What the exam covers
Domain 1Fundamentals of AI and ML20%
- 1.1Explain basic AI concepts and terminologies
- 1.2Identify practical use cases for AI
- 1.3Describe the ML development lifecycle
Domain 2Fundamentals of Generative AI24%
- 2.1Explain the basic concepts of generative AI
- 2.2Understand the capabilities and limitations of generative AI for solving business problems
- 2.3Describe AWS infrastructure and technologies for building generative AI applications
Domain 3Applications of Foundation Models28%
- 3.1Describe design considerations for applications that use foundation models
- 3.2Choose effective prompt engineering techniques
- 3.3Describe the training and fine-tuning process for foundation models
- 3.4Describe methods to evaluate foundation model performance
Domain 4Guidelines for Responsible AI14%
- 4.1Explain the development of AI systems that are responsible
- 4.2Recognize the importance of transparent and explainable models
Domain 5Security, Compliance, and Governance for AI Solutions14%
- 5.1Explain methods to secure AI systems
- 5.2Recognize governance and compliance regulations for AI systems
Sample questions
A vineyard wants to estimate how many kilograms of grapes each plot will produce this season. It has ten years of harvest weights for every plot, along with rainfall, temperature, and soil readings. Which type of AI model produces this kind of output?
- Multi-modal
- Transformer
- Diffusion
- Regression
Answer
D. The vineyard wants a continuous quantity, a harvest weight in kilograms, and predicting a number from input features is what a regression model does. A transformer is an architecture used mainly for language and other sequence tasks, and a diffusion model generates content such as images; neither term describes a model whose job is to output a numeric estimate.
A grocery cooperative holds years of loyalty card purchase histories. Nobody has assigned shoppers to categories, and the cooperative has no predefined groups in mind. It wants to discover groups of shoppers with similar buying habits so that it can design offers for each group. Which type of learning should the cooperative use?
- Supervised learning
- Reinforcement learning
- Unsupervised learning
- Reinforcement learning from human feedback (RLHF)
Answer
C. The data has no labels and the goal is to discover natural groups, which is clustering, an unsupervised learning task. Supervised learning is the tempting choice because the result looks like a classification, but it needs examples that are already labeled with the correct group. Reinforcement learning, with or without human feedback, trains an agent through rewards and does not group records.
A county records office has decades of property deeds that exist only as scanned page images. Clerks currently retype the content of each deed by hand, and the backlog is growing faster than they can work. The office wants to turn the scanned pages into text automatically so the text can be loaded into a search system. Which AWS service should the office use?
- Amazon Lex
- Amazon Personalize
- Amazon Textract
- Amazon Transcribe
Answer
C. Amazon Textract uses optical character recognition and ML to extract printed text, handwriting and form data from scanned documents. Amazon Transcribe is the common mix-up because it also produces text, but its input is audio, not page images.
An agricultural cooperative is developing an ML model to forecast crop yields from soil readings, weather records, and planting history. The cooperative plans to use feature engineering to build a model that makes accurate predictions. Which approach will meet these requirements?
- Create or select relevant features for model training.
- Understand patterns by providing data visualization.
- Collect data from multiple sources.
- Tune the model's hyperparameters.
Answer
A. Feature engineering is the process of creating, transforming, or selecting the input variables that the model learns from, such as combining soil and weather readings into more useful signals. Hyperparameter tuning adjusts the settings of the training process, not the features. Collecting data and visualizing it are earlier steps in the workflow (data collection and exploratory data analysis).
A grocery chain has developed a generative AI application by using Amazon Bedrock. The application produces promotional flyers that are updated weekly, in a single scheduled run rather than in response to live shopper requests. The chain must select a pricing model for the weekly flyer updates. Which Amazon Bedrock pricing model meets these requirements MOST cost- effectively?
- Provisioned Throughput
- On-Demand
- Batch
- Model evaluation
Answer
C. Batch inference in Amazon Bedrock processes a large set of prompts as one asynchronous job and is priced lower than On-Demand for supported models, which suits a weekly scheduled run that needs no immediate response. On-Demand would work but costs more for the same tokens. Provisioned Throughput reserves capacity for a commitment period and would sit unused for most of the week.
A museum keeps its membership records in a relational database. Fundraisers who have never written SQL want to type questions such as 'which members renewed in March but have not visited since' and get back a query that they can run. Which type of model can turn these questions into SQL?
- WaveNet
- Generative pre-trained transformers (GPT)
- Residual neural network
- Support vector machine
Answer
B. Generative pre-trained transformers understand natural language and generate text, and SQL is a form of structured text, so they can translate a question into a query. A support vector machine is the tempting classic ML option, but it classifies data and cannot write text. WaveNet generates audio, and residual neural networks are mainly used for images.
An engineer is loading 300-page equipment manuals into a vector index for a Retrieval Augmented Generation (RAG) application. Instead of embedding each manual as one item, she first breaks every manual into short passages. What is this chunking step meant to achieve?
