A used-car marketplace wants to predict the exact sale price, in dollars, of each vehicle listed on its platform based on features such as mileage, age, brand, and condition. How should the machine learning problem be framed?

As a regression problem that predicts a continuous numeric value

As a binary classification problem that predicts whether a car will sell

As a clustering problem that groups similar vehicles together

As a multi-class classification problem that assigns each car to a price category

A retailer has a trained propensity model and needs to score its entire database of 20 million customers once per week. There is no requirement for low-latency, on-demand predictions, and the team wants to minimize cost and infrastructure management. Which inference option is the best fit?

Run a SageMaker batch transform job on the customer dataset each week

Deploy the model to a SageMaker real-time endpoint that runs 24/7

Build an Amazon Lex bot that returns a prediction for each customer

Expose the model through API Gateway and invoke it once per customer record

A bank is evaluating a binary classification model that flags credit-card transactions as fraudulent or legitimate. The business states that failing to catch a fraudulent transaction costs far more than sending a legitimate transaction for manual review. Which evaluation metric should the team prioritize when tuning the model?

Recall, because it measures how many actual fraud cases the model catches

Accuracy, because it measures overall correctness across all transactions

Precision, because it measures how many flagged transactions are truly fraudulent

Root mean square error (RMSE), because it penalizes large prediction errors

A team is building a document search and question-answering application on Amazon Bedrock using the Amazon Titan model family. The application must convert documents into numerical vectors for similarity search and also generate natural-language answers for users. Which statement correctly describes how Titan models are used in this application?

Amazon Titan Text Embeddings converts text into vectors for similarity search, while a Titan text generation model produces the natural-language answers.

A single Titan text generation model performs both the vectorization and the answer generation.

Amazon Titan Image Generator creates the vectors, and Titan Text Embeddings writes the answers.

Titan Text Embeddings generates the answers because embeddings models are optimized for producing fluent text.

Users of a customer-facing chatbot built on Amazon Bedrock complain that after submitting a question they stare at a blank screen for many seconds before the full answer suddenly appears. Answers are often several paragraphs long. Which change most directly improves the perceived responsiveness of the chatbot?

Invoke the model with response streaming so tokens are displayed to the user as they are generated.

Increase the maximum output token limit so the model can write longer answers.

Submit user questions through a Bedrock batch inference job.

Raise the temperature parameter so the model produces answers more quickly.

A legal team uses Amazon Bedrock Knowledge Bases over long contracts stored in Amazon S3. Retrieval returns relevant text, but answers are often incomplete because a clause and its surrounding definitions end up in separate retrieved fragments. Which adjustment to the knowledge base is most likely to fix this?

Change the chunking configuration used during ingestion so more surrounding context is kept together in each chunk.

Lower the temperature of the generation model so it stops splitting clauses.

Re-run the data source sync more frequently so the fragments merge over time.

Switch the vector store to a different supported database to change how documents are divided.

A company has customized a foundation model in Amazon Bedrock and now needs to serve it in production, where traffic is steady and high-volume and the business requires guaranteed model throughput. Which consumption option should they choose?

Purchase Provisioned Throughput for the model, committing to a number of model units.

Use standard on-demand invocation, which guarantees dedicated capacity for customized models.

Run all production traffic through batch inference jobs to reserve capacity.

Enable response streaming, which reserves throughput for the application.

A designer uses an image generation model on Amazon Bedrock to create product artwork. The generated images keep including unwanted elements such as text overlays and watermark-like patterns. The designer wants to explicitly tell the model what should NOT appear in the output, in addition to describing what should appear. Which prompting technique accomplishes this?

Negative prompting — supplying a separate instruction listing elements the model should avoid generating

Increasing the maximum output length so the model has room to avoid the unwanted elements

Lowering the random seed value to make the output reproducible

Switching to an embeddings model to filter the unwanted concepts

A retail company is prioritizing candidate projects for its first generative AI initiatives. Which project is the LEAST appropriate fit for a generative AI solution?

Computing the exact sales tax owed on each customer transaction, where results must be precise, deterministic, and auditable

Summarizing lengthy customer-service call transcripts into short case notes

Drafting first versions of product descriptions for new catalog items

Powering a conversational chatbot that answers common shipping and returns questions

A business analyst with no coding experience wants to quickly prototype a generative AI app idea — such as a marketing-slogan generator — experiment with prompts, and share the app with colleagues, all without writing code or provisioning AWS infrastructure. Which AWS offering is designed for this kind of hands-on experimentation?

