AWS AIF-C01 Practice Test: AI Practitioner

AIF-C01 is AWS’s AI Practitioner certification for candidates who need practical AI, generative AI, service-selection, and responsible-AI judgment without going deep into custom model building. If you are searching for AIF-C01 sample questions, a practice test, mock exam, or exam simulator, this is the main IT Mastery page to start on web and continue on iOS or Android with the same IT Mastery account.

Free diagnostic: Try the 65-question AWS AIF-C01 full-length practice exam before subscribing. Use it as one AI Practitioner baseline, then return to IT Mastery for timed mocks, topic drills, explanations, and the full AWS AI question bank.

What this AIF-C01 practice page gives you

• a direct route into IT Mastery practice for AIF-C01
• topic drills and mixed sets across AI/ML fundamentals, generative AI, responsible AI, and governance
• detailed explanations that show why the best AWS AI answer is correct
• a clear free-preview path before you subscribe
• the same IT Mastery account across web and mobile

AIF-C01 exam snapshot

• Vendor: AWS
• Official exam name: AWS Certified AI Practitioner (AIF-C01)
• Exam code: AIF-C01
• Items: 65 total, including 50 scored and 15 unscored
• Exam time: 90 minutes
• Question types: multiple choice, multiple response, ordering, and matching
• Passing score: 700 scaled

AIF-C01 questions usually reward the option that applies the right AWS AI service, understands generative AI limits, and protects responsible-AI and governance expectations without overengineering the solution.

Topic coverage for AIF-C01 practice

AIF-C01 AI decision filters

Use these filters when two AWS AI answers sound reasonable:

• Use case before service: identify whether the task is prediction, classification, transcription, translation, document extraction, search, summarization, generation, or agentic workflow automation.
• GenAI vs traditional ML: choose foundation-model patterns for generation and reasoning tasks, and traditional ML patterns for structured prediction, forecasting, and classification.
• RAG and grounding: look for retrieval, embeddings, vector search, knowledge bases, and source attribution when the scenario needs current or private information.
• Responsible AI: watch for bias, toxicity, hallucination, human review, explainability, privacy, transparency, and safe-use constraints.
• Governance boundary: match IAM, encryption, auditability, data residency, compliance, and model-access controls to the sensitivity of the workload.

AIF-C01 readiness map

How to use the AIF-C01 simulator efficiently

1. Start with domain drills so you can separate AI basics from AWS-specific service choices.
2. Review every miss until you can explain why the best answer fits the use case, the data, and the governance constraint.
3. Move into mixed sets once you can switch between model types, prompt patterns, RAG, guardrails, and compliance scenarios without hesitation.
4. Finish with timed runs so the 90-minute pace feels normal before test day.

Final 7-day AIF-C01 practice sequence

When AIF-C01 practice is enough

If you can score above roughly 75% on several unseen mixed attempts and explain why each answer fits the use case, service boundary, and governance constraint, further drilling may produce memorization more than readiness. At that point, schedule the exam and keep the final review light.

Focused sample questions

Use these child pages when you want focused IT Mastery practice before returning to mixed sets and timed mocks.

• Free AWS AIF-C01 Full-Length Practice Exam
• Fundamentals of AI and ML
• Fundamentals of Genai
• Applications of Foundation Models
• Guidelines for Responsible AI
• AI Security and Governance

Free study resources

Need concept review first? Read the AWS AIF-C01 Cheat Sheet on Tech Exam Lexicon, then return here for timed mocks, topic drills, and full IT Mastery practice.

Free preview vs premium

• Free preview: a smaller web set so you can validate the question style and explanation depth.
• Premium: the full AIF-C01 practice bank, focused drills, mixed sets, timed mock exams, detailed explanations, and progress tracking across web and mobile.

24 AIF-C01 sample questions with detailed explanations

These are original IT Mastery practice questions aligned to AI and ML fundamentals, generative AI, foundation models, responsible AI, AWS AI services, and business-use-case decisions. They are not AWS exam questions and are not copied from any exam sponsor. Use them to check readiness here, then continue in IT Mastery with mixed sets, topic drills, and timed mocks.

Question 1

Topic: Content Domain 2: Fundamentals of GenAI

Which option best describes a GenAI agent use case (for example, using Agents for Amazon Bedrock) rather than a simple Q&A assistant?

• A. Answers HR policy questions from an uploaded PDF
• B. Classifies customer emails into support categories
• C. Automatically completes a refund by calling business APIs
• D. Forecasts next month’s product demand from sales history

Best answer: C

Explanation: A GenAI agent is designed to go beyond answering questions by selecting tools and taking actions to achieve a user goal. The best example is a workflow where the model calls approved business APIs (for example, ticketing or payments) to complete a multi-step task on the user’s behalf.

The core concept is the difference between an assistant and an agent. A GenAI assistant primarily generates text responses (Q&A, summarization, drafting) and may use retrieval (RAG) to ground answers in enterprise content. A GenAI agent adds “action-taking” behavior: it can break a request into steps, decide which tools to use, and invoke APIs or workflows to complete tasks with guardrails and permissions. In AWS, this pattern is commonly implemented with Agents for Amazon Bedrock, where the model orchestrates tool calls to perform operations like checking order status, creating tickets, or processing returns. The key distinguishing feature is autonomous tool use to accomplish an outcome, not just producing an answer.

