-model domains are 52% of the exam. If you can confidently answer questions about Amazon Bedrock, prompt engineering, RAG, fine-tuning, and agents, you are most of the way to 700.

Think like a business decision-maker, not an engineer. Questions ask “which AWS service / approach fits this scenario,” not “write the code.”
Learn the service-to-job mapping: Bedrock = build with foundation models; SageMaker AI = build/train custom ML; Amazon Q = AI assistant; Bedrock AgentCore / Strands Agents = build and run AI agents; Guardrails = safety filters; Macie = sensitive-data discovery.
For “which customization approach” questions, remember the cost/effort ladder: prompt engineering < RAG < fine-tuning < continued pre-training.
On the day, do an easy first pass, flag the hard ones, and come back. Never leave a question blank.

Master the domains

Domain weights — spend your study time in proportion.

Domain 1 — Fundamentals of AI and ML (20%). Definitions and the AI/ML lifecycle. Know AI/ML/deep-learning/GenAI/agentic-AI distinctions, learning types, inferencing types (batch, real-time, asynchronous, serverless), and model metrics (accuracy, precision, recall, F1, AUC). Also know when traditional ML vs. foundation models fit a use case. Tested mostly as definition and “match the concept” questions.
Domain 2 — Fundamentals of Generative AI (24%). Foundation models, tokens and token-based pricing, context engineering, foundational agentic AI concepts (multi-agent patterns, MCP, memory, tool use, orchestration), capabilities and limitations (including hallucinations), business value/ROI, and AWS GenAI services (Amazon Bedrock, Amazon SageMaker AI, SageMaker JumpStart, Amazon Q, Strands Agents, Bedrock AgentCore). Tested as use-case selection.
Domain 3 — Applications of Foundation Models (28%, the biggest). Designing FM applications: prompt engineering, prompt versioning (Bedrock Prompt Management), RAG, fine-tuning vs. pre-training vs. model distillation, evaluation (ROUGE, BLEU, BERTScore, LLM-as-a-judge), the business role of AI agents, and business-objective alignment metrics. Heavy on scenario questions.
Domain 4 — Guidelines for Responsible AI (14%). Fairness, bias, transparency, explainability, and human-centered design. Know SageMaker Clarify, Model Cards, Bedrock Model Evaluations, and Bedrock Guardrails.
Domain 5 — Security, Compliance, and Governance (14%). IAM, encryption at rest/in transit, the shared responsibility model, data privacy, prompt injection, data-leakage prevention, output filtering/validation, audit logging, hallucination detection and grounding (RAG grounding, output validation, confidence scoring), and governance. Distinguish security (protecting the system) from governance (policies and oversight).

Common pitfalls

Underestimating Domains 2 and 3. They are 52% of the exam; don’t burn your time on definitions in Domain 1.
Confusing Bedrock with SageMaker. Bedrock = pre-built foundation models via API; SageMaker AI = build/train/host your own ML. Expect questions that hinge on this.
Leaving questions blank. No penalty for guessing — always answer.
Mixing up the customization approaches. RAG adds external knowledge at query time; fine-tuning changes the model’s weights; model distillation creates a smaller, cheaper model from a larger one. These are different answers.
Ignoring the recent content update. The current exam (guide v1.1) adds agentic AI, MCP, AgentCore, Strands Agents, context engineering, token-based pricing, model distillation, and grounding — older study guides miss these.

After you pass

Your certification is valid for 3 years. Before it expires you recertify either by passing the latest version of AIF-C01, or by earning AWS Certified Machine Learning Engineer – Associate, which automatically recertifies this credential. AWS gives you a 50% discount voucher in your Certification Account to use toward recertification.

Natural next steps: AWS Certified Machine Learning Engineer – Associate (the deeper, hands-on ML cert) or, if you want broader cloud breadth, AWS Certified Cloud Practitioner or an Associate-level cert like Solutions Architect.

The week before, and exam day

In the final week:

Take a full, timed practice exam and review every miss.
Re-read the official exam guide and skim the revision history so no topic surprises you.
Drill the service-to-use-case mappings — that’s where easy points live.

On exam day:

For an online proctored exam, test your webcam/mic, clear your desk, and have valid ID ready. For a test center, arrive early with ID.
65 questions in 90 minutes ≈ 80 seconds each. Do an easy first pass, flag and return, and leave nothing blank.
Trust your preparation, read each scenario for the key requirement (cost? latency? safety? privacy?), and pick the most AWS-recommended answer.

You’ve got this — aim for 700, lead with the GenAI domains, and answer every question.

