CertSafari

    Free PMI Certified Professional in Managing AI (PMI-CPMAI) Sample Questions

    35 free sample questions from our bank of 335+, covering every exam domain, with answers and detailed explanations. Updated August 2026.

    Practice all 335 questions in quiz mode

    Domain 1: Support Responsible and Trustworthy AI Efforts

    Subdomain 1.4: Monitor regulatory and policy compliance

    1.Which of the following artifacts provides the strongest evidence that an AI project has systematically evaluated and documented fairness compliance?

    1. A.The project risk register with entries tracking identified fairness risks and mitigations.
    2. B.A fairness assessment report containing bias testing methodology, metrics, and validation results.
    3. C.Email correspondence between the team and legal about fairness concerns and resolutions.
    4. D.The official AI ethics policy document stating the organization's commitment to fairness principles.
    Show answer & explanation

    Correct answer: B — A fairness assessment report containing bias testing methodology, metrics, and validation results.

    • A. Incorrect. While the project risk register is important for tracking risks, it does not directly demonstrate compliance with fairness requirements. It focuses on risk management rather than providing detailed evidence of bias testing, metrics, and validation results needed for systematic fairness evaluation.
    • B. Correct. A fairness assessment report provides concrete evidence of compliance by documenting the methodology, metrics, and results of bias testing. This is the strongest artifact for proving systematic evaluation and documentation of fairness, which is critical for regulatory and policy adherence.
    • C. Incorrect. Email correspondence may show communication about fairness concerns but is informal and incomplete evidence. It does not serve as a formal, auditable artifact to demonstrate structured assessment or documented validation of fairness.
    • D. Incorrect. The AI ethics policy document outlines the organization's commitment to fairness principles but does not provide evidence of actual compliance or implementation of fairness measures in a specific project. It is a governance document rather than project-level assessment evidence.

    Subdomain 1.2: Manage AI/ML transparency (e.g., data selection, algorithm selection)

    2.In the context of AI transparency, what does "data provenance" refer to?

    1. A.It refers to the statistical distribution of the dataset features.
    2. B.It refers to the documented origin and processing history of the data.
    3. C.It refers to the process of handling missing values in the dataset.
    4. D.It refers to the accuracy metrics of models trained on the data.
    5. E.It refers to the storage infrastructure used for the data.
    Show answer & explanation

    Correct answer: B — It refers to the documented origin and processing history of the data.

    • A. Incorrect. Data provenance is not about the statistical distribution of dataset features. Statistical distribution describes the spread and characteristics of data values, not their origin or history. Provenance focuses on where data came from and how it was transformed over time, which is essential for transparency and traceability in AI systems.
    • B. Correct. Data provenance refers to the documented origin, lineage, and processing history of the data, including collection, transformations, and any handling steps. This record is critical for transparency, auditability, and trust in AI systems, as it provides a clear trail of data's journey from source to use.
    • C. Incorrect. Handling missing values is a data preprocessing step, not data provenance. While preprocessing activities may be part of the data's history, provenance encompasses the broader record of origin and all changes, not just cleaning or imputation tasks.
    • D. Incorrect. Accuracy metrics of models evaluate performance after training, not the source or lineage of the data. Provenance is about traceability of the dataset itself, not the performance of models trained on it.
    • E. Incorrect. Storage infrastructure describes where data is physically or logically stored, which is separate from provenance. Provenance is concerned with the data's origin, ownership, and processing history, not its storage system.

    Subdomain 1.3: Conduct bias checks (e.g., model, data, algorithm)

    3.An e-commerce recommendation system is suspected of economic bias: it recommends higher-priced items to affluent users and lower-priced items to less affluent users. To quantitatively monitor this issue, which metric should the project manager include in the fairness dashboard?

    1. A.Monitor the demographic parity of recommended items across different user groups.
    2. B.Track equality of opportunity by comparing the rate of high-value purchases across segments.
    3. C.Evaluate the average profit generated per user across different income brackets to detect economic bias.
    4. D.Compare the average price of recommended items between high-income and low-income user segments.
    5. E.Assess the click-through rate for recommendations specifically for users in different income groups.
    Show answer & explanation

    Correct answer: D — Compare the average price of recommended items between high-income and low-income user segments.

    • A. Incorrect. Demographic parity measures whether the recommendation system selects items at the same rate across groups, but it does not directly address economic bias in pricing. The concern is about the price distribution of recommended items, not just recommendation frequency.
    • B. Incorrect. Equality of opportunity focuses on true positive rates (e.g., high-value purchases) across groups, but it does not directly measure the pricing bias in recommendations. A metric tied to recommendation price differences is more appropriate.
    • C. Incorrect. Average profit per user may be a useful business KPI, but it does not directly quantify whether the recommendation system is treating income groups differently in the prices it suggests. It reflects downstream financial outcomes rather than the bias itself.
    • D. Correct. Comparing the average price of recommended items between high-income and low-income segments directly measures whether the system is systematically recommending more expensive or cheaper items based on income. This is a clear quantitative fairness indicator for the suspected economic bias.
    • E. Incorrect. Click-through rate measures user engagement, not the economic bias in the pricing of recommended items. A group difference in CTR could exist for many reasons and would not directly answer the fairness concern regarding item prices.

    Subdomain 1.5: Manage accountability documentation and audit trail

    4.Which of the following documents best establishes accountability for a deployment decision in an AI project?

    1. A.The project schedule that documents the planned deployment date and includes sign-offs on all milestone completions.
    2. B.The signed approval form from the executive sponsor accompanied by the risk assessment report from that decision point.
    3. C.The slide deck used to present the project status, risk analysis, and deployment recommendation to stakeholders.
    4. D.A signed confirmation from the DevOps team that the production infrastructure is fully prepared and meets all security requirements.
    5. E.A report detailing the final model's accuracy on the test set and comparing it to baseline performance metrics.
    Show answer & explanation

    Correct answer: B — The signed approval form from the executive sponsor accompanied by the risk assessment report from that decision point.

    • A. Incorrect. The project schedule and milestone sign-offs track progress but do not document accountability for deployment decisions or risk acceptance.
    • B. Correct. The signed approval form from the executive sponsor, along with the risk assessment report, directly ties accountability to a decision-maker and records the rationale, creating a formal audit trail.
    • C. Incorrect. A slide deck is a communication artifact and lacks formal sign-off, making it insufficient as a durable record of accountability.
    • D. Incorrect. DevOps confirmation addresses technical readiness but does not capture governance approval or decision accountability.
    • E. Incorrect. Model performance reporting is important for validation but does not by itself document accountability or the rationale for deployment.