Question 2

Topic: Content Domain 1: Fundamentals of AI and ML

A team is using Amazon SageMaker AI to build a binary classification model. The team is writing a standard operating procedure (SOP) that describes what happens during model training, hyperparameter tuning, and final model evaluation, along with basic governance controls.

Which TWO SOP statements are INCORRECT (anti-patterns or unsafe practices)? (Select TWO)

• A. Tune hyperparameters by comparing multiple runs using a validation set, and record each run’s hyperparameters and metrics
• B. Use the test set during hyperparameter tuning to choose the best hyperparameters faster
• C. Give all developers broad access like s3:* on the dataset bucket to avoid permission issues
• D. Evaluate the final selected model once on a held-out test set and register the model for an approval step before deployment
• E. Encrypt datasets and model artifacts with AWS KMS and ensure access is auditable with AWS CloudTrail
• F. Fit model parameters on a training set and keep the test set untouched until the end

Best answers: B, C

Explanation: Training fits model parameters (weights) using the training data, hyperparameter tuning selects configuration values using validation results, and evaluation measures the final model on an untouched test set. Good governance also requires traceability of experiments and least-privilege access to sensitive datasets. The incorrect statements either compromise evaluation integrity or weaken security controls.

In the ML lifecycle, model training updates the model’s learned parameters using the training dataset. Hyperparameter tuning runs repeated training jobs with different settings (for example, learning rate, regularization strength) and chooses the best configuration based on validation performance. Final model evaluation should be performed on a separate, held-out test set exactly once to estimate real-world performance.

From a governance perspective, you should preserve provenance (track datasets, hyperparameters, metrics, and model versions) and enforce least privilege for access to training and evaluation data. Using the test set during tuning leaks evaluation information into model selection and invalidates the trustworthiness of reported test metrics; overly broad permissions increase the chance of accidental or malicious data exposure or tampering.

Question 3

Topic: Content Domain 1: Fundamentals of AI and ML

Which scenario is the best fit for using speech recognition to create a text transcript?

• A. Detect anomalous network traffic patterns from VPC Flow Logs
• B. Translate an existing text document from English to Spanish
• C. Automatically convert customer support call audio into text for analytics
• D. Identify objects and scenes in recorded security camera footage

Best answer: C

Explanation: Speech recognition is used to convert spoken audio into written text. Creating transcripts from call recordings is a common application because the resulting text can be searched, summarized, and analyzed (for example, for sentiment or compliance checks). The other scenarios involve computer vision, anomaly detection on log data, or text translation rather than transcription.

Speech recognition (also called automatic speech recognition, ASR) converts human speech in an audio stream or recording into text. Transcription is valuable when you need a written record of spoken interactions, such as contact center calls, meetings, or videos, because text is easier to index, search, analyze for themes, and use for captions/subtitles. In AWS, services like Amazon Transcribe provide managed speech-to-text capabilities for these use cases. Tasks like object detection in video, anomaly detection in network logs, or language translation operate on different data modalities (images/video, structured logs, or existing text) and therefore are not speech recognition problems.

Key takeaway: choose speech recognition when the input is audio speech and the desired output is a text transcript.

Question 4

Topic: Content Domain 2: Fundamentals of GenAI

Which TWO statements about choosing a foundation model’s modality are INCORRECT? (Select TWO)

• A. Text-only models are best for transcribing audio
• B. Use a multimodal model when prompts include images
• C. Text-only models can take images as direct input
• D. Audio tasks may need speech-to-text before a text model
• E. Image generation requires an image-capable model
• F. Multimodal models can accept mixed text and images

Best answers: A, C

Explanation: Model modality must match the input and output types your use case requires. Text-only models cannot natively process audio or images, so those modalities require either multimodal models or pre-/post-processing with specialized services. The incorrect statements are the ones claiming text-only models can directly handle images or audio transcription.

A key model selection factor is modality: what the model can accept as input and produce as output. If your use case includes non-text content, you must choose an image/audio/multimodal-capable model, or convert the content into text before using a text-only model.

• The claim that text-only models can take images directly is incorrect; that requires a multimodal model (or extracting text from the image first).
• Using a multimodal model when prompts include images is true because the model can consume both text and images.
• The statement that audio tasks may need speech-to-text before a text model is true when you only have a text model.
• The claim that text-only models are best for transcribing audio is incorrect; transcription needs audio input handling (for example, speech-to-text).
• The statement that image generation requires an image-capable model is true.
• The statement that multimodal models can accept mixed text and images is true.

Always align modality to the data you actually have and the output you need.

Question 5

Topic: Content Domain 3: Applications of Foundation Models

Which TWO statements accurately describe Amazon Bedrock Model Evaluation and when it is used? (Select TWO.)

• A. It automatically deploys the top-scoring model to production endpoints
• B. It compares foundation models using your prompts and chosen metrics
• C. It can only evaluate models that you have custom-trained in SageMaker
• D. It fine-tunes a foundation model as part of the evaluation job
• E. It is used to help select a model before deploying a GenAI use case
• F. It evaluates only cost and latency, not output quality

Best answers: B, E

Explanation: Amazon Bedrock Model Evaluation is used to assess and compare foundation model performance for a specific use case before choosing a model. You provide prompts (and optionally reference answers) and select evaluation approaches/metrics to understand which model best meets your quality requirements.

Amazon Bedrock Model Evaluation is a feature for evaluating and comparing foundation models on your workload, typically during model selection and iteration (for example, before launching a summarization or Q&A application). At a high level, you run evaluation using your prompts/datasets and chosen evaluation approach to measure output quality and help decide which model is the best fit.