Quick-reference: exam tips by domain

Pulled from every term in this subject — a fast last-pass before exam day.

Applications of Foundation Models

Agent — Agents for Amazon Bedrock orchestrate multi-step actions against your APIs.
BERTScore — BERTScore captures meaning, catching paraphrases that ROUGE and BLEU miss.
BLEU — BLEU for translation, ROUGE for summarization is a useful exam shortcut.
Chain-of-Thought — Chain-of-thought improves accuracy on multi-step and math problems.
Continued Pre-training — Continued pre-training uses unlabeled data; fine-tuning uses labeled data.
Few-shot — Few-shot examples steer format and style without any training.
Fine-tuning — Try prompt engineering and RAG before fine-tuning — fine-tuning costs more than both; only continued pre-training or training from scratch costs more.
Human Evaluation — Human evaluation is the gold standard when automatic metrics fall short.
In-Context Learning — In-context learning changes output without changing model weights.
Jailbreaking — Jailbreaking targets safety controls; prompt injection targets app instructions.
Knowledge Base — Choose Knowledge Bases when you want RAG without building the pipeline yourself.
Max Tokens — Setting max tokens controls cost and prevents runaway output.
OpenSearch — OpenSearch Serverless is a common vector backend for Bedrock Knowledge Bases.
pgvector — pgvector lets Aurora PostgreSQL act as a vector store for RAG.
Prompt Engineering — It is the cheapest customization lever — try it before RAG or fine-tuning.
Prompt Injection — Prompt injection is the #1 OWASP risk for LLM applications.
RAG — RAG reduces hallucination and adds private knowledge without retraining the model.
ROUGE — ROUGE is the go-to automatic metric for summarization.
System Prompt — The system prompt frames behavior before any user message is processed.
Temperature — Low temperature for factual tasks, higher for creative ones.
Top-K — Smaller K narrows the candidate pool and reduces randomness.
Top-P — Top-P adapts the candidate pool to the model's confidence, unlike fixed Top-K.
Vector Database — RAG needs a vector store — on AWS, options include OpenSearch and Aurora pgvector.
Zero-shot — Zero-shot relies entirely on the model's pre-trained knowledge.

Fundamentals of Generative AI

Amazon Q — Amazon Q Business connects to enterprise data; Amazon Q Developer assists with code.
Attention — Attention is what makes transformers context-aware.
Bedrock — Bedrock is serverless and the default answer for hosting FMs on AWS.
Chunking — Chunk size is a key RAG tuning knob — too big dilutes relevance, too small loses context.
Context Window — Inputs that exceed the context window must be truncated, summarized, or retrieved via RAG.
Diffusion Model — Diffusion powers text-to-image generation, such as Amazon Nova Canvas and Stable Diffusion models on Amazon Bedrock.
Embedding — Similar meanings map to nearby vectors — the basis of semantic search and RAG.
Foundation Model — Foundation models are the 'FMs' you access through Amazon Bedrock.
Generative AI — Generative AI is built on foundation models; traditional ML usually predicts or classifies.
Hallucination — RAG and grounding reduce hallucination by supplying real source data.
Interpretability — Interpretability is closely tied to transparency and explainability in Domain 4.
JumpStart — JumpStart is for SageMaker users who want to deploy or fine-tune models quickly.
Large Language Model — LLM is a type of foundation model specialized for text.
Multimodal — Multimodal models can accept an image and answer questions about it.
Nondeterminism — Lowering temperature toward 0 makes output more deterministic.
PartyRock — PartyRock is for hands-on learning and prototyping without writing code.
Prompt — Prompt quality strongly affects output quality — hence prompt engineering.
Token — Pricing and context limits are measured in tokens, not words.
Transformer — The 2017 'Attention Is All You Need' transformer is the basis of today's LLMs.
Vector — Embeddings are vectors; similarity is measured by distance between them.