    Domain 2: Identify Business Needs and Solutions

    Subdomain 2.6: Manage adoption/integration risks

    5.Which of the following is considered a leading indicator of user adoption for a newly deployed AI application?

    1. A.End-user satisfaction survey scores collected after six months.
    2. B.Number of active users logging into the application on a weekly basis.
    3. C.Reduction in average handling time for the process being automated.
    4. D.Return on investment calculated over the first year of operation.
    Show answer & explanation

    Correct answer: B — Number of active users logging into the application on a weekly basis.

    • A. Incorrect. Satisfaction survey scores collected after six months are a lagging indicator, reflecting past experiences rather than early adoption behavior.
    • B. Correct. Weekly active users measure real-time engagement and signal early adoption trends before long-term outcomes are realized, making it a leading indicator.
    • C. Incorrect. Reduction in average handling time is an operational outcome metric and a lagging indicator; it reflects efficiency gains that may occur even with uneven adoption.
    • D. Incorrect. ROI over the first year is a lagging financial measure that depends on accumulated results and does not indicate early user engagement.

    Subdomain 2.4: Develop AI project scope statement

    6.Which of the following options best represents a well-defined AI project scope statement for a fraud detection initiative?

    1. A.Achieve 95% precision in fraud detection accuracy, with the model retrained on a weekly basis.
    2. B.Reduce fraud losses by 30% while maintaining a false positive rate below 0.1%.
    3. C.Achieve an average improvement of 20 points in the credit scores of monitored customers.
    4. D.Ensure the AI model can process a throughput of 10,000 transactions per second.
    Show answer & explanation

    Correct answer: B — Reduce fraud losses by 30% while maintaining a false positive rate below 0.1%.

    • A. Incorrect. While precision is a valid model metric, the scope statement should emphasize business outcomes rather than technical specifications like retraining frequency. Precision alone does not address the business impact of fraud detection.
    • B. Correct. This option directly ties the AI project to a measurable business outcome (reducing financial losses) while setting a practical constraint (false positive rate) to ensure cost-effectiveness. It aligns with an AI project scope statement by defining business value and success criteria.
    • C. Incorrect. Improving credit scores is not a relevant business outcome for a fraud detection system. This metric is unrelated to the primary goal of detecting and preventing fraud.
    • D. Incorrect. Throughput is a technical non-functional requirement, not a business outcome. While important for system design, it does not describe the project's scope in terms of business value or impact.

    Subdomain 2.8: Define success criteria (e.g., KPIs, metrics)

    7.A manufacturing plant has deployed an AI-based predictive maintenance system for its assembly line robots. The operations manager's primary objective is to minimize production interruptions. Which success metric should be prioritized?

    1. A.Track the reduction in unplanned production downtime.
    2. B.Monitor the accuracy of failure predictions.
    3. C.Measure the decrease in total maintenance expenses.
    4. D.Assess the increase in overall equipment effectiveness.
    5. E.Evaluate the decline in false maintenance alerts.
    Show answer & explanation

    Correct answer: A — Track the reduction in unplanned production downtime.

    • A. Correct. Reducing unplanned production downtime directly measures the primary objective of minimizing production interruptions. For a predictive maintenance system, this is the most relevant outcome metric because it reflects whether the system is preventing unexpected stoppages.
    • B. Incorrect. Prediction accuracy is an important model performance metric, but it does not directly measure production interruptions. A highly accurate model is only valuable if it leads to fewer interruptions; it is a secondary metric.
    • C. Incorrect. Lower maintenance expenses are a financial benefit but not the stated primary objective. The manager is focused on operational continuity, not cost reduction.
    • D. Incorrect. Overall equipment effectiveness (OEE) is a broader composite KPI that includes availability, performance, and quality. It does not specifically target unplanned downtime and can be influenced by factors other than interruptions.
    • E. Incorrect. Fewer false alerts improve system usability and may reduce unnecessary maintenance, but they do not directly measure the main business impact of preventing production stoppages.

    Subdomain 2.1: Identify problem to be solved (e.g., needs, persona)

    8.A logistics firm has a high volume of customer emails asking for shipment status updates. The team defines the problem as "customers cannot easily get real-time shipment status." After conducting stakeholder interviews, they discover that the majority of inquiries involve complex multi-leg shipments where tracking information is fragmented across carriers. The project manager needs to map this problem to an appropriate AI pattern. Which AI pattern best addresses the underlying need?

    1. A.Apply regression to predict shipment delays using historical data
    2. B.Use natural language processing to extract data and generate status responses
    3. C.Deploy reinforcement learning to optimize complex shipment routing decisions
    4. D.Implement clustering to categorize inquiries for analysis and reporting
    Show answer & explanation

    Correct answer: B — Use natural language processing to extract data and generate status responses

    • A. Incorrect. A regression model predicts numerical outcomes such as delays, but does not address the core need of extracting and synthesizing fragmented tracking data from customer inquiries or generating responses. The problem is about information retrieval and response generation, not forecasting.
    • B. Correct. Natural language processing (NLP) can parse customer emails to extract shipment details, query fragmented tracking systems, and generate coherent status responses, directly solving the problem of providing real-time updates from fragmented multi-carrier data.
    • C. Incorrect. Reinforcement learning is typically used to optimize sequential decision-making, such as routing or operational policies. The problem described is about consolidating and communicating existing shipment status information, not making routing decisions.
    • D. Incorrect. Clustering can group similar inquiries for analysis, but it does not solve the customer-facing need for real-time shipment status access. It is useful for reporting and categorization, not for extracting status data and generating responses.

    Subdomain 2.2: Evaluate initial AI feasibility

    9.What is the primary purpose of assessing data readiness during the initial feasibility evaluation of an AI project?

    1. A.To determine whether the organization has the necessary hardware to store the data.
    2. B.To verify that needed data exists in a usable format and sufficient amount for training.
    3. C.To evaluate the performance of existing machine learning models on sample datasets.
    4. D.To identify potential biases in the data that may affect model fairness.
    5. E.To ensure that data can be accessed in real time for streaming applications.
    Show answer & explanation

    Correct answer: B — To verify that needed data exists in a usable format and sufficient amount for training.

