View Full PMI CPMAI Exam Dumps and Practice Test Dumps.

 

Question 21

Which phase of an AI project lifecycle focuses on understanding the current business environment, objectives, stakeholders, and constraints?

1. Business understanding
2. Model deployment
3. Model monitoring
4. Data preprocessing

Correct Answer: 1

Explanation

Business understanding establishes the context for an AI initiative before technical development begins. During this phase, the project team works to understand the organizational problem, desired outcomes, stakeholders, constraints, assumptions, and measures of success. This information helps determine whether an AI solution is appropriate and provides direction for later activities such as data preparation and model development. Without a clear understanding of the business need, a technically successful model may fail to provide meaningful value. Business understanding should therefore connect the AI initiative to measurable organizational objectives and provide a foundation for making decisions throughout the project lifecycle.

Question 22

What is the primary purpose of a data quality assessment in an AI project?

1. To determine whether the dataset is suitable and reliable for its intended use
2. To guarantee that the final model will be unbiased
3. To replace model testing
4. To eliminate the need for data governance

Correct Answer: 1

Explanation

A data quality assessment examines whether data is sufficiently accurate, complete, consistent, relevant, timely, and reliable for the intended AI application. Poor-quality data can negatively affect model development and may produce unreliable or misleading results. Assessment activities can identify missing values, duplicates, inconsistencies, incorrect labels, unusual records, and other problems that require attention. Data quality assessment does not guarantee that a model will be unbiased because bias can arise from many sources beyond basic data quality. It also does not replace model testing or governance. Instead, it provides important information for deciding how data should be prepared and used.

Question 23

Which approach is most appropriate for determining whether an AI use case should proceed?

1. Selecting the most advanced available model
2. Evaluating business value, feasibility, risks, data availability, and organizational requirements
3. Choosing the solution with the largest training dataset
4. Deploying a prototype before identifying the problem

Correct Answer: 2

Explanation

AI use-case evaluation should consider multiple dimensions rather than focusing only on technical sophistication. The team should examine the expected business value, availability and quality of required data, technical feasibility, operational requirements, costs, risks, regulatory considerations, and organizational readiness. A highly advanced model may not be appropriate if the problem does not require it or if the organization cannot support it effectively. Similarly, having a large dataset does not automatically make a use case valuable. A structured evaluation helps organizations prioritize initiatives that have a reasonable relationship between expected benefits, feasibility, risks, and available resources.

Question 24

What is data leakage in machine learning?

1. When data is encrypted before storage
2. When a model is deployed to multiple environments
3. When information unavailable at prediction time improperly influences model training
4. When duplicate records are removed from a dataset

Correct Answer: 3

Explanation

Data leakage occurs when information that should not be available to the model during prediction is unintentionally included during training or evaluation. This can cause the model to appear more accurate than it will be in real-world use. For example, a feature containing information generated after the target outcome occurs could improperly influence training. Leakage can also occur when training and test datasets are improperly separated. Detecting and preventing leakage is important because misleading evaluation results can lead teams to deploy models that perform poorly in production. Proper dataset design, feature review, and separation of training and evaluation processes can reduce this risk.

Question 25

Which type of AI risk is most directly associated with unauthorized access to training or production data?

1. Security risk
2. Usability risk
3. Scheduling risk
4. Estimation risk

Correct Answer: 1

Explanation

Unauthorized access to training or production data is primarily a security risk because it can compromise the confidentiality, integrity, or availability of information. AI systems may process sensitive business, customer, employee, or operational data, making appropriate security controls important throughout the lifecycle. Controls can include authentication, authorization, encryption, logging, network protections, and appropriate data-access policies. Security risks can also affect model integrity if unauthorized parties modify datasets or model artifacts. Although security can overlap with privacy and compliance concerns, unauthorized access itself is fundamentally a security issue. Organizations should assess these risks during planning and continue monitoring them after deployment.

Question 26

What is the main purpose of a confusion matrix when evaluating a classification model?

1. To calculate database storage capacity
2. To summarize correct and incorrect classification outcomes
3. To determine the amount of training data required
4. To identify the programming language used by the model

Correct Answer: 2

Explanation

A confusion matrix summarizes the outcomes of predictions made by a classification model by organizing them into categories such as true positives, true negatives, false positives, and false negatives. These values provide the foundation for calculating several evaluation metrics, including accuracy, precision, recall, and F1 score. Examining the confusion matrix can help a project team understand the types of errors a model makes rather than relying only on a single overall performance measure. The appropriate interpretation depends on the use case because false positives and false negatives may have different consequences. Therefore, confusion matrices support more detailed classification-model evaluation.

Question 27

Which practice helps ensure that an AI model is evaluated using data that was not used to train it?

1. Increasing the number of model parameters
2. Removing all validation activities
3. Using a separate test dataset
4. Reusing the training dataset for every evaluation

Correct Answer: 3

Explanation

A separate test dataset provides an independent collection of examples for evaluating a trained model. Because the model has not used the test examples during training, the results can provide a more realistic indication of how the model may perform on unseen data. The test dataset should be kept separate from training and should generally not be repeatedly used to guide model-development decisions. If teams repeatedly optimize against the test data, it can effectively become part of the development process and produce overly optimistic results. Maintaining appropriate separation between training, validation, and test data supports more trustworthy model evaluation.

Question 28

What is the purpose of establishing a baseline before introducing an AI solution?

1. To provide a reference point for comparing the AI solution’s performance or business impact
2. To guarantee that the AI model will exceed expectations
3. To eliminate the need for monitoring
4. To determine which programming language must be used

Correct Answer: 1

Explanation

A baseline establishes a reference point against which the performance or impact of an AI solution can be compared. Depending on the project, the baseline might represent an existing manual process, a current business metric, a simple statistical approach, or an existing system. Comparing AI results with the baseline helps stakeholders determine whether the new solution produces meaningful improvement. A baseline does not guarantee that an AI system will perform better, and it does not eliminate monitoring requirements. Establishing the baseline before implementation is particularly useful because it provides objective context for evaluating improvements in areas such as efficiency, accuracy, cost, quality, or customer experience.

Question 29

Which activity helps identify whether an AI system’s outputs remain reliable after deployment?

1. Monitoring model and business performance over time
2. Deleting historical performance records
3. Avoiding user feedback
4. Disabling system logging

Correct Answer: 1

Explanation

Monitoring model and business performance after deployment helps determine whether an AI system continues to produce useful and reliable results. Performance can change as the underlying data, user behavior, operating environment, or business conditions evolve. Monitoring may include model metrics, input-data characteristics, error rates, latency, system availability, user feedback, and relevant business outcomes. Historical performance records can also help identify trends and investigate incidents. Disabling logging or avoiding feedback removes useful information needed for oversight. Continuous monitoring should be aligned with the system’s risks and defined success criteria so that significant changes can be detected and addressed appropriately.

Question 30

What is the primary purpose of cross-validation in machine learning?

1. To encrypt training data
2. To assess how a model is likely to generalize using multiple data partitions
3. To increase the number of database records
4. To eliminate the need for a test dataset in every situation

Correct Answer: 2

Explanation

Cross-validation evaluates a machine learning model by repeatedly dividing available data into training and validation portions. In techniques such as k-fold cross-validation, the dataset is divided into several folds, and the model is trained and evaluated across different combinations of those folds. This provides a more robust estimate of performance than relying on a single training-validation split, particularly when the dataset is relatively limited. Cross-validation can support model selection and help identify overfitting. However, it does not encrypt data or increase the underlying amount of information. A separate final test set may still be appropriate when an unbiased final evaluation is required.

Question 31

Which statement best describes precision in classification?

1. The proportion of predicted positive cases that are actually positive
2. The proportion of actual positive cases that are correctly identified
3. The average numerical error of a regression model
4. The percentage of missing values in the dataset

Correct Answer: 1

Explanation

Precision measures how many of the cases predicted as positive are actually positive. It is calculated using true positives divided by the total number of predicted positives, which includes true positives and false positives. Precision becomes particularly relevant when false-positive results can create significant costs or consequences. Recall addresses a different concern by measuring how many actual positive cases were successfully identified. Neither metric alone necessarily provides a complete picture of model performance. Teams should select and interpret evaluation metrics according to the business context, error costs, and consequences associated with different types of incorrect predictions.

Question 32

What is a key benefit of conducting a pilot or limited AI deployment before full-scale implementation?

1. It guarantees complete regulatory approval
2. It allows the organization to evaluate the solution in a controlled environment
3. It eliminates the need for user training
4. It guarantees that the final model will never require updates

Correct Answer: 2

Explanation

A pilot provides an opportunity to evaluate an AI solution with a limited scope before committing to broader deployment. The organization can observe technical performance, operational processes, user interactions, data quality, integration issues, and unexpected risks. Feedback from the pilot can help identify changes needed before scaling the solution. A pilot does not guarantee regulatory approval or permanent model performance, and users may still require training. Its value comes from reducing uncertainty through controlled real-world experience. Clear pilot objectives and evaluation criteria should be established so that the organization can make informed decisions about whether and how to proceed with a wider implementation.

Question 33

Which factor should be considered when determining whether an AI system requires human review of its outputs?

1. The potential impact and risk associated with incorrect outputs
2. The color of the application’s user interface
3. The programming language used to build the system
4. The number of files in the project repository

Correct Answer: 1

Explanation

The need for human review should be influenced by the potential consequences and risks associated with incorrect AI outputs. Systems supporting high-impact decisions may require stronger human oversight than applications where errors have minimal consequences. Other factors can include model uncertainty, regulatory requirements, operational context, reversibility of decisions, and the ability of users to detect errors. Human review should be designed intentionally rather than added without considering the actual workflow. The programming language or user-interface color does not determine the appropriate level of oversight. A risk-based approach helps organizations establish proportionate review and escalation mechanisms.

Question 34

What does explainable AI primarily seek to improve?

1. The physical storage capacity of a server
2. The speed of internet connectivity
3. Understanding of how or why an AI system produces particular outputs
4. The number of records in a dataset

Correct Answer: 3

Explanation

Explainable AI focuses on making model behavior or outputs more understandable to appropriate users and stakeholders. Depending on the model and technique, explanations may identify influential features, show relationships between inputs and outputs, or provide information about why a particular prediction was produced. Explainability can support debugging, user understanding, governance, auditing, and informed decision-making. It does not directly increase server storage, network speed, or dataset size. The level and type of explanation should be appropriate to the audience and use case. For high-impact applications, understandable explanations can also help stakeholders identify unexpected behavior and determine when additional review is necessary.

Question 35

Which approach is most appropriate when an AI model’s performance decreases significantly after deployment?

1. Ignore the change because the model was previously validated
2. Investigate the cause using performance and data monitoring before deciding on corrective action
3. Immediately delete all production data
4. Automatically replace the model without analysis

Correct Answer: 2

Explanation

A significant decline in post-deployment performance should trigger investigation rather than an automatic response. The team should examine monitoring data, changes in input distributions, changes in business conditions, data quality, infrastructure behavior, and possible concept or data drift. The investigation can help determine whether retraining, feature changes, threshold adjustments, process changes, or model replacement is appropriate. Automatically replacing a model without understanding the underlying issue can introduce additional problems. Likewise, deleting production data would generally remove useful evidence. A structured incident and model-management process helps ensure that corrective actions are proportionate and based on reliable information.

Question 36

What is the primary objective of model validation?

1. To determine whether the model performs appropriately against defined evaluation criteria
2. To make the model more complex regardless of results
3. To remove all project stakeholders
4. To eliminate the need for deployment monitoring

Correct Answer: 1

Explanation

Model validation evaluates whether a model meets established technical and business requirements using appropriate data and metrics. Validation can examine predictive performance, robustness, fairness-related measures, operational behavior, and other criteria relevant to the use case. It provides evidence that supports decisions about whether a model is ready for further testing or deployment. Validation should be based on predefined evaluation criteria rather than changing standards after seeing results. It does not eliminate the need for monitoring because model behavior can change after deployment. Nor should validation automatically favor complexity. The goal is to determine whether the model is appropriate for its intended purpose.

Question 37

Why should AI project assumptions be documented?

1. To ensure that assumptions can be reviewed and challenged as project conditions change
2. To prevent stakeholders from providing feedback
3. To guarantee that project risks disappear
4. To replace all formal requirements

Correct Answer: 1

Explanation

Documenting assumptions creates visibility into conditions that the project team believes to be true when making plans and decisions. Examples can include assumptions about data availability, user behavior, infrastructure capacity, business processes, or regulatory conditions. If an assumption changes or proves incorrect, the team can assess its impact on scope, schedule, cost, risk, and solution design. Documentation also improves communication among stakeholders and provides a record for future review. Assumptions do not eliminate risks or replace formal requirements. Instead, they make uncertainty more visible and support proactive project management when circumstances change during the AI lifecycle.

Question 38

Which practice can help reduce unintended exposure of sensitive information in an AI system?

1. Granting every user unrestricted data access
2. Applying appropriate access controls and data-minimization practices
3. Copying all available data into every environment
4. Disabling security monitoring

Correct Answer: 2

Explanation

Access controls and data minimization can reduce unnecessary exposure of sensitive information within AI systems. Access controls restrict data and system capabilities to authorized users or processes based on appropriate permissions. Data minimization focuses on collecting and using only the information necessary for the intended purpose. Additional measures such as encryption, retention controls, logging, and secure development practices may also be appropriate. Giving everyone unrestricted access increases exposure and can create security and privacy risks. Copying all available information into every environment also expands the potential attack surface. These controls should be incorporated into AI project design rather than treated as an afterthought.

Question 39

What is the main reason for involving subject-matter experts during AI model development?

1. They can provide domain knowledge that helps evaluate data, assumptions, outputs, and business relevance
2. They eliminate the need for data scientists
3. They guarantee perfect model accuracy
4. They are responsible for writing every line of model code

Correct Answer: 1

Explanation

Subject-matter experts contribute knowledge about the business domain, operational processes, terminology, constraints, and practical consequences of AI outputs. Their involvement can help the project team determine whether data represents the real-world process correctly, identify important variables, review assumptions, interpret unusual results, and assess whether model outputs make sense in context. Subject-matter experts do not necessarily replace technical specialists or guarantee perfect accuracy. Their value comes from combining domain expertise with the capabilities of data and AI professionals. Collaboration among these roles can help ensure that the resulting solution is technically appropriate and aligned with actual organizational needs.

Question 40

What should be included in an effective AI project risk-management approach?

1. Only technical risks identified during development
2. A process for identifying, assessing, responding to, and monitoring relevant risks
3. Only risks that have already occurred
4. A rule that all risks must be completely eliminated

Correct Answer: 2

Explanation

Effective AI project risk management involves systematically identifying relevant risks, assessing their likelihood and potential impact, selecting appropriate responses, assigning accountability, and monitoring risks throughout the project lifecycle. Risks can involve data quality, privacy, security, bias, model performance, operational processes, compliance, vendor dependencies, and business outcomes. Risk management does not require every risk to be completely eliminated because some residual risk may remain after reasonable controls are applied. The important objective is to understand risks and manage them according to organizational thresholds and governance processes. Continuous monitoring is also necessary because AI-related risks can change as the system and environment evolve.