Fundamentals of AI and ML

Accuracy — Accuracy is misleading on imbalanced data — use F1 or AUC instead.
Artificial Intelligence — AI is the umbrella; machine learning is a subset of AI, and deep learning is a subset of ML.
AUC — AUC of 0.5 is random; 1.0 is perfect ranking.
Bias — Bias is both a statistical concept and a fairness concern — know both senses.
Comprehend — Reach for Comprehend when you need NLP without training a model.
Computer Vision — Amazon Rekognition is the managed AWS computer vision service.
Deep Learning — Deep learning shines on unstructured data — images, audio, and text.
F1 Score — Use F1 when classes are imbalanced and both error types matter.
Inference — Training builds the model; inference uses it — and inference is where ongoing cost lives.
Machine Learning — If the system improves from data rather than hand-coded rules, it is ML.
Neural Network — Layers between input and output are called hidden layers.
NLP — Amazon Comprehend handles NLP tasks like sentiment and entity extraction.
Overfitting — High training accuracy but low test accuracy is the classic overfitting signature.
Reinforcement Learning — RLHF (reinforcement learning from human feedback) is used to align LLMs.
Rekognition — Rekognition handles object detection, moderation, and facial analysis out of the box.
SageMaker — SageMaker is the 'build it yourself' option; pre-trained AI services are the 'use it' option.
Supervised Learning — Classification and regression are the two main supervised tasks.
Training Data — Quality and representativeness of training data directly drive model quality and bias.
Underfitting — Underfitting shows up as poor performance on both training and test data.
Unsupervised Learning — Clustering and dimensionality reduction are classic unsupervised techniques.

Guidelines for Responsible AI

Clarify — Clarify is the AWS answer for both bias detection and explainability.
Explainability — SageMaker Clarify provides feature-attribution explanations.
Fairness — SageMaker Clarify measures fairness and detects bias.
Guardrails — Guardrails block denied topics, harmful content, and can redact PII.
Human-in-the-Loop — Amazon Augmented AI (A2I) implements human-in-the-loop review.
Inclusivity — Inclusive design overlaps with fairness but focuses on access and representation.
Model Card — SageMaker Model Cards centralize this documentation for governance.
Model Monitor — Drift means production data has shifted from training data — Model Monitor flags it.
Responsible AI — AWS frames responsible AI around dimensions like fairness, safety, and transparency.
Robustness — Robustness includes resisting prompt injection and adversarial examples.
Safety — Bedrock Guardrails enforce safety by filtering harmful content.
Toxicity — Guardrails and content filters reduce toxic output.
Transparency — Model Cards are an AWS tool for documenting and communicating transparency.
Veracity — Hallucination is a veracity failure — RAG and grounding improve it.

Security, Compliance, and Governance for AI Solutions

Artifact — Artifact is where you download AWS's audit reports for your auditors.
Audit Manager — Audit Manager automates evidence collection against frameworks like SOC 2.
AWS Config — Config is about resource state and drift; CloudTrail is about API actions.
CloudTrail — CloudTrail answers 'who did what, when' — the backbone of auditing.
Data Governance — Good governance underpins responsible AI — you cannot govern AI without governing data.
Data Lineage — Lineage supports auditability and reproducibility of ML workflows.
Encryption — Use AWS KMS to manage keys for training data and model artifacts.
IAM — Apply least privilege to AI resources like Bedrock and SageMaker via IAM policies.
Inspector — Inspector finds software vulnerabilities; Macie finds sensitive data.
ISO — AWS ISO certifications are downloadable from AWS Artifact.
Macie — Reach for Macie to find PII in data before it is used for training.
OWASP Top 10 for LLMs — Prompt injection tops the OWASP LLM list — know it by name.
PrivateLink — Use PrivateLink to call Bedrock or SageMaker privately from your VPC.
SOC — SOC 2 reports are available through AWS Artifact.

Frequently asked questions

How many questions are scored, and how long do I get?: 65 questions in 90 minutes, but only 50 count toward your score. The other 15 are unscored questions mixed in and not marked, so treat every question as if it counts.
What score do I need to pass?: 700 on a 100–1,000 scaled-score range. It is a single pass/fail result using compensatory scoring, so a weak domain can be offset by strong ones — you only need to clear 700 overall.
Are there hands-on labs or performance-based questions?: No. AIF-C01 uses only multiple choice, multiple response, ordering, and matching questions. There are no console-based labs.
Do I need experience or prerequisites to take it?: There are no required prerequisites. AWS recommends up to 6 months of exposure to AI/ML on AWS and familiarity with core services like Amazon Bedrock and Amazon SageMaker AI, but anyone can register.
How much does it cost and how long is it valid?: USD $100 per attempt. The certification is valid for 3 years; you recertify by passing the latest version of the exam or by earning AWS Certified Machine Learning Engineer – Associate.
What changed recently — is the exam different now?: Yes. As of exam guide version 1.1 (published April 30, 2026, live on exams from roughly late May 2026), the exam adds significant agentic AI content (AI agents, MCP, multi-agent patterns, plus in-scope services like Amazon Bedrock AgentCore and Strands Agents), context engineering, token-based pricing, model distillation, LLM-as-a-judge evaluation, and hallucination-detection/grounding topics. The five domains and their weights are unchanged.