CertifyAIxAce every AI certification.

AIF-C01 sample questions with answers

10 free practice questions for the AWS Certified AI Practitioner exam. Try each one, then open the answer to see why the right option wins and every other option loses.

Question 1Fundamentals of AI and ML

A bank's fraud investigators close several hundred cases a week. Every case already carries a decision recorded in the case management system, and the analyst adds rough working notes as they go. The bank wants a managed AWS service that takes those structured case facts together with the analyst's notes and drafts the written case narrative in the bank's house style, where tone and length are adjusted by changing the instruction rather than by retraining a model. Which AWS service meets this requirement?

  1. A.

    Amazon Fraud Detector, the fully managed service that trains on the bank's own historical fraud event data and returns a fraud score plus a rule-driven outcome, such as approve or send for review, for each online event that it evaluates.

  2. B.

    Amazon Textract, which extracts printed and handwritten text along with forms and tables from scanned documents and returns them to the calling application as structured data.

  3. C.

    Amazon Bedrock, which gives managed API access to foundation models, so the case facts and the analyst's notes go into a prompt and the drafted narrative comes back, with style changed by editing the instruction.

  4. D.

    Amazon Comprehend, which applies natural language processing to text that already exists in order to detect entities, key phrases, personally identifiable information, dominant language and sentiment.

Show answer

Answer: C

Only Amazon Bedrock generates new prose from a prompt; the fraud-specific service scores events, and the other three analyse, extract from, or speak text that already exists.

  • A. Amazon Fraud Detector detects potentially fraudulent activity such as unauthorised transactions and fake account creation by scoring events against a model trained on your fraud history. Its output is a score and an outcome, never written prose. AWS also closed it to new customers in November 2025 and points new work at SageMaker, AutoGluon and AWS WAF.
  • B. Amazon Textract lifts printed and handwritten text, forms and tables out of scanned documents. It could digitise a handwritten note as an input step, but it generates no narrative of its own.
  • C. Amazon Bedrock provides managed access to foundation models through an API, which is the only option here capable of generating the narrative text, and prompt instructions control tone and length without retraining.
  • D. Amazon Comprehend extracts insight from documents that already exist, returning entities, key phrases, PII, language and sentiment. It analyses text rather than producing any, so it cannot draft the case narrative.
Question 2Fundamentals of AI and ML

A startup needs sentiment classification for product reviews and has a small engineering team, a short timeline, and no labelled training data. What is the sensible starting point?

  1. A.

    Wait until the review volume is large enough to justify a custom model, and handle sentiment classification manually in the meantime by having staff read and categorise each review as it arrives.

  2. B.

    Start from an existing pre-trained model rather than training a custom one from scratch.

  3. C.

    Train a custom model on the unlabelled reviews it already holds, since sentiment can be learned from the text of the reviews themselves without any labels being applied to them first.

  4. D.

    Collect and label a training dataset of its own reviews over the coming months and train a custom model on it, so that the resulting model is tuned precisely to the vocabulary its own customers use.

Show answer

Answer: B

Pre-trained models exist for common tasks; train custom only when you have a reason to.

  • A. Manual handling defers the benefit and consumes the small team.
  • B. A pre-trained model gives an immediate result on a common task with no labelling effort.
  • C. Unlabelled data cannot train a supervised sentiment classifier.
  • D. Months of labelling to rebuild something that already exists delays all value.
Question 3Fundamentals of AI and ML

A clinic built a model that predicts how many minutes a patient will wait, by grouping historical waits into 'under 15', '15 to 45' and 'over 45' minutes and training a three-class model. Reception reports that the display is unhelpful precisely when it matters: patients told '15 to 45' are as likely to wait 16 minutes as 44, and the model flips classes for near-identical situations. What explains this?

  1. A.

    The three classes are imbalanced, and rebalancing them would remove the variation that reception observes inside the middle band.

  2. B.

    The model is underfitting, and adding more capacity would let a three-class model produce minute-level wait times for each patient.

  3. C.

    A continuous quantity was converted into coarse classes, so the model cannot express position within a band or distance from a boundary.

  4. D.

    The feature set omits the patient's arrival time, which is the only feature capable of distinguishing a 16-minute wait from a 44-minute wait.

Show answer

Answer: C

Discretising a continuous target destroys within-band resolution and makes boundary cases unstable; the fix is to predict minutes directly.

  • A. Class imbalance affects which class is favoured, not the model's inability to distinguish values inside one band.
  • B. No amount of capacity lets a three-class model output minutes; the resolution was destroyed in the labels, not the model.
  • C. Banding the target before training removes within-band information and forces an arbitrary hard boundary through a smooth quantity.
  • D. A missing feature cannot restore a distinction the target variable no longer encodes after banding.
Question 4Fundamentals of AI and ML

A bookshop chain has three years of anonymised purchase baskets. The marketing team believes there are distinct shopping patterns among its customers but has never defined what they are and does not want to impose its own assumptions. It wants the data to reveal natural groupings that can then be examined and named. Which technique should the team apply?

