DestrezaAI ← AWS AI Practitioner Study Kit
AIF-C01 · Study guide · Checked November 2026

The AWS AI Practitioner exam is not about hard questions. It is about confusing ones.

Nobody asks you to build a neural network. Instead, the exam shows you look-alikes: two AWS services, or two ideas, that sound almost the same. Then it asks which one fits, and you have about 83 seconds to decide. This guide shows how the exam thinks, what you can safely ignore, and how to split your days. Three interactive parts inside, no filler, and not a single “leverage”.

10 minute read3 interactive partsFacts checked against AWS
01

The AIF-C01 exam, on one ticket

Before any strategy, the facts, exactly as AWS publishes them. The exam is foundational: AWS describes the ideal candidate as someone with up to six months of exposure to AI and ML on AWS who uses these tools but does not necessarily build them.

Exam
AIF-C01
AWS Certified AI Practitioner, foundational level
Questions
65
50 count toward your score, 15 do not
Time
90 min
About 83 seconds per question
Exam fee
100 USD
Set by AWS, before any local taxes
Valid for
3 years
Then renew or recertify
Where
Anywhere
Pearson VUE test center or online proctored
700 / 1000
Scaled pass mark

Never leave a blank

Unanswered questions count as wrong, and there is no penalty for guessing. Answer everything, even when you are unsure.

One weak domain will not sink you

Scoring is compensatory: you need to pass the exam as a whole, not each section. Strong spots can cover weak ones.

15 questions are test runs

They do not count, and nobody tells you which ones they are. Treat every question as real. A strange one may simply be a decoy.

Two details people miss. The score runs from 100 to 1,000, so 700 is not the same as “70% correct”. And the exam is not only multiple choice: it can also ask multiple response (you must pick every correct option to get credit), ordering and matching.

02

Where the points live, and your plan

AWS publishes the weight of each domain as a share of the scored content. With 50 scored questions, that works out to roughly this:

DOMAIN 1Fundamentals of AI and ML
20% · about 10 questions
DOMAIN 2Fundamentals of generative AI
24% · about 12 questions
DOMAIN 3Applications of foundation models
28% · about 14 questions
DOMAIN 4Guidelines for responsible AI
14% · about 7 questions
DOMAIN 5Security, compliance and governance for AI solutions
14% · about 7 questions

Read it like this. Domains 2 and 3 are 52% of the points, and both are about generative AI and foundation models. Domains 4 and 5 add up to 28%, which is more than Domain 1 on its own, so skipping them is expensive.

How many days do you have?

Pick a number: the days are split in proportion to the points, and each one comes with its three tools, in this order. It assumes about an hour a day.

1. Understand20 min with the Concept Map of the day’s domain: see how its terms connect and which ones look alike.
2. Remember15 min with the Flashcards of the same domain, so the terms stick.
3. Prove it25 min of practice questions, to check you can use what you learned.
    03

    Spot the impostor

    This is the real skill the exam tests: telling two look-alikes apart under time pressure. Six pairs below. Each clue describes one of the two. Pick the right one before you see the rule, because on exam day nobody hands you the rule.

    1. A call recording lands in your bucket. You need the words, as text.

      Answer: Transcribe. Recording in, Transcribe. Voice out, Polly.

    2. Scanned invoices arrive as PDFs. You need the values inside the tables.

      Answer: Textract. Text you already have, Comprehend. A scan or a form, Textract first.

    3. Your team wants to call a ready-made foundation model through an API. No training involved.

      Answer: Bedrock. Train or build a model, SageMaker AI. Call a foundation model through an API, Bedrock.

    4. Traffic is irregular, and you do not want to pay while the model sits idle.

      Answer: Serverless inference. Steady traffic and instant answers, real-time. Irregular traffic and low cost, serverless.

    5. Every flagged transaction costs money to investigate. You want fewer false alarms.

      Answer: Precision. False alarms are costly, precision. Missing a case is costly, recall.

    6. You have no labels. You want the data to reveal groups you did not know existed.

      Answer: Clustering. Categories given in advance, classification. Groups to discover, clustering.

    All 10 pairs on the free Concept Map (opens in a new tab) Drill them as Flashcards (opens in a new tab)
    04

    Zoom out: the family tree

    Some look-alikes are not two services but two layers of the same idea. The free Domain 1 Concept Map opens with a family tree: each term sits inside the one before it. Mix up two layers and a question turns into a coin toss. Tap a layer to open its card, exactly as on a Concept Map.

    Artificial intelligence
    Any technique that lets machines do tasks that normally need human intelligence. AI is the umbrella. All ML is AI, but not all AI is ML: rule-based expert systems are AI too. “Broadest category”, “rule-based system”.
    Machine learning
    Systems that learn patterns from data instead of following rules written by people. ML learns the rules from data. Traditional software has rules coded by humans. “Learns from historical data”, “improves with more data”.
    Deep learning
    ML using neural networks with many layers. Strongest on unstructured data like images, audio and text. A subset of ML, not a separate field. Needs more data and compute than classic ML. “Many layers”, “images, speech, large unstructured datasets”.
    Generative AI
    Deep learning models that create new content: text, images, code, audio. Traditional ML predicts or classifies. GenAI produces something new. “Generate”, “create”, “draft”, “summarize”.
    Large language model
    A foundation model trained on huge amounts of text to understand and generate language. Every LLM is a foundation model, but not every foundation model is an LLM: some generate images. “Text generation”, “chat”, “tokens”, “transformer”.
    05

    Four answers that feel right and are not

    The best wrong answers are not random. They are the reflex answer, the one your brain reaches for first. Each card below is one of the free practice questions: the reflex, what the question actually asked, and the rule to keep.