    • A. Incorrect. While hardware capacity can be relevant, the primary feasibility concern is whether the data itself is available and usable, not the infrastructure to store it.
    • B. Correct. A core step in evaluating AI feasibility is confirming that the required data exists, is usable, and is available in sufficient quantity and quality to support model training. Without adequate data, the AI solution is not feasible.
    • C. Incorrect. Evaluating the performance of existing models is part of later stages (e.g., development or validation), not initial feasibility assessment, which focuses on foundational data requirements.
    • D. Incorrect. Identifying biases is important for responsible AI and should be addressed during data preprocessing or model evaluation, but it is not the primary criterion for assessing initial feasibility.
    • E. Incorrect. Real-time data access is a technical requirement for specific use cases (e.g., streaming), but it is not a universal feasibility criterion for all AI projects.

    Subdomain 2.9: Support business case creation

    10.Which approach best estimates the cost of using a third-party API for an AI project given uncertain monthly request volumes and tiered pricing?

    1. A.Calculate the cost using the average monthly request volume from past ticket data and apply the API provider’s list price.
    2. B.Perform a Monte Carlo simulation that models the distribution of request volumes and pricing tiers to generate a probabilistic cost forecast.
    3. C.Negotiate a fixed-price contract with the API provider for a predetermined volume of requests per month to eliminate variability.
    4. D.Use a three-point estimate (optimistic, most likely, pessimistic) based on expert judgment for volumes and pricing.
    Show answer & explanation

    Correct answer: B — Perform a Monte Carlo simulation that models the distribution of request volumes and pricing tiers to generate a probabilistic cost forecast.

    • A. Incorrect. Using average monthly request volume and list price produces a single-point estimate that ignores variability in demand and pricing tiers, leading to an oversimplified and potentially inaccurate cost estimate that underestimates risk.
    • B. Correct. A Monte Carlo simulation models the uncertainty in request volumes and pricing tiers, providing a probabilistic range of costs that accounts for variability and risk. This data-driven approach is well-suited for business case creation, as it shows expected costs and risk distribution.
    • C. Incorrect. A fixed-price contract is a procurement decision to reduce cost uncertainty, not a forecasting method. It may not be feasible or cost-effective if actual usage deviates from the predetermined volume, leading to overpayment or underutilization.
    • D. Incorrect. A three-point estimate accounts for some variability, but it is less robust than a Monte Carlo simulation when historical data is available. It provides a simplified estimate rather than a comprehensive probabilistic forecast.

    Subdomain 2.7: Draft AI solution

    11.Which of the following should be included in the architecture of an AI solution?

    1. A.It should include a data pipeline for ingesting, cleaning, and transforming raw data into features for training and inference.
    2. B.It should include a centralized user interface for data scientists to manually label training data without version control.
    3. C.It should use a monolithic application that tightly couples data storage, model training, and inference code.
    4. D.It should use an isolated on-premises GPU cluster dedicated to training, with no connection to production.
    5. E.It should rely on a spreadsheet-based workflow for tracking model versions and experiment results.
    Show answer & explanation

    Correct answer: A — It should include a data pipeline for ingesting, cleaning, and transforming raw data into features for training and inference.

    • A. Correct. A data pipeline is essential for ingesting, cleaning, and transforming raw data into features, ensuring high-quality, consistent data for both training and inference. This is a core architectural component because model quality depends heavily on reliable, repeatable data processing.
    • B. Incorrect. While labeling tools can be part of an AI solution, a centralized user interface without version control is inefficient and error-prone. Version control is critical for tracking changes, ensuring reproducibility, and governing training data.
    • C. Incorrect. A monolithic design tightly coupling data storage, training, and inference reduces flexibility, scalability, and maintainability. Modern AI solutions should use modular, decoupled architectures to allow independent evolution.
    • D. Incorrect. An isolated on-premises GPU cluster with no connection to production is impractical. AI solutions require integration with production systems for deployment, monitoring, and continuous improvement through feedback loops.
    • E. Incorrect. Spreadsheet-based workflows for tracking model versions and experiments are not scalable or reliable. Dedicated tools like MLflow, Weights & Biases, or DVC provide better experiment tracking, reproducibility, collaboration, and auditability.

    Subdomain 2.7: Draft AI solution

    12.A team has developed a machine learning model to classify support tickets. Over time, they observe a decline in accuracy due to changing ticket patterns. Which approach should they adopt to address model drift?

    1. A.Retrain the model every month from scratch using only the original training data to avoid drift and maintain stability.
    2. B.Set up continuous monitoring for data drift and confidence and implement periodic retraining with new labeled data.
    3. C.Have analysts manually review a random sample of predictions each quarter and update rules if errors are found.
    4. D.Add more layers periodically to the model architecture to increase capacity and capture new patterns without retraining.
    5. E.Replace the model with a static rule-based system using keywords to classify tickets, eliminating model drift.
    Show answer & explanation

    Correct answer: B — Set up continuous monitoring for data drift and confidence and implement periodic retraining with new labeled data.

    • A. Incorrect. Retraining from scratch each month using only the original training data does not incorporate new data patterns, thus failing to address drift and potentially reinforcing outdated biases.
    • B. Correct. Continuous monitoring for data drift and confidence detects when the model is misaligned with current patterns, and periodic retraining with new labeled data is a best practice to adapt the model and maintain accuracy.
    • C. Incorrect. Manual quarterly review of a random sample is too infrequent and limited to reliably detect or correct ongoing drift; it lacks systematic monitoring and retraining.
    • D. Incorrect. Adding more layers increases complexity but does not address outdated training data; without retraining on new examples, the model still degrades.
    • E. Incorrect. A static rule-based system eliminates model drift but sacrifices adaptability and the benefits of machine learning for dynamic environments.

    Subdomain 2.10: Identify project resources (e.g., people, hardware, contractors)

    13.Which team composition is best suited for supporting the full AI/ML lifecycle from data preparation to deployment and monitoring?

    1. A.Include a data engineer, ML engineer, domain expert, and DevOps engineer for full ML lifecycle.
    2. B.Include a data scientist, project manager, business analyst, and software developer for model experimentation.
    3. C.Include a machine learning researcher, data labeler, QA tester, and UX designer for research and testing.
    4. D.Include a cloud architect, security specialist, product owner, and scrum master to ensure compliance and agility.
    Show answer & explanation

    Correct answer: A — Include a data engineer, ML engineer, domain expert, and DevOps engineer for full ML lifecycle.