  1. A.

    Regression, because the model must estimate how strongly each customer belongs to a group and belonging strength is a continuous quantity.

  2. B.

    Clustering, because it groups records by similarity without requiring predefined categories.

  3. C.

    Binary classification, because every customer either belongs to a recognisable shopping pattern or does not belong to any pattern at all.

  4. D.

    Multi-class classification, because marketing can define the categories first and the model will then confirm whether those categories were correct.

Show answer

Answer: B

Clustering groups records by similarity with no predefined categories, which is exactly what discovering shopping patterns requires.

  • A. There is no recorded continuous target; belonging strength is an output of some clustering methods, not training data.
  • B. Grouping records by similarity with no predefined categories is the definition of clustering.
  • C. It collapses the question to one crude distinction and still requires labels the bookshop does not have.
  • D. Classification reproduces a taxonomy supplied in advance, which is exactly the assumption the team wants to avoid making.
Question 5Fundamentals of AI and ML

A ride-hailing company deployed a demand-prediction model to a SageMaker real-time endpoint six months ago. Prediction quality has gradually declined as the city's travel patterns changed after a new transit line opened, and the team only discovered the problem from customer complaints. The team wants an automated way to detect when the statistical properties of incoming requests deviate from the training data so they can trigger retraining. Every capability below is already available in the team’s existing SageMaker AI environment. Which one meets this requirement?

  1. A.

    Use Amazon SageMaker Ground Truth to label a sample of incoming requests each day.

  2. B.

    Use Amazon SageMaker Experiments to compare the training runs from six months ago against the runs from today to see what changed.

  3. C.

    Enable Amazon SageMaker Model Monitor on the endpoint with a baseline from the training data and scheduled monitoring for data drift.

  4. D.

    Use Amazon SageMaker Data Wrangler to re-explore the original training dataset every week.

Show answer

Answer: C

SageMaker Model Monitor compares live endpoint traffic against a training baseline and raises alerts on data drift.

  • A. Ground Truth produces labels; it does not compare input distributions or automate drift alerts.
  • B. Experiments compares training-time runs and parameters; it does not observe the deployed endpoint's traffic.
  • C. Model Monitor is designed to baseline training data, watch live endpoint inputs and alert on drift automatically.
  • D. Re-exploring the original training data cannot reveal changes in live traffic; drift is detected by comparing production inputs to the baseline.
Question 6Fundamentals of AI and ML

A telematics company pre-processes several hundred terabytes of vehicle event data every night with an existing Apache Spark codebase. The job runs for hours, the team has tuned it against specific instance families, and it wants to place much of the capacity on interruption-tolerant Spot capacity to control cost. Which AWS service fits this pre-processing stage BEST?

  1. A.

    AWS Glue, because a serverless ETL service always costs less than a managed cluster regardless of how long the job runs each night.

  2. B.

    AWS Data Exchange, which is the AWS service for transforming large volumes of internally generated event data before training.

  3. C.

    Amazon Redshift, because loading the events into a data warehouse is a prerequisite for any large-scale Spark processing job.

  4. D.

    Amazon EMR, which runs managed Apache Spark clusters and lets the team choose instance types and mix purchase options across the cluster.

Show answer

Answer: D

Amazon EMR runs managed Spark clusters with control over instance types and purchase options, which is what a tuned multi-hour nightly job needs.

  • A. The blanket claim is false: serverless is not always cheaper, and sustained multi-hour jobs are where a sized cluster can win.
  • B. AWS Data Exchange acquires third-party data; it does not transform an organisation's own event data.
  • C. A data warehouse is not a prerequisite for Spark processing and is not a Spark execution environment.
  • D. EMR runs managed Spark and gives control over instance types and purchase options, matching a tuned multi-hour nightly job.
Question 7Fundamentals of AI and ML

A payments processor is planning a fraud-detection capability and wants the design document to state up front what makes this application harder than its other classification work, so that the evaluation plan and the operating model are sized realistically. Which two characteristics genuinely distinguish fraud detection? (Choose TWO.)

Choose 2.

  1. A.

    The positive class is extremely rare, so aggregate accuracy is uninformative and the evaluation must use metrics suited to imbalance.

  2. B.

    Fraud detection requires unsupervised learning exclusively, because labelling a transaction as fraudulent after the event is never possible in practice.

  3. C.

    The cost of a false positive and a false negative are equal by definition, so a single accuracy threshold can be applied across all transaction types.

  4. D.

    The behaviour being detected adapts in response to detection, so the data distribution shifts deliberately and retraining cadence matters more than usual.

  5. E.

    Ground truth is available instantly for every transaction, which makes continuous supervised evaluation simpler than in other classification problems.

Show answer

Answer: A, D

Severe class imbalance and an adversary that adapts to detection are what make fraud detection harder than ordinary classification.

