If your AIF-C01 study material predates April 30, 2026, it is missing objectives the current exam tests. The exam guide was revised to version 1.1 on that date, which added objectives on agentic AI, context engineering, and token-based pricing, and AWS publishes guide updates about a month before they appear on the exam. Plan from the current guide's five domains, take the free assessment first so your misses point at a domain, and give the two generative AI domains the most time, since they carry 52% of the scored content.
What the exam measures, and what it does not test
The official AIF-C01 exam guide, version 1.1 (published April 30, 2026), defines five content domains as shares of scored content:
| Domain | Weight | About, of the 50 scored questions | The decision it tests |
|---|---|---|---|
| Fundamentals of AI and ML | 20% | 10 | Label what the system outputs |
| Fundamentals of GenAI | 24% | 12 | Match the capability to the inputs |
| Applications of Foundation Models | 28% | 14 | Pick the fix that supplies what the model lacks |
| Guidelines for Responsible AI | 14% | 7 | Match the concern to the control |
| Security, compliance, and governance for AI solutions | 14% | 7 | Sort controls by what they govern |
The question counts are derived from the weights and the 50 scored questions; AWS does not publish per-domain counts. The revision history lists version 1.0 (March 26, 2026) and version 1.1 (April 30, 2026). The 1.1 revision added the agentic AI, context engineering, and token-based pricing objectives and put Amazon Bedrock AgentCore, Kiro, Strands Agents, Amazon Q, and SageMaker JumpStart on the in-scope list. If your course or notes predate that revision, check them against those additions before you trust them.
The exam has 65 questions (50 scored, 15 unscored and not identified) in 90 minutes, about 83 seconds per question. The AWS certification page lists the 90-minute duration, the 65-question format, and the USD 100 cost, and the certification is valid for 3 years. Your result is a scaled score from 100 to 1,000 with 700 to pass, on a compensatory model. You do not need to pass each section; you need to pass the exam as a whole. Besides multiple choice and multiple response, the exam can include ordering questions (3 to 5 responses placed in order) and matching questions (responses matched to 3 to 7 prompts), and the complex types need every part right to earn credit. Unanswered questions count as incorrect, and there is no penalty for guessing.
The guide lists these tasks as out of scope for the exam: coding AI/ML models or algorithms, data engineering or feature engineering, hyperparameter tuning or model optimization, and building or deploying AI/ML pipelines. The target candidate has up to 6 months of exposure to AI/ML on AWS, and the guide describes that candidate as someone who uses but does not necessarily build AI/ML solutions. One line from the guide also applies to your practice. "Use caution when you interpret section-level feedback."
What it costs
The exam costs USD 100, delivered at a Pearson VUE test center or as an online proctored exam. Before it expires, you can recertify by passing the AWS Certified Machine Learning Engineer - Associate instead of retaking this exam. Re-check the exam pricing page that the certification page links to, for exchange rates and current offers, before you book.
The free assessment decides where you start
Take the free AWS AIF-C01 assessment before you study any domain. It is free, needs no credit card, and returns a Readiness Report with your readiness score, your top gaps, and a targeted repair preview. Answer it without looking anything up, then sort every miss into one of two kinds. The first kind is a term you did not know. The second kind is a term you knew, matched to the wrong approach. The second kind has recognizable patterns on this exam:
- Labeling a prediction task as generative AI when the system only produces a score for an existing workflow.
- Reaching for retraining when the model just needed the missing context.
- Answering a fairness concern with a security control, or documenting a problem the team has only just noticed.
Write down which pattern each miss had. That list, checked against your top gaps, is your study list.
Four weeks, weighted to the two generative AI domains
This plan assumes about 5 hours a week.
- Week 1: diagnostic, then Fundamentals of AI and ML (20%). Done when you can label a system from its output, and name a case where a fixed business rule, not a prediction, is required.
- Week 2: Fundamentals of GenAI (24%). Done when you can match a capability to a list of input types, name what generative AI does badly, and say how token-based pricing makes long context cost more.
- Week 3: Applications of Foundation Models (28%). Done when you can choose between a prompt change, retrieval augmented generation (RAG), and fine-tuning for a stated need, and name how you would evaluate the result.