PartyRock, an Amazon Bedrock Playground

Amazon SageMaker training jobs on GPU instances

AWS Trainium-based EC2 Trn1 instances

Amazon Bedrock Provisioned Throughput

A compliance officer is comparing two documentation mechanisms: AWS AI Service Cards and Amazon SageMaker Model Cards. Which statement correctly describes the key difference between them?

AI Service Cards are published by AWS to describe intended use cases, limitations, and responsible design of AWS's own AI services, while SageMaker Model Cards are created by customers to document their own models

AI Service Cards are created by customers for custom models, while SageMaker Model Cards are published by AWS for its pre-trained services

Both are auto-generated by SageMaker during training, but AI Service Cards additionally include billing information

AI Service Cards enforce runtime content filtering, while SageMaker Model Cards enforce IAM access control on models

A regulated insurer is building a premium-pricing model. Regulators require that the company be able to trace exactly how each input feature contributes to any individual decision, directly from the model's own structure rather than from an approximation. Accuracy requirements are modest. Which modeling choice best satisfies this requirement?

Use an inherently interpretable model such as linear regression or a decision tree, accepting a possible small trade-off in accuracy

Train a deep neural network and generate post-hoc feature attributions for each prediction

Use a large ensemble of gradient-boosted trees because more trees increase interpretability

Use any high-accuracy model and publish its overall accuracy metrics to regulators

A voice assistant's speech-recognition model was trained mostly on recordings of adult speakers from a single geographic region. After launch, error rates are much higher for elderly users and speakers with regional accents. Which data practice most directly addresses the root cause before the model is retrained?

Curate a training dataset that is representative of the full deployment population, collecting additional samples from the underserved age groups and accents

Train the existing dataset for more epochs so the model learns it more thoroughly

Duplicate the existing majority-group recordings to enlarge the dataset

Switch to a larger model architecture while keeping the same training data

A team routes every task in its application — including simple sentiment classification of short reviews — to the largest, most capable foundation model available. Leadership asks the team to make the system more environmentally sustainable and cost-efficient without sacrificing required quality. Which action best aligns with responsible AI guidance?

Benchmark smaller or task-specific models against the accuracy requirement and use the smallest model that meets it for each task

Keep the largest model but raise its max output token limit so fewer calls are needed

Pre-train a new proprietary foundation model from scratch on the company's own data

Run the largest model on-premises instead of in the cloud to reduce its energy footprint

A governance team must continuously verify that no Amazon SageMaker AI notebook instance in the account has direct internet access enabled, and they want noncompliant notebooks flagged automatically in a compliance dashboard. Which approach achieves this with the least custom development?

Enable the AWS Config managed rule that checks whether SageMaker notebook instances allow direct internet access

Schedule a monthly script that lists all notebook instances and emails the results to the security team

Enable Amazon GuardDuty, which blocks notebook instances from being created with internet access

Turn on Amazon Inspector to scan the notebooks' network settings

An auditor asks a company to show which IAM principals invoked Amazon Bedrock foundation models over the past quarter and when. The team plans to rely on AWS CloudTrail. Which statement correctly describes what CloudTrail provides for this audit?

CloudTrail records the caller identity, timestamp, source IP, and API action for Bedrock calls, but it does not capture the prompt and completion text

CloudTrail captures the full prompt and the model's generated response for every invocation

CloudTrail records only AWS Management Console sign-ins, so API calls to Bedrock are not logged

CloudTrail does not support Amazon Bedrock, so the team must build a custom logging solution

During a regulatory audit, a bank must prove end-to-end provenance for each production model: which dataset location was used, which processing and training steps ran, which trial produced the model, and which endpoint serves it. Which Amazon SageMaker AI capability automatically creates and stores this graph of relationships between ML workflow entities?

Amazon SageMaker ML Lineage Tracking

Free AWS AI Practitioner (AIF-C01) Practice Questions

These are real exam-style AWS Certified AI Practitioner (AIF-C01) questions, each paired with a full explanation of why the correct answer is right and why the others fall short. They're completely free with no signup required, and every question links to the official AWS documentation that backs its answer so you can dig deeper whenever you want. 20 questions below, free with no signup. Last updated 2026-07-11.

