AWS Certified AI Practitioner (AIF-C01) Study Guide: Format, Domains, Cost, and How to Pass

The AWS Certified AI Practitioner (AIF-C01) is Amazon Web Services’ foundational-level certification for artificial intelligence and machine learning. The exam presents 65 questions — 50 scored and 15 unscored — in 90 minutes, and you pass with a scaled score of 700 on a scale of 100–1,000. It costs $100 USD and is delivered through Pearson VUE, either at a test center or online with a remote proctor. This guide covers who the certification is for, exactly what each of the five domains asks, how to register, and a four-week plan to get you from zero to a passing score.

What the AI Practitioner certification is and who it’s for

AIF-C01 sits at the foundational tier of the AWS certification ladder, alongside the Cloud Practitioner. It validates that you understand AI, machine learning (ML), and generative AI concepts and know which AWS services apply to which problems — without requiring you to build, train, or code anything. There are no prerequisites; AWS suggests up to six months of exposure to AI/ML workloads on AWS as helpful background, but plenty of candidates pass with focused study alone.

The audience is deliberately broad. Business analysts, project and product managers, sales and marketing professionals, IT support staff, and executives who need to make informed decisions about AI projects are the target — not ML engineers. If you can already train models in Amazon SageMaker, the Machine Learning Engineer Associate is the more natural fit; AIF-C01 is about conceptual fluency: knowing what a foundation model is, when fine-tuning beats Retrieval Augmented Generation (RAG), what responsible AI requires, and how AWS’s governance tooling fits together.

That conceptual framing shapes the whole exam. Questions rarely ask for syntax or console steps. They describe a business scenario — a retailer wants product descriptions generated, a bank worries about biased loan decisions — and ask which concept, service, or practice applies. Vocabulary discrimination is the core skill: the exam constantly asks you to tell apart adjacent terms like fine-tuning vs. RAG, precision vs. recall, explainability vs. transparency, and inference parameters like temperature vs. top-k.

The certification is valid for three years. Recertification means passing the current version of the exam again, or letting a higher-level AWS certification renew it.

Exam format at a glance

Detail	AIF-C01
Number of questions	65 (50 scored + 15 unscored pilot questions)
Time limit	90 minutes
Passing score	700 on a scaled range of 100–1,000
Exam fee	$100 USD
Question types	Multiple choice, multiple response, ordering, matching, case study
Delivery	Pearson VUE test center or online proctored (OnVUE)
Level	Foundational — no prerequisites
Validity	3 years

Three format notes worth internalizing. First, the 15 unscored questions are indistinguishable from scored ones — AWS uses them to pilot future content — so treat every question as if it counts. Second, AIF-C01 was one of the first AWS exams to introduce the newer interaction types: ordering (arrange steps in sequence), matching (pair items with definitions), and case studies (one scenario feeding several questions). Ordering questions love process sequences, so know your lifecycles cold. Third, the scaled 700 does not mean “70% correct” — scoring is compensatory across the whole exam, so a weak domain can be offset by a strong one, and there are no per-domain minimums.

The five domains and their weights

Domain	Weight	~Scored questions
1. Fundamentals of AI and ML	20%	~10
2. Fundamentals of Generative AI	24%	~12
3. Applications of Foundation Models	28%	~14
4. Guidelines for Responsible AI	14%	~7
5. Security, Compliance, and Governance for AI Solutions	14%	~7

Domains 2 and 3 together are more than half the exam. If your time is limited, generative AI concepts and foundation-model application decisions are where studying pays the highest return.

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

Domain 1 is the classical machine learning foundation everything else builds on. Expect to define AI vs. ML vs. deep learning as nested circles: AI is the broad goal of machines performing intelligent tasks, ML is the subset that learns patterns from data instead of following explicit rules, and deep learning is the subset of ML that uses multi-layer neural networks.

You need the three learning paradigms and their data requirements: supervised learning trains on labeled examples (classification and regression), unsupervised learning finds structure in unlabeled data (clustering, dimensionality reduction, anomaly detection), and reinforcement learning learns from rewards through trial and error. Tied to this is the difference between classification (predicting a category) and regression (predicting a number), and the inferencing split between real-time and batch.