  • A. Severe imbalance makes accuracy meaningless and forces imbalance-aware metrics and a sufficiently populated test set.
  • B. Confirmed fraud cases provide labels, and supervised models are standard in this domain alongside anomaly detection.
  • C. The two error types have very different costs, which is why thresholds are tuned per segment rather than set once globally.
  • D. An adversary that adapts to detection makes distribution shift deliberate, so monitoring and retraining cadence become structural.
  • E. Fraud labels arrive late through chargebacks and investigations; delayed ground truth is one of the domain's defining difficulties.
Question 8Fundamentals of AI and ML

A team must choose one visual data preparation tool. The users are ML practitioners who work inside Amazon SageMaker AI, the flows they build must feed directly into training jobs and be exportable as steps in an ML pipeline, and they need ML-specific preparation such as target encoding and handling of the training target column. Which tool fits, and what is the deciding factor?

  1. A.

    Amazon SageMaker Feature Store, because storing prepared feature values removes the need for any visual preparation tool in the workflow.

  2. B.

    Data Wrangler in Amazon SageMaker Canvas, because it is the ML-oriented preparation experience inside SageMaker AI and its flows feed training and pipelines directly.

  3. C.

    AWS Glue DataBrew, because DataBrew is the only AWS service that offers prebuilt transformations through a point-and-click interface.

  4. D.

    AWS Glue DataBrew, because a visual preparation tool aimed at analysts is always preferable to one aimed at machine learning practitioners.

Show answer

Answer: B

The deciding factor is that the flows must feed SageMaker AI training and pipelines directly, which is Data Wrangler's purpose.

  • A. Feature Store stores and serves features but does not compute them, so it cannot replace a preparation tool.
  • B. Its ML-oriented transformations and direct export into SageMaker AI training and pipeline steps match all three requirements.
  • C. Prebuilt point-and-click transformations are not unique to DataBrew, so the exclusivity claim is false.
  • D. Neither tool is universally preferable; the choice depends on the user and where the output must go.
Question 9Fundamentals of AI and ML

A shipping agent processes about 25 unusual customs cases a year. Each takes an experienced clerk roughly half a day to resolve, and the clerk already handles them reliably. A consultant proposes training a model to classify these cases, which would require assembling and labelling a dataset, building a pipeline, and maintaining it thereafter. What should the AI practitioner advise?

  1. A.

    Do not build the model, because customs classification is a regulated activity and regulated activities can never be supported by ML.

  2. B.

    Build the model, because any manual task that follows repeatable rules should be automated regardless of how often it occurs each year.

  3. C.

    Build the model, because 25 examples a year is ample training data once the clerk's decisions from the past twelve months are labelled.

  4. D.

    Do not build the model: the annual effort saved is small, while the build and ongoing maintenance cost is not.

Show answer

Answer: D

Twenty-five cases a year cannot repay the build and maintenance cost of a model, so the cost-benefit test fails.

  • A. Regulation imposes controls such as explainability and oversight; it does not prohibit machine learning outright.
  • B. Repeatability is not sufficient; frequency determines whether automation earns back its cost.
  • C. Twenty-five heterogeneous examples is far too small a sample for a supervised model to learn from.
  • D. Low frequency caps the benefit well below the build and indefinite maintenance cost, so the investment cannot repay itself.
Question 10Fundamentals of AI and ML

A software company is building an invoicing feature that must calculate the exact sales tax for each order according to published tax tables that vary by jurisdiction and product category. The tax authority audits the results, and any deviation from the published rate is a compliance violation. A product manager proposes using ML because the company already has millions of historical invoices. What should the AI practitioner recommend?

  1. A.

    Implement the published tax tables as deterministic business rules, because an exact, auditable outcome is required rather than a prediction.

  2. B.

    Use a classification model to assign each order to a tax bracket and then apply the average rate for that bracket.

  3. C.

    Train a regression model on the historical invoices to predict the tax amount for each new order.

  4. D.

    Use a large language model to read the order and generate the tax amount, with a low temperature for consistency.

Show answer

Answer: A

When an exact, rule-defined outcome is required, deterministic code is appropriate and a predictive model is not.

  • A. The outcome is fully determined by known rules, so deterministic logic is exact, auditable and cheaper than ML.
  • B. Bracket averages are approximations by design and cannot match jurisdiction-specific published rates exactly.
  • C. Regression outputs an estimate with error; tax must be exact per the published table, so any prediction error is a violation.
  • D. LLMs are nondeterministic and can hallucinate numbers; low temperature reduces variance but does not guarantee correctness.

Keep going with 497 more AIF-C01 questions

Free papers every day, in the real exam formats, with progress by exam domain. Unlock every paper and timed mock exam when you are ready.

Practise freeSee all AIF-C01 papers
AIF-C01 sample questions with answers (10 free) · CertifyAIx