Every question is verified against official AWS documentation, linked below each answer.

Question 1Fundamentals of AI and ML
A used-car marketplace wants to predict the exact sale price, in dollars, of each vehicle listed on its platform based on features such as mileage, age, brand, and condition. How should the machine learning problem be framed?
Reveal answer & explanation
Answer: B. As a regression problem that predicts a continuous numeric value
The target variable is a continuous dollar amount, which is the defining characteristic of a regression problem — the model learns to output a number on a continuous scale from the input features.
Go deeper: What would change if the marketplace only needed to predict a price range instead of an exact dollar amount? Why does the choice of continuous versus discrete target determine which algorithms and loss functions apply?
Source: official AWS documentation · verified 2026-07-11
Question 2Fundamentals of AI and ML
An insurance company receives thousands of scanned claim forms every day. It wants to automatically extract the printed text, tables, and form fields (key-value pairs such as 'Policy Number: 12345') from these document images without building or training any ML models. Which AWS service should the company use?
Reveal answer & explanation
Answer: C. Amazon Textract
Amazon Textract is the pre-trained AI service purpose-built for document processing, extracting structured data such as tables and form key-value pairs from scanned documents without any model training.
Go deeper: What would you use instead if the input were a photo of a street sign rather than a scanned form? Once Textract extracts the text, which service would you reach for to analyze sentiment or entities within it?
Source: official AWS documentation · verified 2026-07-11
Question 3Fundamentals of AI and ML
A retailer has a trained propensity model and needs to score its entire database of 20 million customers once per week. There is no requirement for low-latency, on-demand predictions, and the team wants to minimize cost and infrastructure management. Which inference option is the best fit?
Reveal answer & explanation
Answer: B. Run a SageMaker batch transform job on the customer dataset each week
Batch inference is designed for scoring large datasets on a schedule when no one needs an immediate answer; SageMaker batch transform spins up compute only for the job's duration, writes results to S3, and shuts down, so the retailer pays nothing between weekly runs.
Go deeper: How would the answer change if the retailer needed a prediction back within milliseconds of a customer visiting the site? What happens to cost if an always-on endpoint sits idle between weekly scoring runs?
Source: official AWS documentation · verified 2026-07-11
Question 4Fundamentals of AI and ML
A bank is evaluating a binary classification model that flags credit-card transactions as fraudulent or legitimate. The business states that failing to catch a fraudulent transaction costs far more than sending a legitimate transaction for manual review. Which evaluation metric should the team prioritize when tuning the model?
Reveal answer & explanation
Answer: C. Recall, because it measures how many actual fraud cases the model catches
Recall measures the fraction of real fraud the model catches, and maximizing it minimizes false negatives — the error type the business identified as most costly.
Go deeper: What would happen to recall if the model flagged every single transaction as fraudulent? How would the priority change if the business instead said manual reviews were the expensive part?
Source: official AWS documentation · verified 2026-07-11
Question 5Fundamentals of AI and ML
A manufacturing company tests a defect-detection classifier on 1,000 parts. The confusion matrix shows 40 true positives, 10 false positives, 20 false negatives, and 930 true negatives. What is the model's precision for the defect class?
Reveal answer & explanation
Answer: A. 0.80
Precision is TP/(TP+FP) = 40/(40+10) = 40/50 = 0.80 — when the model flags a part as defective, it is right 80% of the time.
Go deeper: Why does accuracy look so much higher than precision on this dataset? What would precision become if false positives doubled while true positives stayed the same?
Source: official AWS documentation · verified 2026-07-11
Question 6Applications of Foundation Models
A team is building a document search and question-answering application on Amazon Bedrock using the Amazon Titan model family. The application must convert documents into numerical vectors for similarity search and also generate natural-language answers for users. Which statement correctly describes how Titan models are used in this application?
Reveal answer & explanation
Answer: B. Amazon Titan Text Embeddings converts text into vectors for similarity search, while a Titan text generation model produces the natural-language answers.
Titan Text Embeddings converts text into dense numerical vectors for similarity search, while a separate Titan text generation model produces the natural-language answers — a typical search or RAG application uses two different models for these two jobs.