Know the ML development lifecycle as a sequence — business problem framing, data collection and preparation, feature engineering, training, evaluation, deployment, monitoring — because ordering questions draw from it. Evaluation gets its own cluster of questions: confusion matrices and the metrics derived from them. The distinctions between accuracy, precision, recall, F1, and AUC — and which one fits a fraud-detection scenario vs. a spam filter — are among the most reliably tested facts in the whole exam.

Finally, map AWS’s managed AI services to their jobs: Amazon Rekognition for image and video analysis, Amazon Textract for extracting text from documents, Amazon Comprehend for natural language processing, Amazon Transcribe for speech-to-text, Amazon Polly for text-to-speech, Amazon Translate for language translation, Amazon Kendra for intelligent enterprise search, Amazon Personalize for recommendations, and Amazon SageMaker as the build-train-deploy platform for custom models. A large share of Domain 1 questions are pure service matching: scenario in, service name out.

Domain 2: Fundamentals of Generative AI (24%)

Domain 2 shifts from classical ML to generative AI vocabulary and business framing. Start with the core concepts: tokens (the chunks of text a model reads and writes), embeddings (numeric vectors that capture meaning), prompts, inference, context window, and hallucination (confident but fabricated output).

The centerpiece is understanding what a foundation model is — a large model pre-trained on broad data that can be adapted to many downstream tasks — and why that pre-train-once-adapt-many pattern changed the economics of AI. Follow that with the foundation model lifecycle in order: data selection, pre-training, optional fine-tuning, evaluation, deployment, and ongoing feedback — another favorite source of ordering questions. Round out the model-type vocabulary with multimodal models, which accept or produce more than one data type (an image plus a prompt in, text out), and diffusion models, which generate images by iteratively removing noise.

Two business-judgment threads run through this domain. One is generative AI vs. traditional ML: when a generative model’s adaptability and content creation win, and when a cheaper, more predictable traditional classifier is the right call — including generative AI’s inherent drawbacks like hallucination, nondeterminism, and difficulty of interpretation. The other is measuring generative AI business value: tying projects to metrics like conversion rate, average revenue per user (ARPU), customer lifetime value (CLV), and efficiency gains, rather than model-quality metrics alone.

On the services side, know the generative AI stack: Amazon Bedrock (managed API access to foundation models from multiple providers), Amazon Q Business and Amazon Q Developer (generative AI assistants), Amazon SageMaker JumpStart (deployable pre-trained models), and PartyRock as the no-code Bedrock playground.

Domain 3: Applications of Foundation Models (28%)

Domain 3 is the largest slice of the exam and the most decision-heavy: given a business need, pick the right way to apply a foundation model.

The first decision is model selection. Choosing a foundation model means weighing cost, latency, model size, supported input and output modalities, supported languages, context window length, and licensing — and the exam loves scenarios where one constraint (say, a 200-page document that must fit in context) dominates the choice.

The second decision is the adaptation ladder, ordered from cheapest to most expensive: prompt engineering, then RAG, then fine-tuning, then training a model from scratch. Prompt engineering covers zero-shot, one-shot, and few-shot prompting, prompt templates, and negative prompts, plus inference parameters — temperature and top-p control randomness, and lower values make output more deterministic. RAG retrieves relevant documents (typically via embeddings in a vector database, with OpenSearch as an AWS example, orchestrated by Knowledge Bases for Amazon Bedrock) and stuffs them into the prompt — the right answer whenever knowledge changes frequently or must be grounded in company data without retraining. Fine-tuning actually updates model weights on domain data; preparing data for fine-tuning — curation, representativeness, sufficient size, high-quality labeled pairs — is tested on its own.

Three more concepts complete the domain. Transfer learning and catastrophic forgetting explain why adapting a pre-trained model is efficient and what can go wrong when new training overwrites old knowledge. RLHF — Reinforcement Learning from Human Feedback — is the human-preference alignment technique, tested as a step sequence: collect human preference rankings, train a reward model, optimize the model against it. And prompt-level attacks — injection, jailbreaking/hijacking, poisoning, and prompt exposure — appear here as application risks, with Guardrails for Amazon Bedrock as the mitigation service. Expect a few questions on evaluating foundation-model output too: ROUGE for summarization, BLEU for translation, BERTScore for semantic similarity, plus human evaluation and benchmark datasets.