It is not a training or deployment feature: it does not fine-tune models as part of evaluation, it is not limited to models you trained yourself, and it does not automatically push a “winner” to production. The key takeaway is that Model Evaluation supports structured comparison of model quality for your specific task so you can pick an appropriate FM with evidence.

Question 6

Topic: Content Domain 2: Fundamentals of GenAI

A company is building a retrieval-augmented generation (RAG) chatbot using Amazon Bedrock. The team generates embeddings for a user query and searches a vector index.

Exhibit: Vector search results (cosine similarity)

Which option best interprets the exhibit in terms of vectors and vector similarity search?

• A. Embeddings are class labels; retrain the model using chunk-09
• B. Embeddings are numeric vectors; retrieve chunk-17 as the nearest match
• C. Embeddings are keywords; retrieve chunk-04 because it is shortest
• D. Embeddings are encrypted text; select a chunk at random to avoid bias

Best answer: B

Explanation: An embedding represents text as a numeric vector in a high-dimensional space. Vector similarity search compares the query vector to stored vectors using a metric such as cosine similarity and returns the closest matches. Because the exhibit’s highest similarity is 0.82, that item is the best retrieval candidate.

Vectors (embeddings) are arrays of numbers that represent the meaning of text so that semantically similar text ends up “near” each other in vector space. Vector similarity search works by comparing a query embedding to stored embeddings with a distance/similarity metric (here, cosine similarity) and returning the nearest neighbors.

In the exhibit, the system should pick the document chunk with the largest cosine similarity value because that vector is closest to the query vector:

• chunk-17 has 0.82
• chunk-09 has 0.76
• chunk-04 has 0.31

The key takeaway is that the score is produced by comparing vectors, not by keyword matching or model retraining.

Question 7

Topic: Content Domain 4: Guidelines for Responsible AI

A company is building a customer-support chatbot using an Amazon Bedrock foundation model. The company needs a managed control that can screen both user prompts and model responses to help prevent harmful content and to detect/redact sensitive data (for example, PII) without changing the underlying model.

Which solution meets this requirement?

• A. Use AWS KMS encryption for prompts and responses at rest
• B. Use prompt instructions that tell the model to avoid unsafe topics
• C. Use IAM policies to allow only bedrock:InvokeModel for the chatbot role
• D. Use Amazon Bedrock Guardrails with content and sensitive-data filters

Best answer: D

Explanation: Amazon Bedrock Guardrails are designed to apply policy-based controls to both incoming prompts and outgoing model responses. They can filter categories of harmful content and detect or redact sensitive information such as PII. This directly supports responsible AI by constraining inputs and outputs without changing the foundation model.

The discriminating requirement is a managed mechanism that enforces safety constraints on both the user input and the model output. Amazon Bedrock Guardrails provide centralized, configurable policies (such as content filters, topic restrictions, and sensitive-information/PII handling) that are evaluated on prompts and on generated responses.

Encryption and IAM address security and access control, but they do not evaluate text for harmful or sensitive content. Prompt-only instructions can reduce risk, but they are not reliable enforcement because the model can still produce disallowed content under some inputs. The key takeaway is to use Guardrails when you need policy enforcement over model inputs and outputs.

Question 8

Topic: Content Domain 3: Applications of Foundation Models

A company uses an Amazon Bedrock text model to generate summaries of customer support tickets. Different agents type their own prompts, and the downstream application fails when the model output is not consistent.

Exhibit: Current workflow

1 User prompt varies (no template)
2 Example output A: "Summary:... Next steps:..."
3 Example output B: "- Issue...\n- Action..."
4 Example output C: "{ 'sum':..., 'priority': 'high' }"
5 Requirement: downstream app expects JSON keys:
6 summary, priority, customer_impact, next_action

Based on the exhibit, what is the best next step to standardize outputs across users and workflows?

• A. Increase the model temperature to make outputs more consistent
• B. Create a prompt template that enforces a fixed JSON schema with those keys
• C. Add a retrieval-augmented generation (RAG) knowledge base for tickets
• D. Fine-tune the foundation model on past ticket summaries

Best answer: B

Explanation: The exhibit shows inconsistent formats across agents (lines 1-4) while the downstream app requires specific JSON keys (lines 5-6). The most direct way to standardize output is to use a shared prompt template that fixes the instructions and output structure (for example, “return only JSON with these keys”) and parameterizes only the ticket content.

Prompt templates standardize results by separating variable user inputs from the fixed instructions and response format the model must follow. In the exhibit, outputs vary widely because agents use different prompts (line 1), producing different structures (lines 2-4). Meanwhile, the consuming application expects a specific schema with four JSON keys (lines 5-6).

A best-practice next step is to define a single reusable prompt template that:

• Inserts the ticket text into a placeholder (for example, {{ticket_text}}).
• Explicitly requires “return only valid JSON” with exactly summary, priority, customer_impact, and next_action.

This addresses the formatting failure directly without changing the model or adding extra data sources.

Question 9

Topic: Content Domain 5: Security, Compliance, and Governance for AI Solutions

A company collects customer chat transcripts to evaluate a generative AI assistant. The company defines how long each data type is kept (for example, raw transcripts for 30 days, redacted aggregates for 1 year) and automates transition and expiration using storage lifecycle rules to ensure timely deletion.