Go deeper: Why can't a single generation model efficiently search across thousands of documents for similar content? What role does the embeddings model play once the generation model has retrieved matching text?
Source: official AWS documentation · verified 2026-07-11
Question 7Applications of Foundation Models
Users of a customer-facing chatbot built on Amazon Bedrock complain that after submitting a question they stare at a blank screen for many seconds before the full answer suddenly appears. Answers are often several paragraphs long. Which change most directly improves the perceived responsiveness of the chatbot?
Reveal answer & explanation
Answer: C. Invoke the model with response streaming so tokens are displayed to the user as they are generated.
Response streaming returns the model's output in chunks as tokens are generated instead of waiting for the full completion, so users see the first words almost immediately, dramatically improving perceived responsiveness for long answers.
Go deeper: Why does streaming reduce perceived latency even if the total generation time stays the same? What would happen to time-to-first-token if the application kept waiting for the full response before displaying anything?
Source: official AWS documentation · verified 2026-07-11
Question 8Applications of Foundation Models
A legal team uses Amazon Bedrock Knowledge Bases over long contracts stored in Amazon S3. Retrieval returns relevant text, but answers are often incomplete because a clause and its surrounding definitions end up in separate retrieved fragments. Which adjustment to the knowledge base is most likely to fix this?
Reveal answer & explanation
Answer: A. Change the chunking configuration used during ingestion so more surrounding context is kept together in each chunk.
The chunking configuration set during ingestion determines how much surrounding context travels with each retrievable unit, so adjusting it to keep related clauses and definitions together directly fixes fragmented retrieval.
Go deeper: At what stage of the pipeline — ingestion or generation — does chunk fragmentation actually occur? What tradeoff might larger chunks introduce even as they preserve more context?
Source: official AWS documentation · verified 2026-07-11
Question 9Applications of Foundation Models
A company has customized a foundation model in Amazon Bedrock and now needs to serve it in production, where traffic is steady and high-volume and the business requires guaranteed model throughput. Which consumption option should they choose?
Reveal answer & explanation
Answer: B. Purchase Provisioned Throughput for the model, committing to a number of model units.
Provisioned Throughput lets you purchase model units and receive a guaranteed level of throughput, fitting steady high-volume production workloads, and it is also the required mechanism for serving customized models in Bedrock.
Go deeper: Why can't a customized model be served through shared on-demand capacity? What tradeoff does a business accept by committing to provisioned model units instead of paying per token?
Source: official AWS documentation · verified 2026-07-11
Question 10Fundamentals of Generative AI
A designer uses an image generation model on Amazon Bedrock to create product artwork. The generated images keep including unwanted elements such as text overlays and watermark-like patterns. The designer wants to explicitly tell the model what should NOT appear in the output, in addition to describing what should appear. Which prompting technique accomplishes this?
Reveal answer & explanation
Answer: A. Negative prompting — supplying a separate instruction listing elements the model should avoid generating
Negative prompting lets the user specify concepts or elements the model should exclude from its output alongside the positive prompt, and Bedrock image models like Titan Image Generator expose a negativeText parameter for exactly this purpose.
Go deeper: What would happen to the watermark problem if the designer only fixed the seed without changing the prompt? How does negative prompting differ from simply describing the desired image in more detail?
Source: official AWS documentation · verified 2026-07-11
Question 11Fundamentals of Generative AI
A developer is tuning a text-generation model on Amazon Bedrock. At each generation step, the developer wants the model to sample only from the smallest set of candidate tokens whose cumulative probability exceeds a threshold such as 0.9, so that the size of the candidate pool shrinks when the model is confident and grows when it is uncertain. Which inference parameter controls this behavior?
Reveal answer & explanation
Answer: A. Top-p
Top-p (nucleus sampling) restricts sampling to the smallest set of most-likely tokens whose cumulative probability meets a threshold, so the candidate pool adapts dynamically — small when one token dominates, larger when many tokens are plausible.
Go deeper: How would the candidate pool size differ between top-p and top-k when the model is very confident in one token? What would happen to output diversity if top-p were set close to 1.0?
Source: official AWS documentation · verified 2026-07-11
Question 12Fundamentals of Generative AI
A retail company is prioritizing candidate projects for its first generative AI initiatives. Which project is the LEAST appropriate fit for a generative AI solution?