- To improve the contextual relevancy of results retrieved from the vector index
- To decrease the cost of storage by storing parts or chunks of the text
- To avoid database storage limitations for large text documents by storing parts or chunks of the text
- To improve efficiency by avoiding the need to convert large text into vector embeddings
Answer
A. A short passage produces an embedding that represents one focused idea, so searching the vector index returns the passages most relevant to the question instead of a whole manual that covers hundreds of topics. Chunking is not done to save storage or to get around database size limits. Every chunk is still converted into a vector embedding, so no conversion work is avoided.
A consumer electronics company has launched a generative AI support assistant. Staff notice that when two customers ask the identical question, the assistant sometimes words its answers differently, and now and then it states product specifications that are simply false. Which pair of issues does this behavior represent? (Choose two.) (Choose two.)
- Nondeterminism
- Toxicity
- Hallucination
- Bias
- Interpretability
Answer
A, C. Different wording for the identical question is nondeterminism, because the model samples its output and can vary from run to run. Stating product specifications that are false is hallucination. Bias would mean systematically unfair treatment of groups and toxicity would mean offensive content, and neither appears in the scenario.
A gaming studio is benchmarking a pre-trained foundation model (FM) for a live commentary feature. The studio measures how long the model takes to process each incoming request and return its output to players. Which term is the speed at which a pre-trained foundation model (FM) processes requests and delivers output?
- Inference latency
- Fine-tuning
- Context window
- Model size
Answer
A. Inference latency is the time a model takes to process a request and return its output, which is exactly what the studio is measuring. Model size affects latency but is not the measure of speed itself. The context window is how much input the model can consider at once, and fine-tuning is a training method.
A media company wants to use foundation models (FMs) across several workloads: summarizing articles, generating code for its internal tools, and powering a reader support assistant. Which approach implements FMs in the MOST cost-effective way?
- Select the smallest model that meets the performance requirements for each specific use case.
- Build and train a custom FM for each individual use case.
- Use the largest and most capable model available for all tasks to ensure quality.
- Use batch processing for all inference requests to reduce the costs and latency for each request.
Answer
A. Smaller models cost less per token and respond faster, so choosing the smallest model that meets the quality bar for each workload gives the lowest cost. Using the largest model for everything means paying for capability that simple tasks do not need. Batch processing lowers cost but adds delay, so it does not suit a live reader support assistant, and building custom models is the most expensive option.
A restaurant franchise wants an assistant that answers store managers' questions about food safety, staffing, and refund policies. Head office revises the policy documents many times a year, and managers must always get answers that match the current versions. Which strategy meets this need MOST cost-effectively?
- Use Retrieval Augmented Generation (RAG) combined with prompt engineering techniques.
- Fine-tune the model again every time a policy changes.
- Pre-train and benchmark the model by using the policy documents as context data.
- Train the model from scratch by using the policy documents as context data.
Answer
A. RAG looks up the current policy documents when a question is asked and passes the relevant text to the model, so answers follow the latest versions with no training cost. Fine-tuning again after every policy change could also keep answers current, but it means paying for a training job many times a year. Pre-training or training a model from scratch on the documents is the most expensive route of all.
A seismology institute adapted a foundation model (FM) to its field through fine-tuning. Each quarter, the institute also continues pre-training the model on newly published field reports. What does the institute gain from this ongoing pre-training?
- Decreases the training time requirement
- Improves model performance over time
- Optimizes model inference time
- Helps decrease the model's complexity
Answer
B. Ongoing pre-training keeps exposing the model to new text from the domain, so its knowledge of the field grows and its performance improves over time. It adds training work instead of reducing it, and it does not make the model simpler or make inference faster.
Engineers at a shipyard will use a chatbot built on a large language model (LLM) to ask about welding procedures. The yard has an extensive library of its own standards in digital form, and the answers must come from that library. Which technique should the yard use to improve the chatbot's responses?
- Use Retrieval Augmented Generation (RAG).
- Decrease the token size.
- Use few-shot prompting.
- Set the temperature to 1.
Answer
A. RAG searches the yard's standards for passages relevant to each question and gives them to the model with the prompt, so the answers are based on those documents. Few-shot prompting places a handful of examples in the prompt to show the desired format, which cannot cover an extensive library. A temperature of 1 only makes the wording more varied.
A publishing platform is building an application that summarizes research papers, which range from short abstracts to documents hundreds of pages long. In testing, the application successfully summarizes the shorter papers but fails on the longest ones. What is the most likely reason?
- The input tokens exceed the model's context size.
- The temperature is set too high.
- The selected model does not support fine-tuning.
- The Top P value is too high.
Answer
A. Every model has a context window, a maximum number of tokens it can take in at once, and a paper hundreds of pages long exceeds it, so the request fails while shorter papers work. Temperature and Top P only affect how random the wording is, not how much input fits. The usual fix is to split long documents into chunks and summarize them in stages, or to use a model with a larger context window.
A law library uses a model in Amazon Bedrock to condense long court rulings into one-paragraph digests. The library starts an automatic model evaluation job in Amazon Bedrock and selects text summarization as the task type. Which metric does the job use to measure the accuracy of the digests?