Which governance principle does this practice MOST directly demonstrate?

• A. Least privilege access control
• B. Transparency and explainability
• C. Defense in depth security
• D. Data lifecycle governance (retention and deletion policies)

Best answer: D

Explanation: This practice is about governing data from creation through storage, retention, and end-of-life deletion. Defining retention durations by data type and automating expiration are core data governance lifecycle strategies that reduce risk and support compliance expectations for AI systems.

The core principle is data lifecycle governance: managing data from collection to disposal. In AI solutions, this includes defining what data is stored, how long it is retained, and how it is securely deleted when it is no longer needed. Automating retention and deletion (for example, with storage lifecycle/expiration rules) helps ensure policies are consistently enforced, supports compliance requirements, and reduces exposure of sensitive training/evaluation data. This is distinct from security access controls or model interpretability, which address different risk areas.

Question 10

Topic: Content Domain 2: Fundamentals of GenAI

A company deploys a GenAI-powered virtual agent (using Amazon Bedrock) to help customers resolve support issues before creating a ticket. The team agrees to evaluate the rollout using average issue resolution time, self-service deflection rate, and customer satisfaction (CSAT).

Which principle is the team applying?

• A. Define outcome-based success metrics aligned to the business goal
• B. Increase robustness by stress-testing with adversarial prompts
• C. Apply least privilege access to the model and data
• D. Improve transparency by explaining model outputs to users

Best answer: A

Explanation: The team is focusing on how to measure whether the GenAI assistant improves the support experience. Resolution time, deflection rate, and CSAT are outcome-oriented KPIs that indicate efficiency, reduced agent workload, and customer sentiment. This reflects the principle of defining business-aligned success metrics for the use case.

For a customer support GenAI use case, you should evaluate success with metrics that reflect the desired business outcomes. Resolution time indicates how quickly issues are solved, deflection rate indicates how often self-service avoids creating a ticket or contacting an agent, and CSAT measures customer-perceived quality of the experience.

Defining these KPIs up front helps you:

• Set a baseline and measure change after rollout
• Compare versions (prompt changes, different models) against the same goals
• Decide whether the solution is meeting business expectations

Other principles like least privilege, transparency, and robustness are important, but they don’t directly represent the act of selecting and using these support KPIs.

Question 11

Topic: Content Domain 2: Fundamentals of GenAI

A company is building an internal HR assistant using Amazon Bedrock. The team wants to steer the foundation model to (1) answer only from an approved HR policy excerpt provided in the prompt, (2) refuse requests for employee PII, and (3) return answers in a consistent JSON format for downstream processing.

Which TWO actions are examples of prompt engineering to achieve these goals? (Select TWO.)

• A. Use IAM least privilege for bedrock:InvokeModel and encrypt the HR documents with AWS KMS.
• B. Add a system prompt that defines the assistant’s role, refusal rules for PII, and “answer only from the provided excerpt.”
• C. Enable Bedrock Guardrails to block PII and enforce topic restrictions.
• D. Include a few-shot example showing the required JSON keys and an example refusal response.
• E. Turn on AWS CloudTrail data events for Bedrock and store logs for audits.
• F. Switch to a different foundation model that has lower latency and cost.

Best answers: B, D

Explanation: Prompt engineering means designing and iterating on prompts (instructions, context, and examples) to steer a model’s behavior and outputs. Using clear system-level instructions and providing examples of the desired JSON and refusal patterns are direct prompt changes that shape how the model responds. Security, governance, and model selection are important, but they are not prompt engineering techniques.

Prompt engineering is the practice of crafting the input prompt (instructions, context, formatting constraints, and examples) to steer an FM’s responses without changing the model weights. In this scenario, the goal is to constrain the assistant to an approved excerpt, refuse PII-related requests, and produce consistent structured output; those are best addressed by prompt instructions and in-prompt examples.

Key takeaway: prompt engineering changes what you ask and show the model to influence outputs; other controls wrap or govern the system but don’t redesign the prompt.

Question 12

Topic: Content Domain 5: Security, Compliance, and Governance for AI Solutions

A team runs a generative AI customer-support app on AWS. The application stores chat transcripts in Amazon S3 and must meet an internal compliance requirement to continuously prove that data stores remain encrypted and not publicly accessible.

During a review, the team discovers that an S3 bucket briefly became public, but they have no reliable configuration history or automated compliance evaluation to detect and document this drift.

Which action addresses the root cause?

• A. Enable AWS Config to record resource configurations and evaluate them with managed rules
• B. Enable Amazon Macie to discover and classify sensitive data in the S3 bucket
• C. Enable Amazon GuardDuty to detect anomalous or suspicious access to the S3 bucket
• D. Enable AWS CloudTrail to capture S3 API calls for auditing bucket access

Best answer: A

Explanation: The issue is missing continuous configuration recording and compliance evaluation for AWS resources. AWS Config provides configuration history (configuration items) and can automatically evaluate resources against rules such as S3 public access and encryption settings. This produces ongoing compliance status and evidence, not just point-in-time findings.

This is a governance/compliance gap: the team cannot demonstrate what the S3 bucket configuration was over time or automatically assess whether it stayed compliant. AWS Config is designed to record resource configuration changes (creating a history) and to evaluate those configurations against compliance requirements using AWS Config Rules or conformance packs. In this scenario, Config can track when the bucket’s public access settings changed and continuously evaluate encryption and public access posture, generating a compliance timeline and alerts. Logging or threat-detection services help with auditing access or detecting suspicious activity, but they do not provide continuous configuration state and compliance evaluation as the primary control.