Reveal answer & explanation
Answer: B. Computing the exact sales tax owed on each customer transaction, where results must be precise, deterministic, and auditable
Generative AI models produce probabilistic output rather than exact calculations, so tasks demanding precise, deterministic, auditable answers — like computing exact sales tax — are best handled by conventional rule-based software, making this the least appropriate generative AI use case.
Go deeper: Why does probabilistic token prediction make exact deterministic math risky for a generative model? What kind of system would you pair with a generative model to handle the tax calculation reliably?
Source: official AWS documentation · verified 2026-07-11
Question 13Fundamentals of Generative AI
A business analyst with no coding experience wants to quickly prototype a generative AI app idea — such as a marketing-slogan generator — experiment with prompts, and share the app with colleagues, all without writing code or provisioning AWS infrastructure. Which AWS offering is designed for this kind of hands-on experimentation?
Reveal answer & explanation
Answer: A. PartyRock, an Amazon Bedrock Playground
PartyRock is AWS's shareable, no-code generative AI playground built on Amazon Bedrock, letting anyone build, remix, and share small generative AI apps and experiment with prompts through a web interface with no infrastructure or code required.
Go deeper: Why does a no-code playground matter specifically for a business analyst without a development background? What would the analyst need to add before this same idea became a production application?
Source: official AWS documentation · verified 2026-07-11
Question 14Guidelines for Responsible AI
A compliance officer is comparing two documentation mechanisms: AWS AI Service Cards and Amazon SageMaker Model Cards. Which statement correctly describes the key difference between them?
Reveal answer & explanation
Answer: A. AI Service Cards are published by AWS to describe intended use cases, limitations, and responsible design of AWS's own AI services, while SageMaker Model Cards are created by customers to document their own models
AWS AI Service Cards are transparency documentation authored and published by AWS describing intended use cases, limitations, and responsible design of AWS's own AI services, while SageMaker Model Cards are a tool customers use to document their own models' training data, evaluation results, and intended uses.
Go deeper: Who would you consult to understand the limitations of Amazon Rekognition itself versus limitations of a model your own team built? Why might a regulator ask for both types of documentation on the same AI application?
Source: official AWS documentation · verified 2026-07-11
Question 15Guidelines for Responsible AI
A regulated insurer is building a premium-pricing model. Regulators require that the company be able to trace exactly how each input feature contributes to any individual decision, directly from the model's own structure rather than from an approximation. Accuracy requirements are modest. Which modeling choice best satisfies this requirement?
Reveal answer & explanation
Answer: B. Use an inherently interpretable model such as linear regression or a decision tree, accepting a possible small trade-off in accuracy
Because the regulator requires that contributions be traceable directly from the model's own structure, an inherently interpretable model such as linear regression or a decision tree is the right choice, and the modest accuracy requirement makes the interpretability-versus-accuracy trade-off acceptable.
Go deeper: What's the difference between a model being interpretable by design versus explainable after the fact? How would the answer change if the regulator only needed aggregate fairness statistics rather than per-decision traceability?
Source: official AWS documentation · verified 2026-07-11
Question 16Guidelines for Responsible AI
A voice assistant's speech-recognition model was trained mostly on recordings of adult speakers from a single geographic region. After launch, error rates are much higher for elderly users and speakers with regional accents. Which data practice most directly addresses the root cause before the model is retrained?
Reveal answer & explanation
Answer: A. Curate a training dataset that is representative of the full deployment population, collecting additional samples from the underserved age groups and accents
The root cause is a non-representative training dataset, so the remedy is curating a dataset that is representative of the full deployment population by collecting additional samples from underserved age groups and accents.
Go deeper: Why can't more training epochs compensate for a demographic group being absent from the data in the first place? What sampling strategy would ensure each subgroup is adequately represented going forward?
Source: official AWS documentation · verified 2026-07-11
Question 17Guidelines for Responsible AI
A team routes every task in its application — including simple sentiment classification of short reviews — to the largest, most capable foundation model available. Leadership asks the team to make the system more environmentally sustainable and cost-efficient without sacrificing required quality. Which action best aligns with responsible AI guidance?