- F1 score
- Area Under the ROC Curve (AUC) score
- Real world knowledge (RWK) score
- BERTScore
Answer
D. For the text summarization task type, Amazon Bedrock automatic evaluation measures accuracy with BERTScore, which uses embeddings to compare the meaning of each generated summary with a reference summary. F1 score is the accuracy metric for question answering, and the real world knowledge score applies to general text generation.
A manufacturer that is a long-time Amazon SageMaker AI customer runs a model that predicts when factory machines are likely to fail. Plant managers want to understand which sensor readings push the model toward a failure prediction. Which SageMaker feature that the manufacturer already uses meets this requirement?
- SageMaker Clarify
- SageMaker Ground Truth
- SageMaker Canvas
- SageMaker Feature Store
Answer
A. SageMaker Clarify produces feature attributions (based on SHAP) that show how much each input contributes to a prediction. Feature Store only stores and serves features; it does not explain model behavior. Note: Amazon SageMaker Clarify has been closed to new AWS customers since July 2026; existing customers can keep using it.
An auto insurer is training a generative AI model to predict the outcomes of insurance claims. The training dataset is small. The dataset groups policyholders as "new drivers," "mid-career drivers," or "senior drivers." Most individuals in the dataset are characterized as "mid-career drivers." The insurer removes the driver age-band feature from the training dataset. Which model behavior will likely happen as a result of this change to the dataset?
- The model will inaccurately predict outcomes for new and senior driver groups.
- The model will accurately predict outcomes for all driver groups.
- The model will predict accurate outcomes for only new driver groups.
- The model will require less training data.
Answer
A. The dataset is small and dominated by mid-career drivers, so the model mostly learns patterns for that group. After the age-band feature is removed, the model can no longer tell the groups apart and will apply the majority pattern to everyone, which makes predictions inaccurate for new and senior drivers. Removing the feature does not fix the imbalance, so accurate results for all groups are not likely.
An auto insurance company that has built its models on Amazon SageMaker AI for several years, including its bias and explainability tooling, runs an ML model that recommends whether a claim should be approved. Insurance regulators demand transparency, so the company must be able to show how the model arrives at each recommendation. Which AWS service or feature satisfies this demand?
- Amazon SageMaker Clarify
- Amazon SageMaker Model Monitor
- Amazon Rekognition
- Amazon Comprehend
Answer
A. SageMaker Clarify provides feature attributions that show how much each input contributed to a prediction, which lets the company explain each recommendation to regulators. SageMaker Model Monitor tracks drift in a deployed model but does not explain individual predictions. Note: Amazon SageMaker Clarify has been closed to new AWS customers since July 2026; existing customers can keep using it.
An airline uses a generative AI system to draft maintenance summaries from aircraft inspection records. Some summaries state inspection findings that sound convincing but never actually occurred. Which approach will address this problem?
- Give the system a larger model backed by more computing capacity.
- Enforce rigid validation rules on the inputs and a fixed format on the outputs.
- Apply Retrieval Augmented Generation (RAG) so the summaries are grounded in the actual inspection records.
- Switch to smaller models whose inner workings are easier to inspect.
Answer
C. Retrieval Augmented Generation (RAG) supplies the actual inspection records as context, so the summaries are grounded in real findings and not invented ones. A larger model with more computing capacity can still hallucinate because it still lacks the specific records. Rigid input and output formats control structure but do not make the content true.
A retail chain uses Amazon Comprehend to study customer reviews and has built several custom models, each running on its own assigned endpoint. The chain wants an automated report that flags any endpoint that has gone unused for more than 15 days. Which service will meet these requirements?
- Amazon CloudWatch
- AWS Trusted Advisor
- AWS Config
- AWS CloudTrail
Answer
B. AWS Trusted Advisor includes a cost optimization check for underutilized Amazon Comprehend endpoints that flags endpoints with no real-time inference activity over the past 15 days. Amazon CloudWatch holds the underlying usage metrics, but you would have to build the alarms and reporting yourself.
A debt counseling service plans to save the messages it receives from clients in Amazon S3 and build a reporting application on top of them. Some messages include personal details such as national ID and bank account numbers, and the service wants to be alerted automatically whenever an uploaded file contains this kind of information. Which approach finds the sensitive information with the LEAST development effort?
- Write a set of regex patterns for sensitive data and host them in an Amazon SageMaker notebook.
- Configure Amazon Macie to scan the objects uploaded to Amazon S3 for sensitive information.
- Deploy a large language model (LLM) on an Amazon SageMaker endpoint to redact the sensitive data.
- Instruct clients not to include sensitive information in their messages.
Answer
B. Amazon Macie is a managed service that scans S3 objects with ML and pattern matching, finds sensitive data such as ID and bank account numbers, and raises findings that can trigger alerts, so almost nothing has to be built. Regex patterns in a notebook and an LLM hosted on a SageMaker AI endpoint could both find some of this data, but each is a custom solution to build and maintain. Asking clients to leave out sensitive details detects nothing.