Question 13

Topic: Content Domain 5: Security, Compliance, and Governance for AI Solutions

Which statement best describes AWS Key Management Service (AWS KMS) and how it supports encryption key management for AI solutions on AWS?

• A. AWS KMS lets you create and manage encryption keys and control their use for encrypting data used by AI services.
• B. AWS KMS replaces IAM by providing authentication and authorization for all AI service API calls.
• C. AWS KMS trains foundation models using encrypted datasets without requiring any other AWS service.
• D. AWS KMS automatically labels training data and stores the labels in encrypted form for model training.

Best answer: A

Explanation: AWS KMS is a managed service for creating and managing encryption keys and defining who can use them through key policies and IAM. In AI solutions, KMS keys are commonly used to encrypt data at rest (such as datasets, prompts, and outputs) and to control and audit key usage.

AWS KMS provides centralized encryption key management in AWS. You can create AWS managed keys or customer managed keys and then use them with integrated AWS services to encrypt and decrypt data used in AI solutions (for example, data stored in Amazon S3 or encrypted volumes and databases that store model inputs and outputs). Access to use a KMS key is controlled with KMS key policies and IAM permissions, which helps enforce least privilege. KMS also supports auditing of key usage through AWS logging, helping with governance and compliance evidence. KMS manages key material lifecycle and helps you avoid building and operating your own key management infrastructure.

Question 14

Topic: Content Domain 3: Applications of Foundation Models

A company wants an LLM-based assistant to draft responses to customer disputes in a highly regulated insurance domain. A team proposes training a new model from scratch on Amazon SageMaker AI, but they have only 8,000 labeled examples, a 4-week deadline, and a limited GPU budget. Compliance also prohibits creating a persistent vector index for retrieval (no embeddings store).

Which change best optimizes cost and time while still adapting the model to the insurance domain?

• A. Increase sampling temperature to improve domain-specific accuracy
• B. Train a new transformer model from scratch on the 8,000 examples
• C. Fine-tune a pretrained foundation model on the domain examples
• D. Use RAG with a new vector database built from policy documents

Best answer: C

Explanation: Transfer learning adapts a pretrained model to a new domain by fine-tuning on domain-specific examples, which usually needs much less data and compute than training from scratch. Given the small labeled dataset, tight timeline, and limited GPU budget, fine-tuning is the most practical way to achieve domain adaptation while controlling cost and delivery time.

Domain adaptation is the goal of making a model perform well on a specific target domain (here, regulated insurance language and dispute-response style). Transfer learning is the common way to do this in practice: start from a pretrained foundation model that already learned general language patterns, then fine-tune it on representative domain examples so it learns domain terminology, tone, and decision patterns with far less compute and data.

In this scenario, fine-tuning optimizes:

• Cost and time (reuse pretrained weights instead of full training)
• Feasibility with only 8,000 labeled examples
• Operational effort compared with building and validating a model from scratch

Because a persistent vector index is prohibited, approaches that rely on retrieval-backed embeddings don’t meet the constraints; fine-tuning adapts the model without needing an embeddings store at inference time.

Question 15

Topic: Content Domain 2: Fundamentals of GenAI

A company is building a GenAI assistant on AWS. Users can submit chat text, screenshots of error dialogs, and short voice messages. The team will use Amazon Bedrock and wants to choose appropriate model modalities while following good governance.

Which TWO approaches are INCORRECT? (Select TWO.)

• A. Grant one shared IAM role bedrock:* on * so all developers can use any model
• B. Use Amazon Textract to extract text from screenshots when only OCR is needed
• C. Transcribe voice messages to text, then use a text model for summarization
• D. Use a multimodal foundation model to interpret text plus screenshots
• E. Apply Bedrock Guardrails to reduce sensitive-data leakage in responses
• F. Store prompts, transcripts, and screenshots with customer PII in a public, unencrypted S3 bucket for debugging

Best answers: A, F

Explanation: Modality selection should match the input type (text, image, audio, or multimodal) while still enforcing basic security and governance controls. Broad access to all Bedrock actions and storing sensitive inputs in public, unencrypted storage are unsafe practices that increase the risk of unauthorized use and data leakage.

When selecting GenAI model modalities, align the model (or preprocessing) to the data type: text models for text, transcription before text modeling for audio, and multimodal models (or OCR services) for images. In AWS, you can combine managed services (for example, Amazon Transcribe or Amazon Textract) with Amazon Bedrock models to fit the modality needed.

Governance must be applied regardless of modality:

• Use least-privilege IAM permissions scoped to required Bedrock actions and specific models.
• Protect sensitive prompts/inputs/outputs with strong access controls and encryption (for example, S3 SSE-KMS) and avoid public exposure.
• Add safety controls such as Bedrock Guardrails to reduce leakage of sensitive data.

The key takeaway is that correct modality choices don’t justify weakening access control or data protection.

Question 16

Topic: Content Domain 2: Fundamentals of GenAI

A company built an internal HR chatbot using an Amazon Bedrock foundation model. The team currently runs a weekly fine-tuning job so the model “learns” the latest policy PDFs stored in Amazon S3.

Constraints: Policies change multiple times per day, updates must be reflected within hours, operational effort and cost must decrease, and the solution must reduce hallucinations by grounding answers in approved content.

