AIF-C01 · Domain 1 of 5

AIF-C01 Domain 1: Fundamentals of AI and ML

Domain 1 builds the vocabulary of the AWS Certified AI Practitioner exam: core AI and ML terms, learning and inference types, when ML fits a problem, the ML pipeline, MLOps and the metrics used to judge a model.

Exam weight
20%
Official task statements
3
Estimated questions estimate
~10
Suggested study time estimate
10–14 hrs

About 12 min read Reviewed October 2026

Overview

Fundamentals of AI and ML has three task statements. The first covers terminology: AI, ML, deep learning, neural networks, computer vision, NLP, LLMs, GenAI and agentic AI, plus inference types (batch, real-time, asynchronous, serverless), data types and supervised, unsupervised and reinforcement learning. The second covers use cases, including when ML is the wrong tool, how to pick regression, classification or clustering, and what managed services such as Amazon SageMaker AI, Transcribe, Translate, Comprehend, Lex and Polly do. The third covers the ML development lifecycle, MLOps and model and business metrics.

AIF-C01 remains the current exam as of October 2026, and no AIF-C02 has been announced. AWS revised the guide during 2026 (version 1.2 is dated October 29, 2026), adding agentic AI and newer services such as Kiro and Amazon Quick. Italian and German delivery ended after October 15, 2026.

How it's tested

Domain 1 questions are short business scenarios at the awareness level. They describe a goal, a dataset or a prediction need and ask which learning type, ML technique, inference option, managed service or metric fits. Configuration detail is not tested; precise definitions and correct service-to-task matching are.

  • Pick the technique

    A prediction goal is described, such as a number, a category or natural groupings, and you choose regression, classification or clustering.

  • Match the managed service

    A task such as speech to text, translation, sentiment or a chatbot is described, and you name the AWS AI service built for it.

  • Choose the right metric

    A model's error costs are described, and you decide whether precision, recall, F1 score, accuracy or a business metric matters most.

Reasoning it through

A bank wants to flag fraudulent card transactions, and missing a real fraud case costs far more than reviewing a false alarm. The task is binary classification. Because false negatives are the expensive error, recall is the metric to maximize, even if precision drops and analysts review more alerts. Accuracy would mislead here because fraud is rare, so a model that predicts no fraud would still score very high.

Task statements 3

Official task statements from the Amazon Web Services outline, in outline order.

1.1 Explain basic AI concepts and terminologiesCore AI terms, inference types, data types and learning types.
What it means
This task statement covers the basic vocabulary of AI: AI, ML, deep learning, neural networks, computer vision, NLP, models, algorithms, training and inferencing, bias, fairness, fit, LLMs, GenAI and agentic AI. It also covers inference modes (batch, real-time, asynchronous, serverless), data types such as labeled, unlabeled, tabular, time-series, image and text, and the main learning approaches.
What to understand
The terms nest: AI is the broadest field, ML learns patterns from data, deep learning uses multi-layer neural networks, and GenAI uses large deep learning models to create new content. Agentic AI goes a step further by planning and taking actions with tools toward a goal. Supervised learning needs labeled data, unsupervised learning finds structure in unlabeled data, and reinforcement learning learns from rewards. Real-time inference serves low-latency requests, batch scores large datasets offline, asynchronous handles large payloads with queued responses, and serverless suits intermittent traffic.
How it shows up
A short description is given, and you name the term, learning type, data type or inference mode it describes.
Common confusion
Treating reinforcement learning as unsupervised; reinforcement learning learns from rewards for actions, not from finding clusters.
1.2 Identify practical use cases for AIWhere AI adds value, when it does not, technique selection and AWS AI services.
What it means
This task statement covers recognizing where AI and ML add value (supporting decisions, scaling, automation), deciding when they are not appropriate, choosing regression, classification or clustering, recognizing real-world applications such as fraud detection and recommendations, knowing what AWS managed AI services do, and judging when a traditional ML model fits better than a foundation model.
What to understand
ML is a poor fit when a deterministic, exact outcome is required, when a simple rule solves the problem, or when the cost of data and training outweighs the benefit. Regression predicts a number, classification predicts a category and clustering groups unlabeled records. Amazon Transcribe converts speech to text, Polly does the reverse, Translate converts languages, Comprehend extracts sentiment and entities, Lex builds conversational bots and SageMaker AI builds custom models. Traditional ML often wins where regulators demand explainability or where latency and cost are tight.
How it shows up
A business need is described, and you decide whether AI fits and which technique or managed service applies.
Common confusion
Mixing up Amazon Polly and Amazon Transcribe; Polly turns text into speech, Transcribe turns speech into text.
1.3 Describe the AI/ML development lifecycleML pipeline stages, model sources, production options, MLOps and metrics.
What it means
This task statement covers the stages of an ML pipeline (data collection, exploratory analysis, preparation, feature engineering, training, tuning, evaluation, deployment and monitoring), sources of foundation models, ways to run a model in production, AWS services for each stage, MLOps concepts and the model and business metrics used to evaluate ML.
What to understand
Each stage maps to AWS tooling: SageMaker AI covers data preparation, training, tuning, deployment and monitoring, Amazon Bedrock provides managed foundation models, Kiro supports agentic, spec-driven development and Amazon Quick gives business users agentic research and analysis. A managed API service removes hosting work, while a self-hosted endpoint gives more control. MLOps brings repeatable pipelines, experiment tracking, monitoring for drift and scheduled re-training. Accuracy, precision, recall and F1 judge the model; cost per user, ROI and customer feedback judge the business result.
How it shows up
A lifecycle step or production concern is described, and you identify the stage, MLOps practice, service or metric.
Common confusion
Assuming deployment ends the lifecycle; monitoring and re-training continue because data drift degrades models over time.