Agents get lighter coverage: know that Agents for Amazon Bedrock can break a task into steps and call APIs and knowledge bases to complete multi-step business workflows.

Domain 4: Guidelines for Responsible AI (14%)

Domain 4 asks whether AI systems are fair, explainable, safe, and legally defensible. Its anchor topic is AI bias and fairness: where bias enters (unrepresentative training data, historical prejudice baked into labels, proxy features), what it does to model decisions, and how to counter it with diverse datasets, subgroup evaluation, and human oversight. Amazon SageMaker Clarify is AWS’s tool for detecting bias and explaining predictions; Amazon Augmented AI (A2I) adds human review loops for low-confidence predictions.

Explainability is its own cluster. Explainable AI covers transparency (openness about how a system was built and trained, including training-data disclosure) and interpretability (understanding why an individual prediction happened), while the black-box vs. interpretable model tradeoff captures the exam’s favorite tension: complex models like deep neural networks tend to be more accurate but harder to explain, while linear models and decision trees are transparent but sometimes weaker — and regulated decisions (credit, hiring, healthcare) tilt the choice toward interpretability. Documentation tooling belongs here too: SageMaker Model Cards for recording a model’s intended use, risk rating, and evaluation results, and AWS AI Service Cards as AWS’s own transparency documentation for its services.

Model-quality failure modes appear under the responsible AI banner as well. Overfitting vs. underfitting — diagnosing each from the gap between training accuracy and validation accuracy — is tested with small numeric scenarios (high training accuracy plus low validation accuracy means overfitting).

The legal thread rounds it out: generative AI and copyright (infringement exposure from both training data and generated output) and the broader legal risks of generative AI — liability for hallucinated claims, harmful advice delivered to customers, and compliance exposure when unvetted output reaches the public. Guardrails for Amazon Bedrock (content filtering, denied topics, personally identifiable information redaction) is the recurring control answer.

Domain 5: Security, Compliance, and Governance for AI Solutions (14%)

Domain 5 applies AWS security and governance fundamentals to AI workloads. On the security side, know the standard toolkit: AWS Identity and Access Management (IAM) roles and least-privilege policies for controlling who can invoke models, encryption at rest and in transit with AWS Key Management Service (KMS), Amazon Macie for discovering sensitive data in Amazon S3, and the shared responsibility model — AWS secures the infrastructure, you secure your data, access, and configurations.

Data governance carries several question clusters. Data lineage — tracking where training data came from, its licensing, and who accessed and transformed it from ingestion through training — supports both compliance audits and reproducibility. Privacy-enhancing technologies such as differential privacy, anonymization, and data masking protect personal data used in AI systems, each with a utility-vs-privacy tradeoff you should be able to articulate.

Compliance tooling is heavily service-matched: AWS Artifact for downloading AWS’s own compliance reports, AWS Audit Manager for continuously collecting evidence against frameworks, AWS Config for tracking resource configuration compliance, and AWS CloudTrail for recording API activity. Distinguishing Artifact (AWS’s compliance documents) from Audit Manager (your audit evidence) is a classic exam discrimination.

Finally, the regulatory and framework layer: the EU AI Act and algorithmic accountability laws with their risk-tier approach, bias audits, and impact assessments, and the NIST AI Risk Management Framework with its four functions — Govern, Map, Measure, Manage — plus ongoing governance practices like review cadences and role definitions. Neither is an AWS product, and the exam tests them as concepts you’d align an AI program to.

How to register

Registration runs through your AWS Certification account: sign in at the AWS Training and Certification portal, choose the AWS Certified AI Practitioner exam, and schedule through Pearson VUE. You pick between a physical test center and online proctoring (OnVUE) at scheduling time. The fee is $100 USD, paid when you book.

Two logistics worth knowing. If English is not your first language, request the “ESL +30” accommodation before scheduling — it adds 30 minutes to your exam time and costs nothing. And if you test online, run Pearson VUE’s system check on the same computer and network you’ll use on exam day, clear your desk completely, and expect a room scan; a phone must be nearby but out of reach for check-in.

If you fail, you must wait 14 days before rebooking, with no limit on total attempts (each attempt costs the full fee).

