AIF-C01 Exam Guide and AI/ML Fundamentals

Every AWS certification starts with a document: the official exam guide. For AIF-C01 that document does two jobs at once; it tells you exactly how the exam is built (codes, minutes, questions, cut score, weights) and it defines the conceptual vocabulary you must command in Domain 1 (Fundamentals of AI and ML, 20% of the score). Students who skip the guide and jump straight to flashcards usually fail for the same three reasons: they misjudge the time budget, they confuse statistical bias with fairness bias, and they cannot tell when AI is the wrong answer to a business problem.

In this lesson you will:

• read the exam guide the way an exam writer does: code, timing, cost, scoring, languages;
• convert domain weights into a concrete study-hour budget;
• master the four question types and the compensatory scoring model;
• separate AI, ML, deep learning and generative AI using AWS's own definitions;
• choose between supervised, unsupervised, reinforcement and semi-supervised learning;
• lock in the vocabulary: model, algorithm, training, inference, epoch, batch, feature, label;
• diagnose underfitting and overfitting through the bias/variance lens;
• compute and interpret accuracy, precision, recall and F1;
• map the six-phase ML lifecycle to real AWS services;
• absorb the 2025 to 2026 exam-guide updates (v1.1 objectives) and separate verified from non-verified claims;
• apply the exam strategy essentials: pacing, elimination and the recurring trap families;
• study real-world AWS case studies: services, numbers and sources;
• practice with 12 exam-style questions plus interactive checks.

---

1. What AIF-C01 actually measures

1.1 The target candidate

The exam guide defines the target candidate in one sentence, and that sentence is exam-relevant material in its own right: a candidate has up to 6 months of exposure to AI/ML technologies on AWS and uses but does not necessarily build AI/ML solutions. This is a practitioner credential, not an engineer credential. You are being tested on judgment, vocabulary, service selection and governance: not on writing training code.

1.2 In scope and explicitly out of scope

The guide draws a hard line around scope. The out-of-scope list is not decoration: items from it have appeared as distractors on the official practice question set, because a distractor that sounds like "real ML work" is the most tempting wrong answer of all.

Explicitly out of scope (exam guide):

• coding AI/ML models;
• data engineering and feature engineering;
• hyperparameter tuning;
• building ML pipelines;
• mathematical and statistical analysis;
• implementing security and compliance protocols;
• developing governance frameworks.

Explicitly in scope (recommended AWS knowledge):

• core cloud services: Amazon EC2, Amazon S3, AWS Lambda;
• AI services: Amazon Bedrock, Amazon SageMaker AI;
• shared responsibility model and AWS Identity and Access Management (IAM);
• AWS pricing and cost-management fundamentals.

1.3 The AWS service surface you must recognize

The guide publishes a non-exhaustive service list, grouped by category. You are not required to know every console path; you are required to know what each service is for and when to choose it.

• 📚 Did you know? The AIF-C01 exam was announced in June 2024 and its beta opened on 13 August 2024 with 85 questions, 170 minutes and a 75 USD fee in English and Japanese. The standard exam went GA in October 2024 at 100 USD with 65 questions and 90 minutes, so if you find an old study guide describing a 170-minute exam, you are reading beta-era material.

---

2. Exam logistics: the numbers that change your strategy

2.1 The logistics table

Every row below is verified against the AWS certification page, the exam guide or the Pearson VUE store, with the verification year shown.

2.2 Example E1: the per-question time budget

The single most useful number in the whole guide is arithmetic nobody publishes:

$$t = \frac{90 \times 60}{65} = \frac{5{,}400}{65} \approx 83 \text{ seconds per question}$$

If you reserve 10 minutes for review, the real budget is $80 \times 60 / 65 \approx$ 73.8 seconds per item. Compare this with the 2024 beta: $170 \times 60 / 85 = 120$ seconds per item: the standard exam is roughly 31% tighter per question. Ordering and matching items legitimately consume more than average; multiple-choice items should be finished in under a minute so the surplus pays for them.