High-yield concepts

The terms, rules and formulas this domain turns into questions most often.

  • AI vs. ML vs. deep learning vs. GenAI

    Nested fields from broad AI down to content-generating deep learning models.

    Remember
    GenAI is a subset of deep learning, which is a subset of ML.
    Tested as
    Terminology questions.
    Trap
    Treating GenAI as separate from ML.

  • Agentic AI

    AI that plans steps and uses tools to act toward a goal with limited supervision.

    Remember
    It acts, not only generates.
    Tested as
    New 2026 guide wording.
    Trap
    Calling any chatbot an agent.

  • Supervised, unsupervised, reinforcement

    Learning from labels, from unlabeled structure, or from rewards.

    Remember
    Labels mean supervised.
    Tested as
    Learning type questions.
    Trap
    Placing reinforcement learning under unsupervised.

  • Regression, classification, clustering

    Predict a number, predict a category, or group similar records.

    Remember
    Clustering needs no labels.
    Tested as
    Technique selection questions.
    Trap
    Calling a yes/no prediction regression.

  • Inference modes

    Real-time, batch, asynchronous and serverless ways to get predictions.

    Remember
    Asynchronous suits large payloads and long processing.
    Tested as
    Deployment questions.
    Trap
    Choosing real-time for overnight scoring of millions of rows.

  • Overfitting vs. underfitting

    Memorizing training data versus failing to learn the pattern.

    Remember
    Overfit: good on training, poor on new data.
    Tested as
    Fit questions.
    Trap
    Fixing overfitting by training longer.

  • Precision vs. recall

    Precision: how many flagged items are right. Recall: how many real positives were caught.

    Remember
    Costly misses mean favor recall.
    Tested as
    Metric questions.
    Trap
    Using accuracy on imbalanced data.

  • F1 score

    Harmonic mean of precision and recall.

    Remember
    Use when both errors matter.
    Tested as
    Metric questions.
    Trap
    Confusing it with accuracy.

  • MLOps

    Practices for repeatable, monitored, maintainable ML in production.

    Remember
    Includes drift monitoring and re-training.
    Tested as
    Lifecycle questions.
    Trap
    Thinking MLOps ends at deployment.

  • AWS managed AI services

    Transcribe, Translate, Comprehend, Lex, Polly and others solve common tasks without training models.

    Remember
    Pick the purpose-built service first.
    Tested as
    Service matching questions.
    Trap
    Building in SageMaker AI when a managed service fits.

  • When not to use ML

    Exact, rule-based outcomes or poor cost-benefit make ML the wrong choice.

    Remember
    Predictions are probabilistic.
    Tested as
    Suitability questions.
    Trap
    Assuming AI improves every process.

Side by side

Concepts candidates mix up, compared on the point that decides the answer.

Learning types

ConceptMeaningKey differenceCommon trap
Supervised Learns from labeled examples Needs labels Using it with no labeled data
Unsupervised Finds patterns in unlabeled data No labels; clustering, anomalies Confusing it with reinforcement learning
Reinforcement Learns by reward and penalty Agent acts in an environment Assuming it needs a labeled dataset

Inference options

ConceptMeaningKey differenceCommon trap
Real-time Persistent endpoint, low latency Immediate single predictions Using it for large offline jobs
Batch Scores a whole dataset offline No endpoint kept running Choosing it for interactive apps
Asynchronous Queues requests, returns results later Large payloads, long processing Confusing it with batch
Serverless Scales to zero between requests Intermittent, unpredictable traffic Ignoring cold-start latency

AWS AI services by task

ConceptMeaningKey differenceCommon trap
Amazon Transcribe Speech to text Audio input Confusing it with Polly
Amazon Polly Text to speech Audio output Choosing it for transcription
Amazon Comprehend NLP insights such as sentiment and entities Analyzes text Choosing it for translation
Amazon Lex Conversational voice and text bots Builds chat interfaces Confusing it with an LLM service

Exam traps

Each trap is an answer that looks right. Learn the clue that should change your mind.

  1. Accuracy on rare events

    What candidates pick
    Pick the model with the highest accuracy for fraud or disease detection.
    Why it looks right
    Accuracy sounds like overall quality.
    Why it's wrong
    With rare positives, a model that always predicts the majority class scores high accuracy while missing every case that matters.

    Clue to watch for The scenario mentions imbalanced data or costly misses.

  2. ML for deterministic rules

    What candidates pick
    Recommend training a model to apply a fixed tax rule.
    Why it looks right
    AI seems modern and capable.
    Why it's wrong
    When a specific, exact outcome is required, a rule or calculation is cheaper and always correct; ML only predicts.

    Clue to watch for The requirement is an exact, known calculation.

  3. Clustering as classification

    What candidates pick
    Choose classification to group customers with no existing labels.
    Why it looks right
    Both put records into groups.
    Why it's wrong
    Classification needs labeled categories in advance; clustering discovers groups in unlabeled data.

    Clue to watch for No labels exist and segments are unknown.

  4. Custom model over managed service

    What candidates pick
    Build a speech recognition model in SageMaker AI.
    Why it looks right
    Custom models feel more flexible.
    Why it's wrong
    Amazon Transcribe already converts speech to text with no training, which is faster and cheaper for a standard need.

    Clue to watch for A common task with no unusual requirements.

  5. FM where explainability is mandated

    What candidates pick
    Use a large foundation model for a regulated credit decision.
    Why it looks right
    FMs are powerful and current.
    Why it's wrong
    Regulators often require clear reasons for each decision, and simpler traditional models are easier to explain and audit.

    Clue to watch for The scenario stresses regulation or explainability.

  6. Real-time for nightly scoring

    What candidates pick
    Deploy a real-time endpoint to score a full customer table overnight.
    Why it looks right
    Real-time sounds faster.
    Why it's wrong
    Batch inference processes large datasets without paying for an always-on endpoint.

    Clue to watch for Results are needed later, not instantly.

Answer framework

Use these steps to work through any Domain 1 scenario about AI concepts, use cases or the ML lifecycle.

  1. Decide first whether the problem needs a prediction at all, because exact rule-based outcomes or poor cost-benefit rule ML out.
  2. Identify the output: a number points to regression, a known category to classification, and unknown groupings in unlabeled data to clustering.
  3. Check whether a managed AWS AI service such as Transcribe, Comprehend or Lex already solves the task before choosing SageMaker AI.
  4. Match the inference mode to latency and volume: real-time for instant answers, batch for large offline jobs, asynchronous for big payloads, serverless for sporadic traffic.
  5. Pick the metric from the cost of errors, favoring recall when misses are expensive and precision when false alarms are, and add a business metric such as ROI.

Practice question

A retailer has two years of purchase history with no customer labels. Marketing wants to discover distinct shopper groups so it can design targeted campaigns, but it does not know in advance how many groups exist. Which ML approach should the team use?

Want more like this? The AIF-C01 practice test has questions for every domain.

Checklist

0 of 13

Tick each one you can do without notes. Your ticks stay in this browser.

Study strategy

Start with vocabulary, because Domain 1 is mostly definitions applied to short scenarios. Build a one-page sheet that nests AI, ML, deep learning, GenAI and agentic AI, then add the three learning types and the three core techniques. Write one everyday example beside each so you can recognize it when a question hides the term.

Next, learn the managed services as pairs of input and output: audio in, text out for Transcribe; text in, audio out for Polly; text in, other language out for Translate. Add Comprehend, Lex and SageMaker AI, and note where Kiro and Amazon Quick appear in the current guide.

Then work through the ML pipeline and MLOps. Draw the stages from data collection to monitoring and mark where drift triggers re-training. Finish with metrics: build a small confusion matrix by hand, calculate precision, recall and F1, and explain which one a fraud team or a marketing team would care about. Practice questions should take most of your remaining time, since speed in recognizing patterns matters more than depth here.

FAQ

How much of AIF-C01 is Domain 1?

Fundamentals of AI and ML is 20% of the scored content, about ten of the fifty scored questions, across three task statements.

Is AIF-C01 being replaced by AIF-C02?

As of October 31, 2026, AWS has not announced an AIF-C02. The exam guide was updated in 2026 (version 1.2, October 29, 2026) but the exam code is still AIF-C01.

Do I need to know how to code or train models?

No. AIF-C01 tests understanding of concepts, use cases and services. You should know what training and tuning do, not how to write the code.

What is the difference between deep learning and GenAI?

Deep learning uses multi-layer neural networks for tasks like image recognition. GenAI is a branch of deep learning that produces new content such as text, images or code.

When is ML not the right answer?

When the outcome must be exact and a rule can produce it, when there is not enough quality data, or when cost outweighs the benefit.

Which metric should I use for imbalanced data?

Use precision, recall or F1 rather than accuracy. Choose recall when missing a positive is costly and precision when false alarms are costly.

Which new topics did the 2026 guide add to this domain?

Agentic AI terminology, comparisons between AI, GenAI and agentic AI, and services such as Kiro and Amazon Quick in the ML pipeline.

Other Artificial Intelligence & Machine Learning certifications

Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.

Test your Domain 1 knowledge

You have covered AI terminology, learning and inference types, use cases, AWS AI services, the ML lifecycle and metrics. Check whether you can match each scenario to the right concept quickly.

Take the AIF-C01 practice test Check your AIF-C01 readiness

Keep going