Which change best optimizes the solution while meeting all constraints?

• A. Increase temperature and max tokens to improve answer quality
• B. Fine-tune the model daily with the latest policy documents
• C. Pre-train a new model from scratch on all HR documents
• D. Use Bedrock Knowledge Bases for RAG over S3 documents

Best answer: D

Explanation: Replacing frequent fine-tuning with retrieval-augmented generation (RAG) lets the application use updated S3 documents immediately while keeping the underlying model unchanged. This shifts the problem from the fine-tuning stage to data selection and deployment-time retrieval, which typically lowers operational effort and cost. Grounding responses on retrieved, approved passages also reduces hallucinations.

In the foundation model lifecycle, pre-training and fine-tuning are model adaptation stages that are slower and operationally heavier than updating the knowledge used at inference time. Because the policies change multiple times per day, repeatedly fine-tuning to “bake in” new content is inefficient and risks serving stale answers.

Using Amazon Bedrock Knowledge Bases implements RAG during deployment:

• Select and curate the source documents (data selection)
• Keep the FM choice the same (model selection)
• Retrieve relevant passages at inference and generate grounded responses (deployment)
• Validate with a small test set and iterate (evaluation/feedback)

Key takeaway: when content changes frequently, prefer RAG over fine-tuning for freshness and lower ops.

Question 17

Topic: Content Domain 3: Applications of Foundation Models

Which statement is INCORRECT about high-level qualitative criteria used to evaluate GenAI outputs (helpfulness, accuracy, relevance, safety)?

• A. Helpfulness evaluates whether the response addresses the user’s intent and is actionable.
• B. Accuracy evaluates whether claims are factually correct and can be validated against reliable sources.
• C. Relevance evaluates whether the response stays on-topic and uses any provided context appropriately.
• D. Safety checks can be skipped if an answer is helpful and accurate.

Best answer: D

Explanation: The incorrect statement is the one that treats safety as optional when other criteria look good. Safety is a separate qualitative dimension that checks for harmful, disallowed, or sensitive outputs even when the response is useful, on-topic, and factually correct. Evaluations typically consider all four criteria together to understand overall output quality.

Qualitative evaluation of GenAI outputs commonly scores multiple dimensions because they capture different failure modes. In the scenario, helpfulness, accuracy, relevance, and safety are complementary: a response can be helpful but irrelevant to the user’s request, relevant but inaccurate (hallucinated), or accurate and relevant but still unsafe (for example, providing instructions for wrongdoing or disclosing sensitive data).

A simple way to apply these criteria is:

• Helpfulness: does it solve the user’s task?
• Accuracy: are the claims correct and verifiable?
• Relevance: is it on-topic and grounded in provided context (if any)?
• Safety: does it avoid harmful or prohibited content and sensitive disclosures?

Key takeaway: safety is not implied by the other criteria and must be evaluated explicitly.

Question 18

Topic: Content Domain 2: Fundamentals of GenAI

A company wants a GenAI chatbot for its internal HR help desk. The answers must include citations back to the exact policy paragraphs, and policies change weekly so updates must be reflected quickly without retraining a model. The HR documents must remain in the company’s AWS account.

Which solution is the BEST fit?

• A. Build and train a custom language model from scratch in Amazon SageMaker AI
• B. Fine-tune a foundation model in Amazon Bedrock using the HR policy documents
• C. Use Amazon Bedrock with Knowledge Bases to implement RAG over the HR documents in Amazon S3
• D. Rely only on prompt engineering and in-context examples sent with each request

Best answer: C

Explanation: Retrieval-augmented generation (RAG) is designed to ground an LLM on up-to-date enterprise content while keeping the model unchanged. Using Amazon Bedrock Knowledge Bases lets the chatbot retrieve relevant HR policy passages from documents stored in the company’s AWS account and return responses with citations. This best matches the requirements for source attribution and frequent content updates.

The key decision is choosing between in-context prompting, RAG, and fine-tuning. When the business requires answers to be based on specific, changing internal documents with citations, RAG is typically the best strategy because it retrieves authoritative passages at request time and uses them as context for generation. With Amazon Bedrock Knowledge Bases, the HR documents can stay in Amazon S3 in the company account, the knowledge base can be updated as documents change, and the model can cite the retrieved sources.

Fine-tuning is better for changing the model’s style/behavior or improving performance on a stable task, but it does not inherently provide per-answer citations and is not ideal when content changes weekly. Pure prompting can help format and tone, but it cannot reliably ensure factual grounding or citations from a large document set.

Question 19

Topic: Content Domain 4: Guidelines for Responsible AI

Which TWO statements about training dataset characteristics that support responsible AI are INCORRECT?

• A. Improve class balance by collecting more data or using careful re-sampling
• B. Check label quality and consistency across different groups
• C. Document data sources, collection methods, and known coverage gaps
• D. Historical ticket outcomes are safe labels even if past handling was biased
• E. Include data from diverse customer segments, including underserved groups
• F. Reduce bias by removing records from an underrepresented group

Best answers: D, F

Explanation: Responsible datasets aim for diverse, inclusive coverage and balanced representation so model performance is more equitable across groups. It is unsafe to drop an underrepresented group to “fix” bias, and it is also unsafe to assume historical outcomes are neutral when they may reflect biased processes. Better approaches include improving coverage, validating labels, and documenting known limitations.