2.3 Example E5: cost of failure

A failed attempt costs another 100 USD plus a mandatory 14-calendar-day wait before you can retake. There is no limit on attempts and the full fee applies each time; conversely, you cannot retake a exam you have passed for 2 years. A limited "Foundational Retake" promotion ended on 15 February 2025, so budget for one shot.

• 📚 Did you know? You do not need to "pass every domain". Because scoring is compensatory, a strong Domain 3 (28%) can carry a weak Domain 4 (14%): AWS states you need a passing score only on the overall exam, and the score report may still show section-level feedback to guide your retry.

---

3. Domains, weights and task statements

3.1 Official weights and the derived question budget

3.2 What each domain really tests

3.3 Turning weights into a study plan

Suppose you have 40 hours before exam day. The weights convert directly into hours:

Domains 2 + 3 together are 52% of the score: more than half the exam is generative-AI work. A candidate who spends 80% of study time on classic ML has already misallocated the majority of their effort.

---

4. Question types and scoring mechanics

4.1 The four question types

There is no case-study type listed in the current guide. A legacy 2024 PDF (v1.4, 31 July 2024) contains case-study wording, and AWS does not state which numbering series governs exam day: see the caveats in Section 11.3.

4.2 Compensatory scoring

Compensatory means one thing: only the overall result must clear 700. You do not need to reach 700 inside each individual domain. A candidate who scores 650-worth in Domain 4 but 800-worth overall-equivalent in Domains 1 to 3 can still pass, and vice versa: dominating Domain 3 cannot rescue an overall score below the cut.

4.3 Example E2: scaled, not a raw percentage

700 out of 1,000 looks like 70%, and the naive conversion $0.70 \times 50 = 35$ correct answers is arithmetically clean, and wrong as an exam fact. AWS scales scores "across multiple exam forms that might have slightly different difficulty levels", so the raw number of correct answers required for 700 is not fixed at 35 and is not published. The only deterministic rule you can rely on is behavioural: unanswered items are scored incorrect, so you must answer all 65.

---

5. AI, machine learning, deep learning and generative AI

5.1 The four-layer map, in AWS's own words

5.2 When AI/ML is the wrong answer

Domain 1 explicitly tests cost: benefit analysis and the decision not to use AI. The two canonical reasons to decline an AI project:

1. You need a specific, deterministic outcome: if the requirement is "apply exactly these 12 business rules", a rules engine gives the same answer every time, is auditable line by line and costs nothing to train.
2. The cost, benefit does not close, data collection, labeling, evaluation and monitoring can exceed the value of the prediction, especially where a spreadsheet or a human review already achieves the target accuracy.

• 📚 Did you know? AWS's own "what is overfitting" page uses a definition that surprises people: "Underfit models experience high bias, they give inaccurate results for both the training data and test set." Bias in Domain 1 is a statistical term (high bias = too simple) and Domain 4 uses bias as a fairness term (demographic skew). The exam tests both senses, usually in different domains.

---

6. Learning paradigms: supervised, unsupervised, reinforcement, semi-supervised

6.1 The decision table

AWS documents the three paradigms precisely: supervised learning uses features plus target values (categorical target → classification; continuous target → regression); unsupervised learning has no labels and discovers groupings (PCA, clustering); reinforcement learning has an agent learn by trial and error to maximize long-term reward, balancing exploration against exploitation.

6.2 Choosing the paradigm from the question stem

---

7. Core terminology: the vocabulary of Domain 1

7.1 Terms with AWS-sourced definitions

7.2 Example E6: one epoch, one batch, 1,000 rows

Take 1,000 rows with a batch size of 100: one epoch = 10 training iterations, each consuming 100 samples. Ten epochs = 100 iterations total, and the model has now seen every row ten times. More epochs are not automatically better; they reduce bias up to a point and then start raising variance (memorization). AWS guidance for batch size is unusually practical: start at 1 and increase until you hit an out-of-memory error, because larger batches raise OOM risk on the training instance.