- Week 4: Responsible AI (14%) and Security, Compliance, and Governance (14%), then mixed retests. Done when you can pick the control that addresses a stated fairness, access, or residency concern, and the miss patterns from your diagnostic do not show up on fresh questions.
If your diagnostic shows Applications of Foundation Models is your weakest area, start it in week 2 alongside the fundamentals. It is the largest domain, and its vocabulary (prompt, retrieval, fine-tuning) is what the later domains build on.
Fundamentals of AI and ML (20%): label the system by what it outputs
The Domain 1 outline covers terminology, practical use cases, and the AI/ML development lifecycle. The decision the exam makes you take is a label. Ask what the system outputs:
- A score, a class, or a forecast learned from historical examples is machine learning.
- New text, images, or code is generative AI.
- A system that plans steps, calls tools, and coordinates toward a goal is agentic AI.
The agentic label is new in version 1.1, and it is where the mislabels happen. The cue is not "automation". The objective names tool usage, memory management, and workflow orchestration toward a goal. A system that runs a fixed workflow and produces a number for someone else's process is not planning or acting. It is a prediction.
Two more boundaries come from the same domain. Task statement 1.2 lists "situations when a specific outcome is needed instead of a prediction" as cases where AI/ML is not appropriate, so a stem that requires a fixed business rule to be followed is a cue to step back from ML entirely. Objective 1.2.6, also added in version 1.1, asks you to pick between traditional ML models and foundation models based on regulatory concerns, explainability requirements, or operational constraints. When the stem names one of those, the answer is the traditional model, not the foundation model.
For hands-on practice, map the lifecycle stages of task statement 1.3 to the services that objective names (Amazon Bedrock, Amazon Quick, Kiro, and SageMaker AI), and work through one managed service, such as Amazon Comprehend, on your own text.
Fundamentals of GenAI (24%): match the capability to the inputs
The Domain 2 outline covers generative AI concepts, what generative AI does well and badly, and the AWS technologies for building these applications. The decision the exam makes you take is capability matched to input types. List the inputs in the stem first, then check that the capability you choose handles all of them. When a stem mixes a written note with a photo and asks for one judgment across both, a text-only model cannot make that judgment, and the outline names multi-modal models for exactly this.
The outline also names what generative AI does badly. Hallucinations, inaccuracy, nondeterminism, and poor interpretability. When the stem requires output that must not invent, or behavior that must be deterministic, plain generative AI is the wrong tool on its own, and the stem points at grounding, evaluation, or a non-ML answer.
Version 1.1 added two cost-and-context objectives here. Objective 2.1.4 covers the token-based pricing model and its effect on inference cost, so a question asking what makes a generative AI workload more expensive is pointing at token counts, including the context you send. Objective 2.1.5 covers the role of context engineering in foundation model applications. What you put in the context is a design decision, and the exam can test it.
Applications of Foundation Models (28%): supply what the model is missing before you change the model
The Domain 3 outline is the largest, and its core decision is choosing between prompt engineering, RAG, and fine-tuning, with in-context learning and model distillation as the other customization approaches the outline names. Start from what the model is missing, and pick the cheapest fix that supplies it:
- The model lacks current or private facts that exist in a document. Use RAG, which retrieves the relevant content at request time. On AWS that is Amazon Bedrock Knowledge Bases, named in objective 3.1.
- The model lacks a style, format, or framing. Use prompt engineering. The outline names context, instruction, and negative prompts as the concepts, and zero-shot, few-shot, and chain-of-thought as the techniques.
- The model needs a persistent domain behavior that does not fit in the context, or you must cut latency or cost at scale. Use fine-tuning or continued pre-training, and know the cost tradeoffs that objective 3.1.5 asks you to explain.
The tempting alternatives fail for specific reasons, and the exam tests that:
- Fine-tuning bakes data into the model's parameters, so it does not fetch new source files. For documents that change weekly, or that are customer-specific, RAG is the fit, because the document is retrieved at request time. The stem cue is change, privacy, or a citation requirement.
- A prompt change fixes tone and layout, not facts. If the stem says the relevant document is not part of the request and the answers vary and cite assumptions, rewording the instruction does not give the model the policy text. The fix is to include the document, as context or through RAG.