A key responsible-AI dataset characteristic is whether the data represents the real-world populations and use cases the model will affect, including groups that are often under-sampled. “Balanced representation” typically means reducing harmful skews that cause systematically worse performance for some groups.

Two dataset practices are unsafe here:

• Removing data from an underrepresented group makes representation worse and can hide or amplify disparities.
• Treating historical outcomes as inherently safe labels can bake existing process bias into the dataset, even if labels are internally consistent.

Stronger signals of responsibility include collecting more representative data (or carefully re-sampling), validating label quality across groups, and documenting data provenance and known gaps so stakeholders understand limits and risks.

Question 20

Topic: Content Domain 2: Fundamentals of GenAI

A company wants to add generative text summarization to an existing application. The team wants to call a fully managed service that provides API access to multiple foundation models and does not require provisioning or managing model training infrastructure or inference endpoints.

Which AWS service best fits this requirement?

• A. Amazon Bedrock
• B. Amazon Comprehend
• C. Amazon SageMaker AI
• D. Amazon Q Business

Best answer: A

Explanation: The discriminating factor is whether the service offers managed foundation model inference through a simple API without managing model endpoints. Amazon Bedrock is designed for accessing and invoking multiple foundation models as managed APIs. This matches the team’s requirement to avoid provisioning and operating inference infrastructure.

The deciding attribute is “managed FM APIs vs. broader ML platform.” Amazon Bedrock is the AWS service that provides managed access to multiple foundation models through a consistent API, so teams can integrate generative capabilities (like summarization) without standing up or operating their own model hosting.

Amazon SageMaker AI is a broader ML platform used to build, train, customize, and deploy ML models (including hosting endpoints you manage). While it can be used for generative workloads, it is not primarily the single-API, multi-FM managed access layer described in the scenario. Key takeaway: choose Bedrock when you want managed FM invocation; choose SageMaker AI when you need a full ML lifecycle platform.

Question 21

Topic: Content Domain 4: Guidelines for Responsible AI

A retail company wants to use a generative AI solution on Amazon Bedrock to draft marketing copy. Legal leadership is most concerned about intellectual property (IP) infringement claims and reputational damage from outputs that are too similar to copyrighted third-party content.

Which approach best reduces this specific risk?

• A. Encrypt all data with AWS KMS while using a dataset with unclear licensing terms
• B. Use RAG with a knowledge base built only from licensed or company-owned content
• C. Add Bedrock Guardrails for toxicity while grounding on unvetted third-party documents
• D. Fine-tune the model on web-scraped product pages to improve writing style

Best answer: B

Explanation: The deciding factor is whether the content used to generate responses is licensed or company-owned with clear usage rights. Building a RAG solution that retrieves only approved, properly licensed material reduces exposure to copyright infringement allegations and the reputational harm that can follow. Controls like safety filters or encryption do not address IP ownership of the source content.

A key legal and reputational risk in GenAI is generating content that infringes on third-party intellectual property, especially when the system is trained on or grounded with copyrighted material that the organization does not have rights to use. The most direct mitigation is controlling data provenance: restrict the system’s grounding/training sources to content that is company-owned or properly licensed and maintain traceability of what sources are used.

Using Amazon Bedrock with a RAG pattern (for example, a knowledge base backed by approved documents in Amazon S3) helps because the model is guided by retrieved, authorized references rather than unvetted third-party text. In contrast, safety guardrails and security controls improve safety and confidentiality but do not resolve whether the underlying content can be legally used.

Question 22

Topic: Content Domain 2: Fundamentals of GenAI

A company is building a customer-support chatbot using Amazon Bedrock. Most users are in Germany and France, and a new compliance requirement states that all customer content (prompts and model responses) must be processed and stored only within the EU.

Which approach best satisfies this requirement?

• A. Deploy the application and invoke Amazon Bedrock in an EU Region, storing content in EU-region data stores
• B. Invoke Amazon Bedrock in us-east-1 for broader model choice and replicate stored data to an EU Region
• C. Invoke Amazon Bedrock in an EU Region but enable automatic failover to a non-EU Region for higher availability
• D. Use Amazon CloudFront to reduce latency for EU users while invoking Amazon Bedrock in us-east-1

Best answer: A

Explanation: The deciding attribute is regional data residency: prompts and responses must not leave the EU. Invoking Amazon Bedrock in an EU Region and keeping storage in the EU meets the compliance constraint and typically improves latency for users located in nearby EU countries.

Regional coverage affects GenAI solutions because the AWS Region you choose determines where requests are processed (latency), what happens during regional disruptions (availability), and whether data residency/compliance requirements are met. In this scenario, the single discriminator is the requirement that customer content (prompts and responses) be processed and stored only within the EU.

To satisfy that requirement at a high level:

• Run the application in an EU Region.
• Invoke Amazon Bedrock in an EU Region.
• Store any persisted content (for example, transcripts) in EU-region services.

Any design that sends prompts/responses to a non-EU Region, even as a performance optimization or failover path, violates the stated compliance requirement.

Question 23

Topic: Content Domain 1: Fundamentals of AI and ML

A company is building a customer-support chatbot using Amazon Bedrock. Two foundation models are being considered.

• Volume: 100,000 chats/month
• Each chat resolved by the bot avoids an average of $4 in agent cost; escalated chats avoid $0
• Estimated model usage per chat and vendor-quoted inference cost:
• Model A: 92% resolution rate in tests; 1,500 tokens/chat; $0.003 per 1,000 tokens
• Model B: 87% resolution rate in tests; 1,000 tokens/chat; $0.0015 per 1,000 tokens

The product owner wants to choose the model with the best expected business ROI while minimizing negative customer impact.