A 4-week study plan

This plan assumes roughly one hour on weekdays and a longer weekend block. Adjust the ratio, not the order — it follows the exam’s weighting.

Week 1 — Domains 1 and 2 (fundamentals). Nail the definitions: AI/ML/deep learning, supervised vs. unsupervised vs. reinforcement learning, classification vs. regression, the ML lifecycle, and the confusion-matrix metrics. Then move to generative AI vocabulary — tokens, embeddings, context windows, hallucination, foundation models, multimodal and diffusion models. Build a two-column service map (service → job) for the managed AI services and the generative stack (Bedrock, Amazon Q, JumpStart). End the week with a timed quiz over both domains and log every miss.

Week 2 — Domain 3 (the big one). Spend the full week on foundation-model application: the adaptation ladder (prompt engineering → RAG → fine-tuning → train from scratch) and the cost logic behind it, inference parameters, RAG architecture with vector databases and Knowledge Bases, fine-tuning data preparation, transfer learning, RLHF’s step sequence, output-evaluation metrics (ROUGE, BLEU), and prompt attacks with Guardrails as the control. Practice with scenario questions daily — this domain is decided by scenario reading, not memorization.

Week 3 — Domains 4 and 5 (responsibility and governance). Cover bias sources and mitigations, explainability vs. interpretability, the black-box tradeoff, overfitting diagnosis, and the copyright/legal risk landscape, then switch to security and governance: IAM, encryption, Macie, the compliance tool quartet (Artifact, Audit Manager, Config, CloudTrail), data lineage, PETs, the EU AI Act, and the NIST AI RMF functions. These two domains are only 28% combined but are dense with service-name discriminations — flashcards work well here.

Week 4 — Full practice exams and gap repair. Take at least two full-length, 65-question timed practice exams under real conditions. Score each by domain, then spend your remaining sessions exclusively on your two weakest objective areas. Re-test until you’re comfortably clearing 80% on fresh questions — a margin above the roughly 70% scaled bar. The final two days, review your accumulated miss log rather than new material.

How the AIF-C01 exam tests this

• Service matching. A one-sentence business need maps to exactly one AWS service (“extract text from scanned invoices” → Amazon Textract). Distractors are neighboring services, so learn services by job, not by description.
• Cheapest-adequate-option scenarios. Several options would technically work (fine-tuning, RAG, prompt engineering); the correct answer is the least costly approach that meets the stated requirement. Watch for signal phrases like “minimal operational overhead” or “without retraining.”
• Term discrimination. Two adjacent concepts appear as answer choices — precision vs. recall, transparency vs. interpretability, Artifact vs. Audit Manager — and the scenario contains one detail that settles it.
• Process ordering. Lifecycles (ML development, foundation model lifecycle, RLHF) appear as arrange-the-steps items. If you can’t recite the sequence, you can’t partial-credit your way through.

Exam-day tips

• Budget about 80 seconds per question; flag anything slow and return. Case-study sets share one scenario, so read it once carefully and answer all attached questions before moving on.
• Read the last sentence of long scenarios first — it contains the actual question and tells you which details matter.
• In multiple-response items, the question states exactly how many options to pick; partial credit doesn’t exist, so treat each selection as its own true/false check.
• Eliminate answers that solve a different problem than the one asked — most distractors are real facts attached to the wrong scenario.
• Don’t leave blanks. There’s no penalty for guessing, and compensatory scoring means every scored question you salvage matters.
• Results aren’t shown on screen at completion; your pass/fail and score arrive in your certification account, typically within a few days.

Quick reference

• 65 questions (50 scored, 15 unscored), 90 minutes, pass at 700 on a 100–1,000 scale, $100 USD.
• Delivered by Pearson VUE — test center or online proctored; ESL +30 accommodation available if requested before scheduling.
• Five domains: AI/ML fundamentals 20%, generative AI fundamentals 24%, applications of foundation models 28%, responsible AI 14%, security/compliance/governance 14%.
• Domains 2 + 3 are 52% of the exam — foundation-model concepts and application decisions are the highest-yield study areas.
• Question types include multiple choice, multiple response, ordering, matching, and case studies.
• Core skill: discriminating adjacent terms and matching AWS services to one-line business needs.
• Certification is valid for 3 years; retakes require a 14-day wait.