7.3 Inference modes

7.4 Data types you must recognize

• Labeled vs unlabeled: determines the paradigm before anything else;
• Structured/tabular: rows and columns; classic ML territory (XGBoost on SageMaker);
• Time-series: ordered observations; forecasting tasks;
• Images: pixels; computer vision (Rekognition);
• Text: unstructured tokens; NLP (Comprehend) and generative AI;
• Structured vs unstructured: the exam uses this pair to decide between SageMaker tabular jobs and Bedrock/Comprehend/Rekognition.

---

8. Model fit: underfitting, overfitting, bias and variance

8.1 The diagnosis table

8.2 Example E3: a 1,000-row split

Legacy default split: 700 training / 300 evaluation rows.

The gap tells you the diagnosis; the absolute level tells you which direction to move. Scenario B is the one students miss: a small gap is not automatically good news when both numbers are bad.

8.3 Bias and variance in two different senses

The exam guide Task 4.1 lists "effects of bias and variance (e.g., effects on demographic groups, inaccuracy, overfitting, underfitting)": deliberately joining the two senses:

• Statistical sense (Domains 1 and 3): high bias = underfit, high variance = overfit; more training reduces bias but can increase variance.
• Fairness sense (Domain 4): bias as skew against demographic groups, addressed through label-quality analysis, human audits and subgroup analysis, with SageMaker Model Cards documenting the model's intended use and limitations.

---

9. Evaluation metrics: accuracy, precision, recall and F1

9.1 The metric set

$$F1 = 2 \cdot \frac{P \times R}{P + R}$$

9.2 Example E4: why F1 is a harmonic mean

With precision $P = 0.80$ and recall $R = 0.50$:

$$F1 = \frac{2 \times 0.80 \times 0.50}{0.80 + 0.50} = \frac{0.80}{1.30} \approx 0.615$$

The arithmetic mean would be $(0.80 + 0.50)/2 = 0.650$. F1 lands lower (0.615) precisely because the harmonic mean is pulled toward the weaker metric; that is the point of the metric. A model that finds only half the relevant items should not be graded as "mostly fine".

9.3 Model metrics versus business metrics

Domain 1 requires you to keep two ledgers separate:

A fraud model at 99% accuracy that blocks 5% of legitimate payments can destroy more value than it saves: the exam expects you to say so.

• 📚 Did you know? F1 is a harmonic mean rather than an arithmetic mean for a reason that shows up in exam options: with P = 0.95 and R = 0.30 the arithmetic mean says 0.625 (passable) while F1 says 0.456 (failing). The harmonic mean refuses to let one excellent number hide a broken one: AWS documents the range as 0 (worst) to 1 (highest).

---

10. The AI/ML development lifecycle

10.1 The six phases (AWS Well-Architected ML Lens)

The ML Lens defines six phases, and warns that they are "not necessarily sequential", with feedback loops running in both directions:

10.2 Phase-to-service map

10.3 Example E7: from business goal to monitored endpoint

A retail team wants to reduce support spend by 15%. Mapping that to the lifecycle:

1. Business goal: cut support cost by 15%: measured in dollars, not accuracy.
2. Problem framing: predict ticket auto-resolution (binary classification), or decide a Bedrock summarizer is cheaper than a classifier.
3. Data: 2 years of tickets in S3, cleaned with Glue DataBrew, labeled by 3 annotators in Ground Truth (target: ≥90% inter-annotator agreement).
4. Development: train in SageMaker AI, evaluate F1 = 0.74 on a held-out 30% split.
5. Deployment: SageMaker real-time endpoint for the agent console; Bedrock API for summarization.
6. Monitoring: CloudWatch alarms on latency and Model Monitor alerts on feature drift; Model Card records intended use and known limits.