- Inference parameters such as temperature change the style and variety of the response, not the facts the model has. Objective 3.1 names temperature and input/output length as inference parameters. If the stem asks why answers are too creative, the parameter is the answer. If it asks why answers miss internal facts, the answer is context or retrieval, not a parameter.
Evaluation sits in the same domain (task statement 3.4). The outline names human-in-the-loop evaluation, benchmark datasets, Amazon Bedrock Model Evaluation, ROUGE, BLEU, BERTScore, and LLM-as-a-judge as approaches and metrics, plus business metrics such as task completion rate and cost per interaction. When a stem asks how you would verify that a model meets the business objective, pick the evaluation approach, not a new customization.
Guidelines for Responsible AI (14%): measure the harm before you document it
The Domain 4 outline covers responsible AI features and transparent, explainable models. The decision is matching the named concern to the right control, because the outline lists several controls that all sound responsible:
- Uneven quality across customer groups. Measure first. Objective 4.1 names subgroup analysis, label quality analysis, and human audits as tools to detect and monitor bias.
- Unsafe or off-policy content. Amazon Bedrock Guardrails, which the outline names as a tool for responsible AI features.
- Bias detection and explanation for a specific model. In version 1.1, the revision history adds SageMaker Clarify to objective 4.2.2.
- Documenting known behavior. SageMaker Model Cards. A model card records intended use and known evaluation results, so documentation comes after the evidence exists.
The trap is the model card. It belongs to responsible AI, so it tempts you in every fairness stem. But when the stem says the team has only "noticed" uneven reliability, the team needs the measurement, not the documentation.
The outline also puts the legal risks of generative AI in this domain. Intellectual property infringement claims, biased model outputs, loss of customer trust, and hallucinations. A stem that asks what legal or trust risk a generative AI feature creates is a domain 4 question even when the feature itself lives elsewhere.
Security, compliance, and governance (14%): sort controls by what they govern
The Domain 5 outline covers securing AI systems and governance and compliance. Governance stems often list four sensible controls and ask which one addresses the stated concern. Sort them by what each governs:
- Geography (residency). The AWS Region where the workload's data is handled.
- People (access). IAM roles and permissions.
- Time (retention). The data lifecycle rules for how long data is kept and when it is deleted.
- Confidentiality. Encryption at rest and in transit, named in the outline's security considerations.
The stem names one of the four dimensions. Pick the control that governs it, and leave the other three alone.
One wrinkle is Amazon Bedrock cross-Region inference. A cross-Region inference profile can route a request to a Region other than the one you called. AWS's documentation recommends a geographic cross-Region inference profile "when you have data residency requirements and need to ensure data processing remains within specific geographic boundaries" (cross-Region inference). Choosing the Region is the residency decision; the geographic profile is what keeps inference inside it. The same documentation notes that CloudTrail logs all cross-Region inference requests in your source Region, which is the audit trail for where the work actually ran.
The outline's governance objective names the Generative AI Security Scoping Matrix as a governance framework. The matrix sorts use cases into scopes 1 to 5 by how much your organization owns the AI model and its data, from consuming a public third-party model to training your own.
How to know a weak area is fixed
Reading a definition after a miss does not prove you can pick the right approach on a fresh stem. At the end of each week, retest that domain with questions you have not seen, and check whether the miss patterns you recorded still show up when the wording changes. A domain counts as fixed when a fresh question in that domain no longer triggers the pattern you logged for it.
When you need a retest you have not already seen, the full AI Practitioner Pass Plan is the paid CramHQ path. It costs US$34.99 for 12 months of access, and it includes the five practice tests and the timed simulation (AWS AIF-C01 course page).
In the final week, decide on fresh questions
In the final days, take a fresh, mixed, timed practice set, and try the official resources the AWS certification page links to. Its Exam Prep Plan on AWS Skill Builder includes the AWS Certification Official Practice Question Set, the Official Pretest, and the Official Practice Exam. Count only questions you had not seen before.
Apply the guide's caution to your own practice. One low domain score matters less than a miss pattern that shows up across domains. If your diagnostic patterns are gone on a fresh mixed set, book the exam. If one is still showing up, spend a week on that domain and re-run this decision on another fresh set.