Reveal answer & explanation
Answer: A. Benchmark smaller or task-specific models against the accuracy requirement and use the smallest model that meets it for each task
Right-sizing the model to the task is a core sustainability and cost practice — benchmarking smaller or task-specific models against the accuracy requirement and using the smallest one that meets it reduces environmental impact while preserving quality.
Go deeper: Why might a small, task-specific model match a giant foundation model's accuracy on simple sentiment classification? What would you measure to confirm a smaller model still meets the required quality bar?
Source: official AWS documentation · verified 2026-07-11
Question 18Security, Compliance & Governance for AI
A governance team must continuously verify that no Amazon SageMaker AI notebook instance in the account has direct internet access enabled, and they want noncompliant notebooks flagged automatically in a compliance dashboard. Which approach achieves this with the least custom development?
Reveal answer & explanation
Answer: A. Enable the AWS Config managed rule that checks whether SageMaker notebook instances allow direct internet access
AWS Config continuously evaluates resource configurations against rules and ships a managed rule that marks any SageMaker notebook instance with direct internet access enabled as non-compliant, surfacing results in a compliance dashboard with essentially zero custom code.
Go deeper: Why does a continuously evaluated managed rule beat a scheduled script for ongoing compliance monitoring? What other AWS service could aggregate this compliance data across multiple accounts?
Source: official AWS documentation · verified 2026-07-11
Question 19Security, Compliance & Governance for AI
An auditor asks a company to show which IAM principals invoked Amazon Bedrock foundation models over the past quarter and when. The team plans to rely on AWS CloudTrail. Which statement correctly describes what CloudTrail provides for this audit?
Reveal answer & explanation
Answer: A. CloudTrail records the caller identity, timestamp, source IP, and API action for Bedrock calls, but it does not capture the prompt and completion text
Amazon Bedrock is integrated with CloudTrail, which captures API activity — caller identity, timestamp, source IP, and action — letting the auditor attribute model usage to principals, but it records API metadata, not the prompt and completion payloads.
Go deeper: If the auditor also needed the actual prompt text, what additional logging capability would be required? Why does separating metadata logging from content logging matter for both security and privacy?
Source: official AWS documentation · verified 2026-07-11
Question 20Security, Compliance & Governance for AI
During a regulatory audit, a bank must prove end-to-end provenance for each production model: which dataset location was used, which processing and training steps ran, which trial produced the model, and which endpoint serves it. Which Amazon SageMaker AI capability automatically creates and stores this graph of relationships between ML workflow entities?
Reveal answer & explanation
Answer: A. Amazon SageMaker ML Lineage Tracking
SageMaker ML Lineage Tracking creates and stores a graph of ML workflow entities — artifacts, actions, and their associations — letting the bank traverse from a deployed model back through training and processing steps to the exact input data, which is the provenance evidence auditors require.
Go deeper: How would an auditor use the lineage graph to trace a deployed endpoint back to its original training dataset? Why is a monitoring tool like Model Monitor insufficient for answering a provenance question?
Source: official AWS documentation · verified 2026-07-11

AIF-C01 exam — common questions

How many questions are on the AIF-C01 exam?

The AWS Certified AI Practitioner (AIF-C01) exam has 65 questions total: 50 scored questions and 15 unscored questions AWS uses for its own evaluation. You won't know which is which, so it pays to treat every question as if it counts. You have 90 minutes to complete the exam.

What score do I need to pass the AIF-C01 exam?

You need a scaled score of 700 out of a possible range of 100 to 1000 to pass. AWS scales raw scores to account for slight difficulty differences across exam forms, so 700 doesn't simply mean 70% of questions answered correctly.

How much does the AIF-C01 exam cost, and are there any prerequisites?

The exam costs $100 USD. There are no prerequisites and no coding is required, which makes AIF-C01 AWS's most accessible certification for demonstrating AI and ML literacy without a technical background.

How hard is the AIF-C01 exam, and how long should I prepare?

Most candidates prepare for about 2-3 weeks. The exam covers five domains: fundamentals of AI and ML, fundamentals of generative AI, applications of foundation models, guidelines for responsible AI, and security, compliance, and governance for AI.

More practice sets: Cloud Practitioner (CLF-C02) · Developer (DVA-C02) · Solutions Architect (SAA-C03)