---

11. Preparation roadmap, resources and caveats

11.1 Official Skill Builder assets

11.2 The official four-step plan

AWS publishes a four-step prep plan, and the order matters; it is diagnostic before it is instructional:

1. Get to know the exam: read the exam guide and this lesson; understand types, weights and scoring.
2. Refresh AWS knowledge and skills: EC2, S3, Lambda, IAM, Bedrock, SageMaker AI, shared responsibility, pricing.
3. Review and practice: question banks, domain review videos, flashcards.
4. Assess readiness: timed official pretest or practice exam at 65 questions / 90 minutes.

11.3 Recertification, retakes and unverified claims

• Validity: 3 years. Recertify by re-passing AIF-C01 or by earning AWS Certified Machine Learning Engineer: Associate, which auto-recertifies.
• Retake: fail → 14 calendar days wait, unlimited attempts, full 100 USD each time; a passed exam cannot be retaken for 2 years.
• Next certifications: Solutions Architect: Associate (cloud path) or Data Engineer: Associate / ML Engineer: Associate (data and AI path).

• 📚 Did you know? The exam guide's change history and the certification page's language list are both moving targets: the current guide series shows v1.0 on 26 March 2026 and v1.1 on 30 April 2026, while Italian and German retire after 15 October 2026, leaving the exam in 12 languages. Check the guide version date on any PDF you download: undated or 2024-dated PDFs circulate widely.

11.4 2025 to 2026 Updates

The exam guide is a living document, and the last two years changed it twice in one month: v1.0 was published on 26 March 2026 and v1.1 on 30 April 2026. AWS states that guide updates appear on the exam about one month after publication, so every v1.1 objective below has been exam-eligible since late May 2026. Everything in this subsection marked verified was checked against the AWS exam-guide change history (docs v1.1) on 6 October 2026; anything not so marked is labelled non-verified.

The seven new objectives in v1.1 (verified):

Other verified v1.1 changes:

• Service list: added Amazon Aurora, Bedrock AgentCore, Kiro, Strands Agents, Amazon Q, SageMaker JumpStart, AWS Transform; removed Amazon MemoryDB;
• Changed examples: 1.1.3 adds asynchronous and serverless inference; 1.1.1 says "agentic AI"; 1.3.4 names Bedrock, Amazon Quick, Kiro, SageMaker AI; 1.3.6 keeps accuracy/precision/recall/F1 and drops AUC; 3.1.5 adds distillation; 3.4.2 adds LLM-as-a-judge; 4.2.2 names Clarify + Bedrock Model Evaluations; 5.1.4 adds data-leakage prevention, output validation, audit trails and toxicity;
• Unchanged: 65 questions (50 scored + 15 unscored), 90 minutes, 100 USD, cut score 700/1,000, 3-year validity, Foundational level, the four question types and the 20 / 24 / 28 / 14 / 14 weights;
• Service status you can meet as a distractor: SageMaker Model Monitor and SageMaker Clarify closed to new customers on 30 July 2026: status is maintenance, not a shutdown and not a rebrand, so existing customers keep them with no new features; Amazon Kendra entered maintenance 30 June 2026, closed to new customers 30 July 2026, with Bedrock Managed Knowledge Base as AWS's directed replacement; lifecycle vocabulary is Maintenance → Sunset → Full Shutdown.

Non-verified, do not memorise as exam facts: circulated Claude 5-series launch dates (third-party catalogue, not an AWS post); the Q Business closure date (AWS docs say only "no longer open to new customers"); first-generation Nova Sonic and Bedrock Managed Knowledge Base launch dates; conflicting Claude Opus 4.6 pricing ($5.00/$25.00 vs $15.00/$75.00); the claim that batch and prompt-caching savings multiply (AWS wording says caching is not supported with batch); the exact Mechanical Turk end-of-support day (General Reference 30 Sep 2026 vs What's New 29 Sep 2026); and any AIF-C02 successor, none has been announced.

• 📚 Did you know? Bedrock Guardrails now carries six policies, content moderation, prompt-attack detection, denied topics, PII/sensitive-information filters, contextual grounding and Automated Reasoning checks (preview 3 December 2024), and AWS publishes up to 88% harmful-content blocking plus up to 99% accuracy for Automated Reasoning, whose job is to validate, correct and explain answers with formal logic rather than merely block them.