Which TWO actions should the team take? (Select TWO)

• A. Enable AWS CloudTrail logging for bedrock:InvokeModel to determine which model has higher ROI
• B. Run a short A/B pilot with Bedrock Guardrails and collect CSAT, escalation rate, and human-review findings on flagged outputs
• C. Choose the model with the higher resolution rate because accuracy is the primary indicator of ROI
• D. Store all prompts and responses in Amazon S3 with SSE-KMS to reduce inference cost per user
• E. Estimate monthly net ROI by converting resolution rate into avoided agent cost and subtracting estimated inference cost
• F. Fine-tune a custom model in Amazon SageMaker AI before calculating ROI so the comparison is fair

Best answers: B, E

Explanation: To compare models by business value, the team should translate model performance (resolution rate) into financial impact (avoided agent cost) and subtract the model’s inference costs to estimate ROI. They should also validate real user impact through a controlled pilot that captures customer feedback and safety/quality signals that can change the realized business outcome.

The core idea is to connect offline model metrics to business outcomes by expressing them in money and customer experience. Here, resolution rate determines how often the chatbot produces measurable savings, while token-based inference pricing determines cost per chat; together they produce an expected net value. Because offline test performance can differ from production behavior, a limited A/B pilot helps measure real CSAT and escalation rates, and adding guardrails plus targeted human review helps quantify safety-related quality issues that can reduce ROI through rework, churn, or brand risk.

Convert resolution rate and token cost into monthly net ROI
A/B pilot with guardrails + CSAT/escalation + human review
Higher resolution alone ignores per-chat cost and net value
CloudTrail supports auditability, not ROI comparison
SSE-KMS improves security, not inference cost per user
Fine-tuning adds dev cost/time and isn't required to compare ROI

Key takeaway: choose actions that quantify net value (savings minus cost) and validate customer impact, not just technical scores or controls.

Question 24

Topic: Content Domain 4: Guidelines for Responsible AI

Which statement correctly describes a legal and reputational risk related to intellectual property (IP) when an organization uses generative AI to produce public-facing content (for example, product images or marketing copy)?

• A. Encrypting prompts and outputs with AWS KMS prevents IP infringement claims.
• B. The model might generate content similar to copyrighted or trademarked work, so the organization must review for IP infringement before publishing.
• C. Using an AWS managed foundation model transfers IP infringement liability to AWS.
• D. IP infringement risk exists only when the organization fine-tunes a model with copyrighted training data.

Best answer: B

Explanation: A key GenAI risk is that outputs can be substantially similar to copyrighted or trademarked material, which can lead to infringement claims and reputational damage if published. This risk exists even when using managed models, so organizations should implement review and usage controls before releasing content.

Generative AI can produce novel-looking text or images that still resemble protected works (copyright) or brand identifiers (trademark). If an organization publishes that output externally, it can face IP infringement claims and associated reputational harm. This is not eliminated simply by choosing a managed foundation model; under AWS’s shared responsibility model, customers remain responsible for how they use outputs and whether published content complies with applicable laws and licensing obligations. Practical mitigations include human review workflows, content policies, and guardrails to reduce risky generations, but these controls reduce risk rather than guarantee zero infringement.

AIF-C01 AI practitioner map

Use this map after the sample questions to connect individual items to the AWS AI Practitioner fundamentals decisions these practice samples test.

    flowchart LR
	  S1["Business AI scenario"] --> S2
	  S2["Identify AI ML or GenAI workload type"] --> S3
	  S3["Choose AWS AI service family"] --> S4
	  S4["Apply responsible AI and security basics"] --> S5
	  S5["Review cost data and governance"] --> S6
	  S6["Select next AWS AI route"]

Quick Cheat Sheet

Mini Glossary

• Foundation model: Large pretrained model that can support many tasks through prompting or adaptation.
• Inference: Using a trained model to produce a prediction or output.
• Prompt: Input instruction or question given to a generative AI system.
• Responsible AI: Practices that manage safety, fairness, transparency, privacy, and accountability.
• SageMaker: AWS service family for building, training, deploying, and managing ML models.

In this section

• AWS AIF-C01: Fundamentals of AI and ML
  Try 10 focused AWS AIF-C01 questions on Fundamentals of AI and ML, with explanations, then continue with IT Mastery.
• AWS AIF-C01: Fundamentals of Genai
  Try 10 focused AWS AIF-C01 questions on Fundamentals of Genai, with explanations, then continue with IT Mastery.
• AWS AIF-C01: Applications of Foundation Models
  Try 10 focused AWS AIF-C01 questions on Applications of Foundation Models, with explanations, then continue with IT Mastery.
• AWS AIF-C01: Guidelines for Responsible AI
  Try 10 focused AWS AIF-C01 questions on Guidelines for Responsible AI, with explanations, then continue with IT Mastery.
• AWS AIF-C01: AI Security and Governance
  Try 10 focused AWS AIF-C01 questions on AI Security and Governance, with explanations, then continue with IT Mastery.
• Free AWS AIF-C01 Full-Length Practice Exam: 65 Questions
  Try 65 free AWS AIF-C01 questions across the exam domains, with explanations, then continue with full IT Mastery practice.