View Full PMI CPMAI Exam Dumps and Practice Test Dumps.

 

Question 361

Which activity is most useful for confirming that an AI solution remains aligned with its original business objective after deployment?

1. Reviewing business outcomes and comparing them with the defined objectives
2. Increasing the number of model parameters
3. Removing user feedback
4. Measuring only server storage capacity

Correct Answer: 1

Explanation

After deployment, an AI solution should be evaluated against the business objectives that originally justified the project. Technical performance alone does not establish whether the solution is delivering meaningful value. Teams can compare actual outcomes with defined indicators such as productivity improvements, cost reductions, processing times, service quality, or other relevant measures. If expected outcomes are not being achieved, the team should investigate whether the issue involves model performance, adoption, workflow integration, data quality, or incorrect assumptions in the original business case. Periodic benefits reviews help ensure that the AI initiative remains aligned with organizational priorities and that corrective actions are taken when outcomes diverge from expectations.

Question 362

Which approach can help reduce the risk of unauthorized changes to an AI dataset used for training?

1. Sharing the dataset through unrestricted public storage
2. Applying access controls, versioning, and audit logging
3. Removing dataset ownership
4. Allowing every project member to modify the original data

Correct Answer: 2

Explanation

Training datasets can directly influence model behavior, so their integrity should be protected. Access controls restrict modification privileges to authorized users, while versioning preserves identifiable dataset states and allows teams to determine which version was used for a particular experiment or deployment. Audit logging records relevant access and modification activities, helping organizations investigate unexpected changes. Additional measures such as data validation, provenance tracking, and secure storage can strengthen protection. Unrestricted modification creates risks of accidental corruption or intentional manipulation. Dataset security should therefore be treated as part of the AI system’s overall security architecture rather than as a separate concern limited to the data-management team.

Question 363

Which metric is particularly useful when evaluating the average magnitude of errors in a regression model without squaring the errors?

1. Mean Absolute Error
2. Precision
3. Recall
4. F1 score

Correct Answer: 1

Explanation

Mean Absolute Error, or MAE, calculates the average absolute difference between predicted and actual values. Unlike Mean Squared Error, MAE does not square the errors, so individual large errors do not receive the same additional weighting that they receive under MSE. This can make MAE easier to interpret because it is expressed in the same units as the target variable. The appropriate regression metric depends on the business problem and the relative importance of different error sizes. Teams may evaluate MAE alongside other measures such as RMSE or R-squared. Using multiple appropriate metrics can provide a more complete understanding of model performance than relying on one measure alone.

Question 364

What should an AI project team do when a newly identified risk could significantly affect the approved project scope?

1. Ignore it until project closure
2. Remove it from the risk register
3. Assess its impact and follow the established escalation and change-control process
4. Automatically expand the project scope

Correct Answer: 3

Explanation

A risk that could materially affect project scope should be assessed promptly to understand its potential consequences. The team should evaluate probability, impact, affected requirements, dependencies, schedule, cost, and expected benefits. If the risk requires a change to the approved scope, the appropriate change-control and governance procedures should be followed. Significant risks may also require escalation to the responsible decision-makers. Automatically changing scope without assessment can create uncontrolled project changes, while ignoring the risk can result in unexpected disruption. Maintaining the risk register and documenting decisions helps stakeholders understand how the risk was handled and provides traceability for subsequent project planning and execution.

Question 365

Which technique can help identify whether a classification model produces different error rates for different groups?

1. Subgroup performance analysis
2. Data compression
3. Model serialization
4. Network optimization

Correct Answer: 1

Explanation

Subgroup performance analysis evaluates model results separately for relevant populations or groups. Teams can compare measures such as false-positive rates, false-negative rates, precision, recall, or other appropriate indicators. This can reveal performance differences that are hidden when results are reported only as an overall average. The interpretation of these differences depends on the use case, data, and applicable organizational requirements. If meaningful disparities are identified, teams should investigate possible causes, including data representation, labeling practices, feature quality, or model behavior. Subgroup analysis is therefore an important component of responsible AI evaluation when differences between populations could materially affect outcomes or user experiences.

Question 366

Which practice can help ensure that AI project documentation remains useful throughout the system lifecycle?

1. Updating relevant documentation when significant changes occur
2. Creating documentation only after the system is retired
3. Preventing all updates after initial approval
4. Deleting previous documentation versions

Correct Answer: 1

Explanation

AI systems evolve as models, datasets, requirements, infrastructure, risks, and operational processes change. Documentation should therefore be maintained throughout the lifecycle rather than treated as a one-time project deliverable. Important documentation may include model information, intended use, limitations, data lineage, configuration, evaluation results, ownership, deployment history, and operational procedures. When significant changes occur, relevant documentation should be updated and versioned so that stakeholders can understand the current state and historical decisions. Keeping previous versions can also support auditability and troubleshooting. Well-maintained documentation improves transparency, knowledge transfer, operational support, governance, and the ability to investigate unexpected behavior after deployment.

Question 367

Which action is most appropriate when an AI model fails an important acceptance criterion before deployment?

1. Deploy it anyway because development is complete
2. Ignore the failed criterion
3. Investigate the cause and determine corrective action before approval
4. Delete the acceptance criterion

Correct Answer: 3

Explanation

An acceptance criterion represents a condition that must be satisfied before the solution is considered acceptable for its intended use. If the model fails an important criterion, the team should investigate why the failure occurred and determine whether remediation is possible. Corrective actions may include improving data quality, modifying features, adjusting the model, changing thresholds, addressing infrastructure issues, or revisiting requirements through formal change control. The model should not automatically proceed to production simply because development activities are complete. If the criterion itself is no longer appropriate, that change should be evaluated and formally approved. This approach ensures deployment decisions are based on evidence and agreed requirements.

Question 368

Which component of a data pipeline is responsible for transforming raw information into a form suitable for model development?

1. Data preprocessing
2. Incident escalation
3. User authentication
4. Model retirement

Correct Answer: 1

Explanation

Data preprocessing transforms raw information into a form that can be appropriately analyzed or used by an AI model. Activities may include cleaning, handling missing values, encoding categorical variables, scaling numerical features, removing or investigating problematic records, and applying other transformations appropriate to the data. Preprocessing should be designed carefully because inconsistent or inappropriate transformations can reduce model quality or introduce leakage. The same relevant transformations should generally be applied consistently during validation and production inference. Teams should document preprocessing steps and, where appropriate, version them alongside the model. A reliable preprocessing pipeline helps create consistent inputs and reduces errors caused by manual data preparation.

Question 369

Which factor is most important when determining whether an AI system requires real-time inference?

1. Whether decisions or responses must be produced within a defined time window
2. The number of pages in the project documentation
3. The age of the development team
4. The number of historical experiments

Correct Answer: 1

Explanation

Real-time inference is appropriate when the AI system must produce predictions or responses within a defined time window to support an operational process. Examples can include interactive applications, transaction decisions, monitoring systems, or other workflows where delayed results reduce usefulness. The required response time should be established from actual business and technical requirements rather than assumed simply because real-time technology is available. Teams should also consider expected traffic, infrastructure, model complexity, reliability, and cost. If immediate responses are not necessary, batch processing may provide a simpler or more economical alternative. Selecting the appropriate inference mode helps align architecture with the actual needs of the business process.

Question 370

Which approach can help reduce the impact of an unexpected model failure in production?

1. Maintaining a tested rollback or fallback procedure
2. Removing monitoring
3. Disabling incident response
4. Deleting the previous model version

Correct Answer: 1

Explanation

A tested rollback or fallback procedure provides an established response when a newly deployed model behaves unexpectedly. Rollback can restore a previously approved version, while a fallback mechanism may direct processing to an alternative model, rule-based process, or human review depending on the application. The appropriate response should be defined according to the system’s risk and operational requirements. Procedures should identify responsibilities, triggers, technical steps, and validation activities. Keeping previous model versions and deployment configurations supports reliable recovery. Testing the rollback process is important because an untested procedure may fail during a high-pressure incident. Recovery planning is a key component of responsible AI operations.

Question 371

Which activity helps identify whether an AI solution has sufficient infrastructure capacity before production deployment?