11.5 Exam Strategy Essentials

Logistics only earn marks when they become behaviour. These are the highest-yield tactics from AWS's own exam-preparation guidance plus the trap families that recur across all four question types.

Pacing plan (90 minutes, 65 items):

Trap families and the defensive habit:

Three more rules worth memorising: never carry two tied options past 90 seconds, flag, sweep, commit; multiple response, ordering and matching are all-or-nothing, so one wrong pair zeroes the item; and the exam wants the best answer, not the trick answer, AWS rewards the managed, practical, cost-aware option. With roughly 83 seconds per item, three minutes spent stalling on one question is two questions lost.

• 📚 Did you know? Admin details that belong in your calendar: your score report posts to the AWS Certification Account within 5 business days, you may reschedule up to 24 hours before the appointment with a maximum of 2 reschedules, and a no-show or a change inside 24 hours forfeits the fee; while English proctoring for online delivery runs 24×7.

---

12. Real-World Case Studies

Domains 1 and 3 are taught with abstractions, but AWS publishes what the same patterns look like in production. Every figure below is customer- or AWS-claimed and unaudited, and the source is named so you can check it: the examinable point is the pattern (which tier of AI was chosen, which service did the work, which number moved), not the marketing.

12.1 Case × services × outcome

12.2 Three cases worth knowing in detail

Case A: Sun Finance: the LLM-only prototype that failed. A lender in 9 countries reviewed 60% of microloan applications by hand, at 10 minutes to 20 hours per file. The first attempt sent ID photos straight to Claude Sonnet 4 for JSON extraction: on a 585-image test set it scored 61.8% overall and just 43% on the ID number, and AWS rejected it because the model's privacy protections block direct PII extraction. The working pipeline separates OCR from reasoning: Textract for OCR, Rekognition as fallback and face masking, the LLM only for structuring, then validation rules plus Titan Multimodal Embeddings in S3 Vectors for fraud similarity. Result: 79.73% → 90.80% accuracy (ID number 74.32% → 89.40%), −91% cost per document, 20 h → <5 s. Exam lesson: document extraction is a prebuilt-service-first job; a general LLM is the structuring layer, not the OCR layer.

Case B: Nippon India: retrieval engineering, not a bigger model. Naive RAG degraded as document volume grew and produced hallucinations. The fix added FM-as-parser parsing, query reformulation, multi-query RAG, reranker models, GraphRAG and metadata filtering on Bedrock Knowledge Bases, plus Guardrails and citations. Result: accuracy +>95%, hallucination −90 to 95%, report generation 2 days → ~10 minutes. AWS flags these as generally available product features, not customer-only code. Exam lesson: when RAG quality is the problem, tune retrieval before you change the model.

Case C, Chronomics: managed AutoML beats four months of DIY. A health-tech team spent 4 months building a custom computer-vision model for reading COVID test results and never reached target. Rekognition Custom Labels shipped in 3 to 4 weeks at 96.5% accuracy / 97.9% F1. Raising the confidence threshold to 0.99 lifted accuracy to 99.6% but discarded 5% of predictions (0.999 → 99.87% with 27% discarded), so the team had to decide what happens to the discarded tail, the human-review path AWS documents for exactly this case. Exam lesson: choose the tier deliberately, and always answer "who reviews the low-confidence output?" (Amazon A2I).

12.3 What the numbers have in common

• 📚 Did you know? The documented failure cases are the most examinable part of this research: Sun Finance's LLM-only OCR attempt (61.8%), Chronomics' four-month custom vision build (3 to 4 weeks with Rekognition Custom Labels instead) and Forethought's self-managed Kubernetes inference (moved to SageMaker for −66%). AWS also publishes the pre-production rule behind them: value first, safeguards before launch, delivered through the Well-Architected Framework and the AWS Cloud Adoption Framework for AI/ML/Generative AI.

---

Practice Questions

---