    • A. Correct. A data engineer builds data pipelines, an ML engineer develops and deploys models, a domain expert provides business context and validation, and a DevOps engineer ensures CI/CD, infrastructure, and monitoring. Together they cover the full ML lifecycle from data to production.
    • B. Incorrect. This team supports early experimentation or prototyping but lacks dedicated data engineering, deployment, and operational roles needed for the full lifecycle. It is more suited to discovery phases.
    • C. Incorrect. These roles focus on research, data labeling, testing, and user experience, but do not cover core data pipeline engineering, model deployment, or ongoing operational monitoring required for production systems.
    • D. Incorrect. While a cloud architect and security specialist handle infrastructure and compliance, and the product owner and scrum master manage delivery, the team lacks data engineering, ML development, and model operations roles essential for the AI/ML lifecycle.

    Subdomain 2.5: Determine ROI

    14.An automotive manufacturer is calculating the ROI for an AI-based quality inspection system on the assembly line. There is a 20% probability that variations in factory lighting could reduce the AI's accuracy, diminishing projected savings. The project manager must incorporate this risk into the financial analysis. Which method is most aligned with standard corporate finance practice?

    1. A.Adjust the discount rate upward to incorporate the risk of lower accuracy.
    2. B.Multiply the annual projected savings by the probability of success to adjust for risk.
    3. C.Extend the required payback period to accommodate potential reductions in accuracy.
    4. D.Add a proportional risk premium to the initial investment cost to cover uncertainties.
    Show answer & explanation

    Correct answer: B — Multiply the annual projected savings by the probability of success to adjust for risk.

    • A. Incorrect. Adjusting the discount rate upward (risk-adjusted discount rate) is used for systematic or market-wide risks, not for a specific, quantifiable operational risk like lighting variability. This approach would mix time-value adjustments with operational risk imprecisely.
    • B. Correct. This method uses expected value analysis or risk-adjusted cash flows, a standard corporate finance practice. By multiplying projected savings by the probability of success (80%), the expected savings reflect the risk directly, providing a more accurate financial assessment.
    • C. Incorrect. Extending the payback period does not quantify the probability or magnitude of the risk. It is a simplistic screening metric that does not adjust the expected savings for probabilistic outcomes, making it unsuitable for precise financial analysis.
    • D. Incorrect. Adding a risk premium to the initial investment cost is not standard practice for modeling uncertainty in projected savings. This approach inflates costs arbitrarily rather than adjusting expected cash inflows, which is the appropriate treatment for this type of risk.

    Subdomain 2.3: Conduct risk assessment(s) (e.g., security, safety, ethics)

    15.A pharmaceutical team uses an AI model to predict drug toxicity. They want to systematically identify and assess potential failures in the AI pipeline before deployment. Which method should they use?

    1. A.Apply FMEA to the AI pipeline, analyzing each part from data intake to final predictions.
    2. B.Conduct clinical trials on the specific drug involved to confirm the model's initial toxicity prediction.
    3. C.Add more computational power to cut inference time, which may incidentally boost accuracy.
    4. D.Switch from the AI model to a conventional QSAR approach for toxicity prediction.
    Show answer & explanation

    Correct answer: A — Apply FMEA to the AI pipeline, analyzing each part from data intake to final predictions.

    • A. Correct. Failure Mode and Effects Analysis (FMEA) is a structured method for identifying potential failure points, their causes, and impacts across a process. Applying it to the AI pipeline from data intake through prediction enables systematic assessment of safety, reliability, and operational risks, making it suitable for AI risk assessment.
    • B. Incorrect. Clinical trials evaluate the actual drug’s toxicity and efficacy, not the AI system’s failure modes or process risks. While they may validate predictions, they do not serve as a structured assessment of where the AI model itself could fail.
    • C. Incorrect. Increasing computational power addresses performance optimization (e.g., speed) but does not directly assess or mitigate risks like bias, data quality, or model failures. It is not a formal risk assessment method.
    • D. Incorrect. Switching to a conventional QSAR approach changes the solution but does not assess risks in the existing AI model. It is a workaround, not a structured method for understanding potential failures before decision-making.

    Domain 3: Identify Data Needs

    Subdomain 3.2: Identify data SMEs

    16.Which action best helps identify data subject matter experts (SMEs) for data needs?

    1. A.Interview daily CRM users to confirm data accuracy and real-world fidelity.
    2. B.Ask data quality analysts to profile the data and highlight outliers.
    3. C.Request system administrators to check database integrity and transaction logs.
    4. D.Have data architects examine the data model for potential inaccuracies.
    Show answer & explanation

    Correct answer: A — Interview daily CRM users to confirm data accuracy and real-world fidelity.

    • A. Correct. Daily CRM users are operational stakeholders who directly interact with data in business processes. Interviewing them not only validates data accuracy but also helps identify individuals with deep domain knowledge, making them a key source for identifying data SMEs.
    • B. Incorrect. Data quality analysts focus on technical profiling and outlier detection, which provides data quality insights but does not directly identify the people who are subject matter experts on data meaning and usage.
    • C. Incorrect. System administrators check database integrity and logs, which is infrastructure-related. Their expertise is technical, not business-domain-specific, and does not lead to identifying data SMEs.
    • D. Incorrect. Data architects examine data models and structures, which can reveal design issues but not necessarily identify the business experts who understand the operational context and data needs.

    Subdomain 3.5: Gather required data

    17.You are managing the migration of 15 TB of labeled image data from on-premises network-attached storage to a cloud-based AI development environment. The source storage has intermittent I/O bottlenecks, and the cloud environment requires the data in a specific bucket structure with metadata files. What transfer strategy maximizes throughput while ensuring data integrity and minimal downtime?

    1. A.Establish a dedicated high-bandwidth network link and use rsync with checksum verification, scheduling the transfer during a maintenance window to minimize downtime.
    2. B.Deploy a cloud storage gateway on premises that caches and asynchronously replicates data to the cloud, using object versioning to detect changes and ensure integrity.
    3. C.Split the dataset into smaller chunks, compress each with integrity hashes, and use parallel threads of a cloud CLI tool to upload, with automatic retries and post-transfer checks.
    4. D.Ship a physical storage device to the cloud provider for initial bulk import, then use differential sync for any updates and verify data integrity after transfer.
    Show answer & explanation

    Correct answer: C — Split the dataset into smaller chunks, compress each with integrity hashes, and use parallel threads of a cloud CLI tool to upload, with automatic retries and post-transfer checks.