1. Capacity and performance testing
2. Removing system monitoring
3. Increasing project documentation
4. Changing the model name

Correct Answer: 1

Explanation

Capacity and performance testing helps determine whether infrastructure can support expected workloads and service requirements. Tests may evaluate response time, throughput, concurrency, resource consumption, stability, and behavior under expected or higher loads. AI systems may require substantial computing resources for model inference, data processing, storage, or supporting services. Testing before production can reveal bottlenecks that are difficult to detect in development environments. Results should be compared with predefined requirements or service-level objectives. If capacity is insufficient, the team can consider infrastructure scaling, architecture changes, model optimization, workload scheduling, or other appropriate solutions before exposing users to unacceptable performance.

Question 372

Which practice can help ensure that AI model evaluation is based on data that reflects actual production conditions?

1. Using a representative validation or test dataset
2. Using only the easiest training examples
3. Evaluating only on synthetic examples regardless of use case
4. Removing difficult cases from evaluation

Correct Answer: 1

Explanation

A representative validation or test dataset should reflect the conditions, populations, and variation that the model is expected to encounter in production. If evaluation data is substantially different from operational conditions, performance estimates may be misleading. Teams should consider factors such as time periods, user groups, geographic variation, data distributions, rare cases, and other characteristics relevant to the intended use. Difficult or unusual cases should not automatically be removed because they may represent important production scenarios. Evaluation data should also remain appropriately separated from training data to prevent leakage. Representative testing provides stronger evidence about expected generalization and helps identify weaknesses before deployment.

Question 373

Which control is most useful for tracking who accessed or changed an AI system’s important resources?

1. Audit logging
2. Feature scaling
3. Model compression
4. Hyperparameter tuning

Correct Answer: 1

Explanation

Audit logging records relevant activities such as access attempts, configuration changes, data modifications, model deployments, administrative actions, and other security-sensitive events. These records can support accountability, incident investigation, compliance activities, and troubleshooting. Logs should be protected from unauthorized alteration and retained according to appropriate organizational requirements. Logging should be designed around meaningful events rather than collecting information indiscriminately without a purpose. Audit records can also help establish a timeline during an incident and identify which accounts or services performed specific actions. For AI systems, logging should cover important components across the lifecycle, including data platforms, model repositories, deployment environments, APIs, and administrative interfaces.

Question 374

Which practice is most appropriate for managing personally identifiable information used by an AI system?

1. Collecting and retaining all available personal information
2. Applying data minimization and appropriate privacy controls
3. Making the data publicly accessible
4. Sharing the data with all project members

Correct Answer: 2

Explanation

Data minimization means collecting, processing, and retaining only the personal information necessary for a legitimate and defined purpose. This can reduce privacy exposure and limit the potential impact of unauthorized access or misuse. Appropriate controls may include access restrictions, encryption, pseudonymization or anonymization where suitable, retention limits, monitoring, and documented data-handling procedures. Teams should understand why personal information is required and evaluate whether less sensitive alternatives could meet the same objective. Broadly collecting and sharing personal information increases exposure without necessarily improving the AI solution. Privacy considerations should be addressed throughout the data lifecycle, from acquisition and preparation through model development, deployment, monitoring, and retirement.

Question 375

Which activity is most useful for determining whether an AI model’s inputs have changed significantly since training?

1. Data drift monitoring
2. Model documentation
3. User interface testing
4. Project scheduling

Correct Answer: 1

Explanation

Data drift monitoring evaluates whether the distribution or characteristics of incoming production data have changed compared with the data used during model development. Significant changes may indicate that the model is operating under conditions different from those it was trained to handle. Monitoring can examine relevant feature distributions, missing-value rates, categorical frequencies, or other appropriate indicators. Data drift does not automatically mean that model performance has degraded, but it can provide an early signal that investigation may be necessary. Teams should establish appropriate thresholds and response procedures. Combining drift monitoring with actual performance monitoring provides a stronger basis for determining whether retraining or other intervention is required.

Question 376

Which strategy can help manage risks introduced by an external AI service provider?

1. Conducting vendor due diligence and defining appropriate contractual and operational controls
2. Assuming the provider accepts all responsibility
3. Avoiding any vendor documentation
4. Giving the provider unrestricted access to all organizational data

Correct Answer: 1

Explanation

Third-party AI services can introduce risks involving data handling, security, availability, model behavior, intellectual property, compliance, and operational dependency. Vendor due diligence can evaluate the provider’s security practices, reliability, data policies, technical capabilities, incident response, support arrangements, and relevant assurance evidence. Contracts can establish responsibilities for data protection, service levels, incident notification, changes, access, and termination. Organizations should also limit vendor access to only what is necessary and monitor the service according to its risk. Assuming that a provider automatically accepts all responsibility can create significant governance gaps. Third-party risk should therefore be assessed before adoption and reviewed when the service or relationship changes.

Question 377

Which approach can help determine whether an AI system is understandable enough for its intended users?

1. Evaluating explainability requirements and testing explanations with relevant stakeholders
2. Measuring only model size
3. Increasing the number of hidden layers
4. Removing all user documentation

Correct Answer: 1

Explanation

Explainability requirements depend on the purpose and context of the AI system. Teams should determine what users need to understand about the model’s outputs, limitations, and decision process and then evaluate whether the available explanations meet those needs. Technical explanation methods may include feature importance, example-based explanations, or other model-specific techniques. However, an explanation that is technically available may not be useful to the intended audience. Stakeholder testing can help determine whether explanations are understandable and actionable. Explainability should be considered alongside accuracy, privacy, security, and other requirements. The appropriate level of explanation depends on the consequences of decisions and the responsibilities of the people using the system.

Question 378

Which project activity is most useful for ensuring that lessons from an AI initiative are available for future projects?

1. Conducting a lessons-learned review and documenting relevant findings
2. Deleting project records after deployment
3. Avoiding stakeholder feedback
4. Recording only successful experiments

Correct Answer: 1

Explanation

A lessons-learned review captures useful knowledge about what worked well, what caused problems, and what could be improved in future AI initiatives. Findings may relate to data preparation, stakeholder communication, model development, governance, vendor management, testing, deployment, adoption, or risk management. Both successful and unsuccessful experiences can provide valuable information. Documentation should be accessible to appropriate teams and incorporated into organizational practices where relevant. Conducting the review near project closure can help preserve knowledge while events remain fresh, although lessons can also be captured throughout the lifecycle. This practice supports organizational learning and can reduce the repetition of avoidable problems across future AI projects.

Question 379

Which factor should be evaluated when determining whether a model can operate reliably under changing workloads?

1. Performance, resource utilization, and behavior under expected load conditions
2. Number of model documentation pages
3. Number of project meetings
4. Length of the model name

Correct Answer: 1

Explanation

An AI model and its supporting infrastructure should be evaluated under workloads that reflect expected operational conditions. Important indicators can include latency, throughput, resource utilization, error rates, availability, and stability as request volumes change. Load and stress testing can reveal bottlenecks that may not appear during normal development testing. Teams should establish performance requirements and determine how the system should respond when demand exceeds expected levels. Possible strategies include scaling infrastructure, optimizing the model, queueing requests, or using alternative processing patterns. Reliability under changing workloads is particularly important for systems with variable traffic because a solution that performs well at low volume may become unstable when demand increases.

Question 380

Which practice best supports responsible retirement of an AI model that is no longer required?

1. Following a controlled retirement process that addresses dependencies, data, access, records, and replacement needs
2. Immediately deleting every related record
3. Leaving the model active without monitoring
4. Removing ownership before retirement planning

Correct Answer: 1

Explanation

AI model retirement should be managed as a controlled lifecycle activity. The team should identify dependencies, communicate the retirement decision, determine whether another system or process will replace the model, revoke unnecessary access, preserve records required for governance or auditing, and address relevant data and infrastructure resources. Users and operational teams may need advance notice and transition support. Simply deleting the model can create unexpected service disruptions or destroy information needed for accountability. After retirement, associated credentials, endpoints, scheduled processes, and infrastructure should be reviewed so that unused components do not remain exposed. A structured retirement process helps ensure that the system is removed safely and that organizational obligations continue to be satisfied.