    Domain 1 · Use cases

    The model already running

    A card issuer must score two million transactions a day for fraud, within 20 milliseconds each, from tabular data, and give a stated reason for every decline. Its team already runs a foundation model for emails.

    Your reflexReuse the foundation model. It is already there, and it can write an explanation.
    What it askedA traditional classifier with reason codes. A generated explanation is not how the model decided, and 20 milliseconds rules out a large model.

    Carry this: a foundation model for language and generation; traditional ML for a narrow prediction on tabular data under latency and explainability rules.

    Domain 4 · Guardrails

    The swapped filters

    An assistant must never name three rival brands, and must refuse any talk about pending lawsuits, however the user words it.

    Your reflexA denied topic for the brands and a word filter for lawsuits.
    What it askedThe reverse: a word filter for the three names, a denied topic for lawsuits.

    Carry this: exact strings go in a word filter, a subject that can be worded many ways goes in a denied topic.

    Domain 2 · Limits

    The setting that proves nothing

    A chatbot words the same answer slightly differently each time, even at temperature zero. Regulators require the lender to show exactly what each customer was told, months later.

    Your reflexLower the randomness settings until the answers are identical.
    What it askedTurn on model invocation logging and keep every prompt and response.

    Carry this: never promise identical outputs from a generative model. When you must prove what it said, log what it said.

    Domain 3 · Retrieval

    The single index

    A knowledge base serves staff in 30 countries, many documents apply to several countries, answers must use only material valid in the user's country, and the team wants one index.

    Your reflexOne knowledge base per country.
    What it askedTag each document with its countries and filter retrieval on the user's country.

    Carry this: restrict what the model can see at retrieval time, not what it is asked to say. Plan the metadata at ingestion.

    Now try one for real

    On the real exam you only learn whether you passed. Here, every option is explained, the wrong ones too, because that is where the learning is. Pick an answer.

    Domain 3 · Free practice question

    A company has 40 support emails labeled with one of six urgency levels, must start triage within a week, and expects to have 5,000 labeled emails in six months. Which plan fits?

    Rule to carry forward: start with the cheapest adaptation that meets the need, and move up the ladder only when more data and a measured gap justify it.

    Try the 10 free questions (opens in a new tab) The full question bank
    06

    A guide is a map. You still need reps.

    Reading about look-alikes is not the same as spotting them. These three are free, with no sign-up, and they cover Domain 1 and the first ten practice questions.

    Understand
    Free Concept Map Every Domain 1 term on one page, with the 10 look-alike pairs the exam swaps and the clue words that give the answer away.
    Open the map →
    Remember
    Free Flashcards Each Domain 1 term as a card that comes back just before you would forget it.
    Flip a card →
    Prove it
    10 free questions Scenario questions with an explanation for every option, including why the tempting one is wrong.
    Start practicing →
    See the full Study Kit → Back to the overview

    Practice improves your chances. Nobody can promise a pass. Domain 1 is free; Domains 2 to 5 come with the full Study Kit.

    07

    What AWS says you do not need

    This is the most reassuring page of the official exam guide. AWS lists job tasks the target candidate is not expected to perform. If a course spends hours on these, you can skim it.

    Out of scope, per the AWS exam guide

    • Developing or coding AI and ML models or algorithms
    • Implementing data engineering or feature engineering techniques
    • Performing hyperparameter tuning or model optimization
    • Building and deploying AI and ML pipelines or infrastructure
    • Conducting mathematical or statistical analysis of models
    • Implementing security or compliance protocols for AI and ML systems
    • Developing and implementing governance frameworks and policies

    AWS notes the list is not exhaustive.

    What you need instead

    Recognize the idea, match it to a business scenario, name the service that fits.

    A fine line: you will not be asked to implement security or governance, but Domains 4 and 5 are 28% of the exam, so you still need to know what each tool is for.

    08

    Five small rules for 90 minutes

    1. Do the clock math.65 questions in 90 minutes is about 83 seconds each. Answer what you know on the first pass, flag the rest, and come back with the time you saved.
    2. Never leave a blank.An empty answer is scored as wrong and a guess costs nothing. In the last minute, fill every gap.
    3. Find the sentence that hurts.Cost, latency, “no servers to manage”, “without training data”. The constraint in the scenario picks the answer, not the shiniest service name.
    4. “Choose two” means both.On multiple response questions you must select every correct option to earn credit. One right answer out of two is not half a point.
    5. Do not panic at the weird one.Fifteen questions do not count and are not marked. If one looks like it came from another exam, flag it, make your best guess and move on.
    Start with the free questions →
    Sources. Exam facts checked on November 10, 2026. AWS can change them, so confirm before you book.