    • A. Incorrect. While a dedicated link and rsync with checksum verification can improve transfer speed and ensure integrity, the source I/O bottlenecks on the NAS remain a constraint, and scheduling during a maintenance window increases downtime rather than minimizing it. This approach is suboptimal for a large one-time migration.
    • B. Incorrect. A cloud storage gateway is designed for ongoing hybrid access and caching, not for maximizing throughput in a one-time bulk migration. Asynchronous replication can delay the availability of the required bucket structure and metadata, and it does not directly address the source I/O bottlenecks or minimize downtime.
    • C. Correct. Splitting the dataset into smaller chunks allows parallel uploads, which maximizes throughput and mitigates the intermittent I/O bottlenecks at the source. Compression with integrity hashes ensures data integrity, and automatic retries with post-transfer integrity checks provide reliability. This method can also be adapted to create the required bucket structure and metadata files with minimal downtime.
    • D. Incorrect. Shipping a physical device is effective for very large datasets when network bandwidth is extremely limited, but 15 TB is often manageable via network transfer with parallelization. This approach incurs additional operational overhead and delays, and the need for initial import followed by differential sync may not minimize downtime or handle I/O bottlenecks effectively.

    Subdomain 3.4: Coordinate AI workspace and infrastructure

    18.An AI team needs to manage data storage costs while ensuring data remains accessible for future model training and audits. Which approach best meets these requirements?

    1. A.Keep all data in the standard storage class perpetually, assuming retrieval costs are insignificant.
    2. B.Transition data to a cheaper storage class after 30 days and archive after 90 days using lifecycle policies.
    3. C.Store all data on SSD-backed block storage and compress files older than 30 days to reduce space.
    4. D.Purge data that is older than 30 days to eliminate storage costs, given low probability of reuse.
    Show answer & explanation

    Correct answer: B — Transition data to a cheaper storage class after 30 days and archive after 90 days using lifecycle policies.

    • A. Incorrect. Keeping all data in the standard storage class perpetually is not cost-effective because standard storage is typically more expensive, and retrieval costs accumulate for infrequently accessed data. This does not follow data lifecycle best practices.
    • B. Correct. Transitioning data to a cheaper storage class after 30 days and archiving after 90 days using lifecycle policies optimizes cost while maintaining accessibility for potential reuse. This aligns with best practices for AI/ML data management by matching storage tiers to data usage patterns.
    • C. Incorrect. SSD-backed block storage is expensive and intended for high-performance workloads, not long-term retention. Compressing files older than 30 days saves space but does not address the cost inefficiency of keeping all data on high-performance storage.
    • D. Incorrect. Purging data older than 30 days risks losing valuable historical data needed for retraining, audits, or compliance. This aggressive approach undermines data governance and business continuity.

    Subdomain 3.6: Check data privacy, compliance, and access

    19.An AI development team has built a predictive maintenance model using a dataset licensed from an industrial consortium. The license prohibits commercial use and redistribution. The company now intends to offer the model as a cloud-based API service to paying customers. What should the project manager do to ensure compliance?

    1. A.Proceed with deployment but implement strict access controls to prevent data leakage, as the license only restricts redistribution of the raw data.
    2. B.Negotiate a new or amended license with the data provider that permits commercial use before deploying the model as a service.
    3. C.Since the model does not directly resell the dataset, deploying it as a service does not constitute commercial use under standard licensing terms.
    4. D.Use model distillation to create a new model that does not directly rely on the original dataset, thereby circumventing the licensing restriction.
    5. E.Anonymize the dataset and then retrain the model on the anonymized version, as anonymized data is no longer subject to the original license.
    Show answer & explanation

    Correct answer: B — Negotiate a new or amended license with the data provider that permits commercial use before deploying the model as a service.

    • A. Incorrect. Access controls protect raw data but do not change the fact that the intended use is commercial and violates the license. Deploying as a paid service constitutes commercial use, and the license explicitly prohibits it.
    • B. Correct. The project manager must obtain a new or amended license that explicitly permits commercial use before offering the model as a paid service. This is the appropriate compliance action given the current license restrictions.
    • C. Incorrect. Offering the model as a paid API service is a form of commercial use, even if the raw dataset is not resold. Standard licensing terms typically govern downstream commercial exploitation, not just literal resale of data.
    • D. Incorrect. Model distillation does not automatically eliminate licensing obligations; the derived model may still depend on the restricted dataset. Using a technical workaround to bypass contractual terms is not a compliance solution.
    • E. Incorrect. Anonymization reduces privacy risks but does not override contractual license restrictions. Even if anonymized, the original license still governs permitted use, including commercial deployment.

    Subdomain 3.3: Identify data sources and locations

    20.A healthcare organization is merging its electronic health record (EHR) system with a newly acquired clinic’s patient management software. The AI project manager must catalog all data sources to build a unified patient risk scoring model while ensuring HIPAA compliance. What is the best strategy to identify and document the combined data landscape?

    1. A.Grant the AI team full access to both systems’ production databases and let them self-discover tables containing patient data.
    2. B.Engage data stewards to jointly map data flows, classify PII, and register sources in a governed catalog with access controls.
    3. C.Extract all data into a de-identified research repository first, then perform analysis to determine what was originally present.
    4. D.Rely on the EHR vendor’s built-in data lineage reports and assume they fully cover the legacy clinic software.
    5. E.Use a network scanning tool to detect all database instances and then assign each to an owner post-hoc.
    Show answer & explanation

    Correct answer: B — Engage data stewards to jointly map data flows, classify PII, and register sources in a governed catalog with access controls.

    • A. Incorrect. Granting full production access violates HIPAA’s principle of least privilege and risks unauthorized exposure of PHI. Self-discovery without governance leads to incomplete or inconsistent documentation and lacks controlled lineage tracking.
    • B. Correct. Engaging data stewards from both organizations enables a collaborative, governed approach to map data flows, classify PII, and register sources in a catalog with access controls. This aligns with HIPAA requirements for data governance and accountability, ensuring comprehensive documentation of the merged landscape.
    • C. Incorrect. De-identifying data first may obscure critical metadata and relationships needed to understand the original landscape. This approach risks losing context required for accurate cataloging and compliance, and does not solve the problem of identifying source locations.
    • D. Incorrect. Relying solely on vendor reports assumes completeness and accuracy, but they may not cover customizations or integrations of the legacy clinic software. This leads to blind spots and does not establish a controlled, organization-wide inventory of data assets.
    • E. Incorrect. Network scanning can detect database instances but does not provide context about data content, sensitivity, ownership, or business meaning. Assigning owners post-hoc is weak governance and does not adequately address HIPAA compliance or comprehensive documentation.

    Subdomain 3.7: Oversee data evaluation

    21.Which data quality dimension is most directly concerned with whether the data is appropriate for the specific AI use case?

    1. A.Completeness
    2. B.Consistency
    3. C.Accuracy
    4. D.Timeliness
    5. E.Relevance
    Show answer & explanation

    Correct answer: E — Relevance

    • A. Incorrect. Completeness refers to whether all required data elements are present. While important, it does not directly address suitability for the specific AI objective; a complete dataset may still be irrelevant.
    • B. Incorrect. Consistency ensures data values are uniform across systems and time. This maintains reliability but does not evaluate whether the data is appropriate for the intended use case.
    • C. Incorrect. Accuracy measures how close data is to true values. Accurate data can still be irrelevant if it does not support the specific decision or model being built.
    • D. Incorrect. Timeliness refers to the currency of data, ensuring it is up-to-date. This is critical for dynamic contexts but does not primarily assess fit-for-purpose.
    • E. Correct. Relevance directly evaluates whether the data is useful and appropriate for the specific AI model or business objective. It ensures the data supports the intended analysis, avoiding noise and improving model performance.

    Subdomain 3.9: Convey data understanding to leadership

    22.As the data strategy lead for an AI-driven customer segmentation project, you have just completed a comprehensive data quality assessment. Your next task is to prepare an executive summary for the Chief Data Officer and other senior leaders. Which of the following best describes the content focus of an effective executive summary in this context?

    1. A.Detailed technical metrics, e.g., missing value percentages, data type distributions, and correlation matrices.
    2. B.An impact analysis highlighting top data issues affecting project success, mapped to risks and recommended actions.
    3. C.A complete inventory of all data sources, including database credentials and access logs, to facilitate technical review.
    4. D.Step-by-step code snippets and transformation scripts used during data cleaning to demonstrate thoroughness.
    Show answer & explanation

    Correct answer: B — An impact analysis highlighting top data issues affecting project success, mapped to risks and recommended actions.

    • A. Incorrect. Detailed technical metrics are too granular for an executive summary; senior leaders need a concise synthesis of business implications, not exhaustive statistical details. Such metrics are better suited for technical appendices or analyst reviews.
    • B. Correct. An effective executive summary translates data quality findings into business impact, highlighting top issues, risks to project success, and recommended actions. This strategic focus enables leadership to make informed prioritization and resourcing decisions.
    • C. Incorrect. A complete inventory of data sources and access logs is operationally useful but too detailed and security-sensitive for an executive summary. Leadership requires a high-level understanding of data readiness, not system-level documentation.
    • D. Incorrect. Step-by-step code snippets and transformation scripts are implementation details for technical audiences. An executive summary should concentrate on key findings, business implications, and decision points, not technical artifacts.

    Domain 4: Manage AI Model Development and Evaluation

    Subdomain 4.1: Oversee AI/ML model technique(s) (e.g., algorithm, selection)

    23.Which statement accurately distinguishes reinforcement learning from supervised learning?

    1. A.Reinforcement learning requires a labeled dataset, while supervised learning learns from environment feedback.
    2. B.Reinforcement learning learns from rewards and penalties via interaction, while supervised learning from input-output pairs.
    3. C.Supervised learning works with unlabeled data, while reinforcement learning always requires a pre-trained model.
    4. D.Reinforcement learning is used exclusively for image recognition, while supervised learning handles sequential decisions.
    5. E.Supervised learning models are always much more interpretable than reinforcement learning policies.
    Show answer & explanation

    Correct answer: B — Reinforcement learning learns from rewards and penalties via interaction, while supervised learning from input-output pairs.

    • A. Incorrect. This reverses the core distinction: supervised learning relies on labeled datasets (input-output pairs), while reinforcement learning learns through interaction with an environment and feedback signals such as rewards and penalties.
    • B. Correct. Reinforcement learning uses rewards and penalties to learn actions through interaction with an environment. Supervised learning, in contrast, learns from labeled examples (input-output pairs) where the correct answers are provided.
    • C. Incorrect. Supervised learning typically uses labeled data, not unlabeled data. Reinforcement learning does not always require a pre-trained model; it often starts from scratch and improves through experience.
    • D. Incorrect. Reinforcement learning is not limited to image recognition; it is commonly used for sequential decision-making tasks like robotics and games. Supervised learning is broader and applies to classification, regression, and many other prediction tasks.
    • E. Incorrect. Interpretability depends on the specific model and use case, not on whether the method is supervised or reinforcement learning. Both can have varying levels of interpretability.

    Subdomain 4.2: Oversee AI/ML model QA/QC (e.g., configuration management, model performance)

    24.Which of the following practices is most important for ensuring configuration management and reproducibility in AI/ML model QA/QC?

    1. A.Automate model deployment through a continuous integration and delivery pipeline.
    2. B.Store model artifacts and preprocessing code together as an immutable package.
    3. C.Increase the retraining frequency of the model to adapt to changes.
    4. D.Implement A/B testing to gradually roll out new model versions to users.
    5. E.Set up automated alerting to detect performance drops in production.
    Show answer & explanation

    Correct answer: B — Store model artifacts and preprocessing code together as an immutable package.

    • A. CI/CD automation supports deployment efficiency and consistency, but it does not directly ensure configuration management or reproducibility. While helpful for operational QA, it is not the core practice for traceable, version-controlled model artifacts.
    • B. Correct. Storing model artifacts and preprocessing code together as an immutable package ensures full reproducibility, traceability, and auditability. This directly addresses configuration management by preserving the exact model version, dependencies, and transformation logic, which is essential for QA/QC.
    • C. Increasing retraining frequency may help adapt to data drift, but it does not inherently enforce configuration management or reproducibility. Without proper validation and monitoring, frequent retraining can introduce instability and does not serve as a QA/QC control.
    • D. A/B testing is a deployment strategy for comparing model versions and mitigating rollout risk, but it does not provide strong configuration management or immutable versioning. It is not the primary practice for reproducibility in QA/QC.
    • E. Automated alerting for performance drops is valuable for production monitoring and proactive issue detection, but it does not address configuration management or reproducibility. It supports performance oversight rather than the core QA/QC requirement of preserving a controlled, reproducible model package.

    Subdomain 4.3: Manage AI/ML model training

    25.During the training of a large language model, the project lead notices that GPU utilization fluctuates between 40% and 60% over several hours, and the training loss is decreasing very slowly. The training job uses a data loader that feeds batches from a remote storage system. What is the most likely bottleneck causing this behavior?

    1. A.The model has too many parameters, leading to GPU memory thrashing and slow training iterations.
    2. B.The learning rate setting is too low, so optimizer steps are too small to drive meaningful loss reduction.
    3. C.Data loading from remote storage is I/O-bound, causing GPUs to sit idle while waiting for the next batch.
    4. D.The batch size is set too large for the GPU memory, causing out-of-memory errors that degrade performance.
    5. E.A training script defect introduces unnecessary synchronization barriers, causing GPUs to wait for each other.
    Show answer & explanation

    Correct answer: C — Data loading from remote storage is I/O-bound, causing GPUs to sit idle while waiting for the next batch.

    • A. Too many parameters would cause GPU memory thrashing, typically resulting in out-of-memory errors or extremely high memory pressure, not the observed fluctuating utilization. The key clue is the remote storage data loader, pointing to I/O rather than model size.
    • B. A low learning rate reduces the step size, slowing loss convergence, but it does not directly cause GPU underutilization. If data were readily available, GPUs would still be fully utilized. The fluctuating utilization is tied to data availability.
    • C. Data loading from remote storage is I/O-bound. When the input pipeline cannot deliver batches fast enough, GPUs sit idle waiting, leading to fluctuating utilization (40-60%) and slowing training progress. This is a common bottleneck in distributed training with remote storage.
    • D. A batch size too large for GPU memory would cause out-of-memory errors or force batch size reduction, not sustained mid-range utilization. The observed symptoms are more consistent with data latency than memory limits.
    • E. Synchronization barriers can reduce parallel efficiency, but the question specifically highlights remote storage data loading, making I/O the likely bottleneck. Barrier issues would typically be diagnosed via distributed training diagnostics.

    Subdomain 4.5: Verify data quality for go/no-go decision to conduct data preparation

    26.Which action best supports verifying data quality for a go/no-go decision to conduct data preparation?

    1. A.Maintain a log of all manual data cleaning steps and decisions for traceability.
    2. B.Store every version of raw, intermediate, and processed datasets with timestamps.
    3. C.Implement the entire data preparation pipeline as version-controlled, parameterized code.
    4. D.Document the data quality findings and preprocessing outcomes in a shared wiki.
    5. E.Use a data versioning tool like DVC to track dataset versions and changes.
    6. F.Automate the execution of the preprocessing pipeline using an orchestrator like Airflow.
    Show answer & explanation

    Correct answer: D — Document the data quality findings and preprocessing outcomes in a shared wiki.

    • A. While maintaining a log of manual cleaning steps ensures traceability, this action occurs after the go/no-go decision is made. It does not directly contribute to verifying data quality before deciding to proceed with data preparation.
    • B. Storing dataset versions with timestamps supports reproducibility but is not specifically aimed at verifying data quality for the go/no-go decision. It is more relevant for tracking changes after preparation begins.
    • C. Implementing the pipeline as version-controlled code ensures consistency and reproducibility, but it does not directly address the initial verification of data quality needed to inform the go/no-go decision.
    • D. Correct. Documenting data quality findings and preprocessing outcomes provides a clear record of the assessment results. This directly supports the go/no-go decision by making the quality status accessible to stakeholders, enabling an informed choice on whether to proceed with data preparation.
    • E. Using a data versioning tool like DVC is excellent for dataset lineage and reproducibility, but it does not directly capture or communicate the data quality findings essential for the go/no-go decision.
    • F. Automating the preprocessing pipeline with an orchestrator improves execution consistency, but it is focused on the execution phase rather than the upfront verification of data quality to support the go/no-go decision.

    Subdomain 4.4: Manage data transformation to conduct data preparation

    27.What are the most effective methods to ensure reproducible and consistent data preparation in AI model development?(Select 2)

    1. A.Mandate Jupyter notebooks with detailed markdown cells describing each transformation, assuming code transfers correctly across environments.
    2. B.Implement a feature store to centralize transformed features and serve them, paired with version-controlled transformation scripts for reproducibility.
    3. C.Copy preprocessing scripts from the data scientist's machine to production and verify each step manually using a checklist.
    4. D.Use Apache Airflow with containerized tasks, pinned dependencies, and parameterized configs to ensure identical execution across environments.
    5. E.Package the model and preprocessing into a monolithic Docker container, so any update triggers a full rebuild and redeployment.
    Show answer & explanation

    Correct answers: B, D — Implement a feature store to centralize transformed features and serve them, paired with version-controlled transformation scripts for reproducibility.; Use Apache Airflow with containerized tasks, pinned dependencies, and parameterized configs to ensure identical execution across environments.

    • A. Incorrect. Jupyter notebooks are not inherently reproducible across environments due to dependency mismatches, hidden state, and execution order issues. Self-documenting code alone does not guarantee consistent data transformation in production.
    • B. Correct. A feature store centralizes transformed features, ensuring consistency across training and serving. Coupled with version-controlled transformation scripts, it provides reproducibility, auditability, and on-demand recreation of features, which is a robust approach for managing data preparation.
    • C. Incorrect. Manually copying scripts and using a checklist introduces human error, drift, and lack of automation. This approach is not scalable or reliable for production environments.
    • D. Correct. Apache Airflow with containerized tasks, pinned dependencies, and parameterized configurations enables repeatable and identical execution of data transformation workflows, ensuring reproducibility across environments.
    • E. Incorrect. While a monolithic Docker container ensures environmental consistency, it reduces flexibility and maintainability. It requires full rebuilds on updates and does not address centralized feature management or transformation reproducibility as effectively as the stronger options.

    Subdomain 4.6: Verify model ready for operationalization go/no-go decision

    28.Which of the following are appropriate items to verify when making a go/no-go decision for operationalizing an AI chatbot model?(Select 4)

    1. A.A detailed report on the model's accuracy for intent classification from testing.
    2. B.Established monitoring dashboards and alerts for model latency and error rates.
    3. C.An assessment of the chatbot's user interface design and user experience.
    4. D.Verified integration with the existing customer support ticketing system.
    5. E.A record of the number of training epochs used during model development.
    6. F.Documented procedures for model versioning, rollback, and incident response.
    7. G.The specific programming language selected for the deployment scripts.
    Show answer & explanation

    Correct answers: A, B, D, F — A detailed report on the model's accuracy for intent classification from testing.; Established monitoring dashboards and alerts for model latency and error rates.; Verified integration with the existing customer support ticketing system.; Documented procedures for model versioning, rollback, and incident response.

    • A. Correct. A detailed accuracy report demonstrates that the model meets performance thresholds, which is a critical input to the go/no-go decision. While not sufficient alone, it is an essential artifact for validating model readiness.
    • B. Correct. Monitoring dashboards and alerts for latency and error rates are key operational-readiness artifacts. They enable proactive issue detection and performance tracking post-deployment, directly supporting the go/no-go decision.
    • C. Incorrect. User interface design and user experience are important for the overall chatbot product but are not primary criteria for the model operationalization go/no-go decision. The decision focuses on model performance, controls, and production support.
    • D. Correct. Verified integration with the existing customer support ticketing system ensures the model can function within the existing ecosystem, which is a prerequisite for operationalization and a key factor in the go/no-go decision.
    • E. Incorrect. The number of training epochs is a development detail and does not directly demonstrate whether the model is ready for production use. Operationalization decisions rely on validated performance and support readiness, not training hyperparameters.
    • F. Correct. Documented procedures for model versioning, rollback, and incident response are essential go/no-go criteria. They show the organization can manage changes and recover from failures safely, which is central to verifying production readiness.
    • G. Incorrect. The programming language used for deployment scripts may affect implementation but is not a key indicator of operational readiness. The go/no-go decision is about whether the solution can be monitored, supported, and controlled in production.

    Domain 5: Operationalize AI Solution

    Subdomain 5.1: Manage creation of AI solution deployment plan

    29.Which of the following is NOT part of a deployment plan for an AI solution?

    1. A.Project charter
    2. B.Deployment checklist
    3. C.Risk register
    4. D.Lessons learned log
    Show answer & explanation

    Correct answer: D — Lessons learned log

    • A. Incorrect. The project charter is a foundational document created at initiation, not part of the deployment plan. It defines project purpose and scope, but is not used to guide deployment steps.
    • B. Incorrect. A deployment checklist is a key component of the deployment plan, ensuring all steps (e.g., testing, rollout, monitoring) are systematically addressed.
    • C. Incorrect. The risk register is critical in the deployment plan as it identifies, assesses, and mitigates risks associated with deploying the AI solution.
    • D. Correct. The lessons learned log is typically created after deployment or during project closure to document insights, not during the planning of deployment.

    Subdomain 5.6: Manage AI solution transition plan

    30.The operations team for an AI-based pricing optimization system observes that the model's prediction accuracy has dropped below the threshold defined in the transition plan's monitoring triggers. What should the team do first?

    1. A.Immediately initiate the model retraining process using the latest data
    2. B.Validate the accuracy deterioration by analyzing data distribution and model logs
    3. C.Escalate the issue to the data science team for investigation and recalibration
    4. D.Roll back the model to the previous stable version to restore accuracy
    5. E.Notify the business stakeholders and request a decision on next steps
    Show answer & explanation

    Correct answer: B — Validate the accuracy deterioration by analyzing data distribution and model logs

    • A. Incorrect. Initiating retraining without validating the root cause may waste resources and fail to address underlying issues like data drift or monitoring errors. The team should first confirm the issue and understand its cause.
    • B. Correct. The first step is to validate that the accuracy drop is real and diagnose potential causes by reviewing data distribution, model logs, and monitoring signals. This aligns with transition plan governance, which expects investigation before corrective action.
    • C. Incorrect. While escalation may be necessary, it should follow initial validation to provide the data science team with actionable insights rather than raw observations.
    • D. Incorrect. Rolling back is a reactive measure that should only be considered after confirming the degradation and assessing options. Immediate rollback could disrupt service unnecessarily if the alert was transient or due to a monitoring anomaly.
    • E. Incorrect. Stakeholders should be informed according to the communication plan, but requesting a decision is not the first technical response. The team should first validate the performance drop and investigate the cause to provide meaningful information.

    Subdomain 5.5: Prepare final report/lessons learned

    31.During the lessons learned session for an AI project, the team expresses conflicting views on the effectiveness of the agile methodology used. What should the project manager do to capture lessons learned effectively?

    1. A.Record the majority opinion and disregard outliers to simplify documentation.
    2. B.Facilitate a discussion to reach consensus and record only agreed points.
    3. C.Document all perspectives with their context and let future teams decide.
    4. D.Escalate the conflict to senior management for resolution before documenting.
    Show answer & explanation

    Correct answer: C — Document all perspectives with their context and let future teams decide.

    • A. Incorrect. Recording only the majority opinion can create bias and omit important insights from minority viewpoints. Lessons learned should capture diverse perspectives to provide a comprehensive view for future projects.
    • B. Incorrect. While facilitating discussion is useful, limiting the record to only agreed points can lose valuable nuance and disagreement that may help future teams understand trade-offs. Lessons learned should reflect all viewpoints, not just consensus.
    • C. Correct. Documenting all perspectives with their context ensures a balanced and thorough lessons learned report. This preserves knowledge, avoids bias, and allows future teams to understand the full range of experiences and make informed decisions.
    • D. Incorrect. Escalation to senior management is not necessary for normal lessons learned conflicts and may delay the process. The project manager should capture the information objectively without resolving every disagreement before recording it.

    Subdomain 5.3: Oversee model governance

    32.Which of the following is the best practice for model governance?

    1. A.Tagging each model artifact with a semantic version and storing all artifacts in a centralized registry with deployment metadata.
    2. B.Using date-based naming for model files and retaining only the last two versions on the production server.
    3. C.Maintaining a Git repository of all training code with branch-based model tracking for version control.
    4. D.Relying on the experiment tracking tool to log performance metrics and experiment configurations for each training run.
    5. E.Keeping all versions live in a multi-armed bandit setup for automatic fallback and traffic splitting.
    Show answer & explanation

    Correct answer: A — Tagging each model artifact with a semantic version and storing all artifacts in a centralized registry with deployment metadata.

    • A. Correct. Semantic versioning with a centralized registry provides full traceability, auditability, and controlled promotion of model artifacts across environments. This ensures you can identify exactly which model version is deployed, where, and with what configuration—core requirements for model governance.
    • B. Incorrect. Date-based file names lack semantic meaning and retaining only the last two versions limits traceability and rollback capabilities. This approach does not provide robust audit trails or compliance support required for governance.
    • C. Incorrect. While a Git repository for training code is useful for code version control, branch-based model tracking alone does not govern model artifacts or deployments. Governance requires explicit artifact versioning, metadata, approval status, and deployment lineage beyond source code tracking.
    • D. