View Full PMI CPMAI Exam Dumps and Practice Test Dumps.

 

Question 241

Which factor should be considered when determining whether an AI solution is technically feasible?

1. Availability of suitable data, infrastructure, skills, and required technical capabilities
2. The number of project meetings
3. The preferred color of the user interface
4. The age of the project manager

Correct Answer: 1

Explanation

Technical feasibility determines whether an organization has, or can reasonably obtain, the technical capabilities required to implement and operate the proposed AI solution. This assessment may include data availability and quality, computing resources, integration capabilities, security requirements, model-development skills, deployment infrastructure, and technical constraints. A technically attractive idea may still be impractical if required data or infrastructure cannot be obtained. Feasibility analysis should occur before significant implementation resources are committed. The assessment should also consider whether the solution can meet required performance, scalability, reliability, and security expectations. Identifying technical limitations early allows the team to modify the approach, obtain missing capabilities, or reconsider the proposed solution.

Question 242

What is the primary purpose of a business case for an AI project?

1. To describe only the machine learning algorithm
2. To justify the initiative by connecting expected value, costs, risks, and strategic objectives
3. To replace model testing
4. To document programming syntax

Correct Answer: 2

Explanation

An AI business case explains why an organization should consider investing in a proposed initiative. It typically connects the business problem with expected benefits, costs, risks, assumptions, resource requirements, strategic alignment, and potential alternatives. A strong business case allows decision-makers to evaluate whether the initiative is worth pursuing relative to other opportunities. It should not focus exclusively on the machine learning algorithm because technical sophistication does not automatically create business value. Financial estimates, operational impacts, implementation considerations, and uncertainties should also be considered. The business case can evolve as new evidence becomes available during feasibility studies, pilots, and implementation.

Question 243

Which type of data consists primarily of free-form text, images, audio, or video rather than predefined tabular fields?

1. Structured data
2. Relational data
3. Unstructured data
4. Normalized data

Correct Answer: 3

Explanation

Unstructured data does not follow a predefined tabular schema in the same way that traditional structured datasets do. Examples include free-form documents, photographs, audio recordings, videos, and natural-language conversations. AI systems can use techniques such as natural language processing, computer vision, speech processing, and embeddings to extract useful information from these sources. Structured data, by contrast, is typically organized into defined fields and records. Unstructured data often requires additional processing, metadata, labeling, storage, and governance considerations. Understanding the type of data involved helps project teams select appropriate technologies, preprocessing methods, quality controls, and evaluation approaches for the intended AI use case.

Question 244

Which practice helps ensure that training and testing datasets remain appropriately separated?

1. Using the test dataset repeatedly during model tuning
2. Maintaining a controlled train-validation-test split
3. Combining all datasets before evaluation
4. Selecting test examples after viewing model predictions

Correct Answer: 2

Explanation

Maintaining a controlled separation between training, validation, and test datasets helps produce a more reliable estimate of model generalization. Training data is used to fit the model, validation data supports development and tuning decisions, and the test dataset should provide an independent final evaluation. Repeatedly using test data during model development can cause the team to indirectly optimize for that dataset, weakening its role as an independent measure. The exact splitting strategy depends on the problem, including whether the data has temporal, grouped, or other dependencies. Proper separation should be established early and protected throughout the development process.

Question 245

Which principle requires an AI system to provide understandable information about how its outputs are produced or used?

1. Transparency
2. Compression
3. Replication
4. Normalization

Correct Answer: 1

Explanation

Transparency involves providing appropriate information about an AI system, including its purpose, capabilities, limitations, data use, and relevant operational processes. The level of transparency required depends on the system and its impact. For some applications, stakeholders may need information about how outputs are generated, while others may require documentation about data sources, limitations, or governance procedures. Transparency does not necessarily mean exposing proprietary source code or every internal model parameter. Instead, it supports informed understanding and responsible use. Appropriate transparency can improve trust, facilitate oversight, help users interpret outputs, and make it easier to identify situations where AI results should be questioned or independently verified.

Question 246

What is the main purpose of cross-validation during model development?

1. To evaluate model performance across multiple training and validation splits
2. To encrypt the training dataset
3. To deploy a model directly into production
4. To remove every outlier

Correct Answer: 1

Explanation

Cross-validation evaluates a model across multiple partitions of available development data. In techniques such as k-fold cross-validation, the data is divided into several folds, with different folds used for validation across repeated training runs. This provides a broader view of model performance than relying on a single split and can help identify whether results are highly dependent on one particular partition. Cross-validation is especially useful when available data is limited. It does not replace a properly reserved final test dataset when independent final evaluation is required. The exact cross-validation method should also account for characteristics such as time ordering, groups, or class distribution.

Question 247

Which approach is most appropriate when a model performs well overall but poorly for a particular subgroup?

1. Ignore the subgroup because overall performance is high
2. Investigate subgroup data, performance metrics, and potential causes of the disparity
3. Remove the subgroup from evaluation
4. Automatically deploy the model without further review

Correct Answer: 2

Explanation

Strong overall performance can conceal significant differences in outcomes across subgroups. When a model performs poorly for a particular group, the team should investigate relevant data quality, representation, feature behavior, model performance, and potential sources of disparity. Appropriate metrics should be calculated separately for relevant groups, and domain experts may help determine whether observed differences are meaningful. Depending on the findings, corrective actions could involve improving data collection, revising features, adjusting the model, changing thresholds, or introducing additional human oversight. Simply ignoring or removing the affected subgroup from evaluation would hide an important issue and could increase the risk of inappropriate deployment.

Question 248

Which document is most useful for recording assumptions that could affect an AI project’s planning and outcomes?

1. Assumptions log
2. Password list
3. Source-code compiler
4. Network routing table

Correct Answer: 1

Explanation

An assumptions log provides a structured record of conditions believed to be true for project planning purposes but that may not yet be fully verified. AI project assumptions might concern data availability, stakeholder participation, infrastructure capacity, expected user adoption, vendor capabilities, or business conditions. Documenting assumptions makes them visible so they can be monitored and validated during the project. If an important assumption changes, the team can assess its impact on scope, schedule, cost, risks, or expected outcomes. An assumptions log complements other project artifacts such as the risk register and project plan. It should be reviewed periodically because assumptions can become invalid as new evidence emerges.

Question 249

Which characteristic is most important when selecting data for an AI model intended for a specific population?

1. The dataset should reasonably represent the population and intended operating conditions
2. The dataset should always be as small as possible
3. The dataset should contain only the easiest examples
4. The dataset should exclude difficult cases

Correct Answer: 1

Explanation

Representative data helps an AI model learn patterns that are relevant to the population and operating conditions where it will be used. If important groups, scenarios, or environments are missing or underrepresented, model performance may differ substantially after deployment. Teams should examine sampling methods, subgroup representation, data quality, historical coverage, and relevant edge cases when evaluating training and validation data. Simply increasing dataset size does not guarantee representativeness. Excluding difficult examples can also produce unrealistic performance estimates. Data selection should therefore reflect the intended use of the system and should be reviewed with appropriate subject-matter experts to identify important populations, scenarios, and potential gaps.

Question 250

What is the primary purpose of model calibration?

1. To make predicted probabilities better correspond to observed outcome frequencies
2. To increase the size of the training dataset
3. To remove all categorical variables
4. To encrypt model parameters

Correct Answer: 1

Explanation

Model calibration concerns whether predicted probabilities correspond appropriately to observed frequencies. For example, among cases assigned a predicted probability of approximately 0.8, a well-calibrated model would have an outcome frequency reasonably close to 80 percent over an appropriate population. Calibration can be important when downstream decisions depend on probability estimates rather than only ranking or classification labels. A model can have strong discrimination while still being poorly calibrated. Calibration methods may be applied after model development when appropriate. Evaluation should consider the intended use and relevant population because calibration can vary across groups or operating environments.

Question 251

Which activity helps identify whether an AI solution is aligned with the organization’s strategic priorities?

1. Strategic alignment assessment
2. Random hyperparameter selection
3. File compression
4. Uncontrolled model deployment

Correct Answer: 1

Explanation

A strategic alignment assessment examines whether a proposed AI initiative supports the organization’s broader goals and priorities. The assessment may consider expected business outcomes, strategic objectives, customer needs, operational priorities, competitive considerations, and available organizational capabilities. An AI project can be technically feasible but still have limited value if it does not contribute meaningfully to important organizational objectives. Strategic alignment should therefore be evaluated alongside technical, operational, economic, and risk considerations. Clear alignment also helps stakeholders understand why the project matters and can support prioritization when several potential AI initiatives compete for limited funding, personnel, infrastructure, or management attention.

Question 252

Which security principle recommends giving users only the permissions necessary to perform their assigned responsibilities?

1. Least privilege
2. Maximum access
3. Open authorization
4. Universal administration

Correct Answer: 1

Explanation

The principle of least privilege limits users, applications, and services to the minimum permissions required to perform authorized activities. In AI environments, this can apply to training datasets, model artifacts, development environments, production systems, monitoring tools, and administrative functions. Limiting privileges reduces the potential impact of compromised accounts, accidental changes, and unauthorized activity. Permissions should be reviewed periodically because responsibilities can change over time. Least privilege should be implemented alongside strong authentication, logging, access reviews, and other security controls. It does not mean that users should be unable to perform legitimate work; instead, it seeks to provide sufficient access while minimizing unnecessary exposure.

Question 253

Which approach is most useful for evaluating a classification model when both precision and recall are important?

1. Considering the F1 score along with the individual precision and recall values
2. Looking only at training time
3. Measuring only infrastructure cost
4. Ignoring false positives and false negatives

Correct Answer: 1

Explanation

The F1 score combines precision and recall into a single measure based on their harmonic mean. It can be useful when both false positives and false negatives matter and a balance between precision and recall is desired. However, the individual precision and recall values should still be reviewed because the same F1 score can arise from different combinations of the two metrics. In some applications, one type of error may be considerably more costly than the other, making a simple F1 score insufficient. Teams should therefore select metrics based on the business context and evaluate the confusion matrix and subgroup performance when appropriate.

Question 254

What is the main purpose of a pilot deployment?

1. To test the solution with limited real-world use before broader implementation
2. To permanently prevent production deployment
3. To eliminate the need for user feedback
4. To guarantee that the system will never require changes

Correct Answer: 1

Explanation

A pilot deployment introduces an AI solution to a limited operational environment before wider rollout. It allows teams to evaluate real-world behavior, user adoption, workflow integration, performance, reliability, and unexpected issues under controlled conditions. Pilot results can provide evidence for improving the solution before broader implementation. The pilot should have defined objectives and evaluation criteria so that stakeholders know what information is needed to make the next decision. A pilot does not guarantee that the system will be successful or eliminate the need for future changes. Instead, it reduces uncertainty by generating practical evidence from actual or representative operating conditions.

Question 255

Which activity is most directly associated with improving the quality of labeled training data?

1. Reviewing annotation guidelines and measuring labeling consistency
2. Increasing server storage only
3. Changing the production network
4. Removing all validation datasets

Correct Answer: 1

Explanation

High-quality labels are important because supervised learning models learn from the examples provided during training. Teams can improve labeling quality by establishing clear annotation guidelines, training annotators, reviewing ambiguous examples, conducting quality checks, and measuring agreement or consistency between annotators when appropriate. Disagreements can reveal unclear definitions or difficult cases that require clarification. Increasing infrastructure capacity does not directly improve label quality. Poorly labeled data can lead to noisy training signals, misleading evaluation results, and reduced model performance. Label-quality processes should therefore be designed according to the complexity and consequences of the AI use case and maintained throughout data preparation.

Question 256

Which type of change should generally be recorded in an AI model’s change log?

1. Significant changes to model versions, configuration, data, or deployment behavior
2. Only employee birthdays
3. Unrelated office maintenance
4. Personal entertainment preferences

Correct Answer: 1

Explanation

A model change log records important changes that can affect an AI system’s behavior or lifecycle status. Examples include new model versions, changes to training data, feature definitions, hyperparameters, deployment configurations, thresholds, infrastructure, or other significant components. Recording these changes supports traceability, troubleshooting, auditing, reproducibility, and controlled governance. The level of detail should reflect the organization’s requirements and the system’s risk. A useful change log should identify what changed, when it changed, who or which process made the change, and where appropriate, why the change was made. This makes it easier to investigate performance changes and understand the history of deployed AI systems.

Question 257

Which practice can reduce the risk of sensitive information being retained unnecessarily in an AI data pipeline?

1. Applying appropriate data-retention and deletion policies
2. Keeping every dataset forever
3. Copying sensitive data into every development environment
4. Removing all access controls

Correct Answer: 1

Explanation

Data-retention policies define how long information should be kept and when it should be securely deleted or otherwise disposed of according to applicable requirements and organizational needs. Limiting unnecessary retention can reduce the amount of sensitive information exposed to potential security incidents and can simplify data governance. Retention periods should consider legitimate business, legal, regulatory, operational, and analytical requirements. Keeping every dataset indefinitely may increase unnecessary risk. Teams should also control where copies of sensitive information exist and apply appropriate access restrictions. Effective retention management should be incorporated into the AI lifecycle rather than treated as an issue only after the model has been deployed.

Question 258

What should a project team do when an AI model fails an important predefined acceptance criterion?

1. Treat the failure as irrelevant if the model is already deployed
2. Investigate the cause and determine appropriate corrective action before accepting the deliverable
3. Delete the acceptance criterion
4. Automatically increase model complexity

Correct Answer: 2

Explanation

When an AI model fails an important acceptance criterion, the team should investigate why the requirement was not met and determine the appropriate response. The issue may result from inadequate data, model limitations, incorrect requirements, implementation defects, integration problems, or environmental conditions. Depending on the findings, corrective action could include modifying the model, improving data, changing configuration, addressing infrastructure issues, or revisiting the requirement through formal change control. Acceptance criteria exist to provide objective conditions for determining whether a deliverable is acceptable. Ignoring a failed criterion simply because the model has already been deployed can allow known problems to continue without appropriate review or accountability.

Question 259

Which deployment approach allows a new AI model to process production inputs while its outputs are not used to make actual decisions?

1. Shadow deployment
2. Canary deployment
3. Full replacement
4. Manual-only deployment

Correct Answer: 1

Explanation

Shadow deployment allows a new model to operate using production inputs while its outputs remain separate from the live decision-making process. This allows teams to compare the new model with the existing system under realistic conditions without immediately affecting users or business outcomes. Teams can evaluate accuracy, latency, resource consumption, failure patterns, and other relevant characteristics before deciding whether to promote the model. Canary deployment differs because a limited portion of actual production traffic may receive decisions from the new model. Shadow deployment is particularly useful when teams want real-world evidence while minimizing the immediate consequences of an unproven model.

Question 260

Which activity best supports continuous improvement of an AI system after deployment?

1. Collecting relevant performance data and user feedback, then using findings to guide controlled improvements
2. Freezing the system permanently
3. Ignoring operational incidents
4. Removing all monitoring after launch

Correct Answer: 1

Explanation

Continuous improvement involves systematically collecting evidence about how an AI system performs in real-world use and using that information to guide controlled enhancements. Relevant evidence can include model performance metrics, user feedback, incidents, data changes, business outcomes, operational metrics, and newly identified risks. Improvements should be evaluated and governed rather than introduced without testing. This may involve retraining, recalibration, feature changes, interface improvements, workflow adjustments, or other interventions. Removing monitoring would make it difficult to identify problems and measure progress. A structured improvement cycle helps ensure that an AI system remains aligned with changing business needs and operating conditions while maintaining appropriate quality, security, and governance controls.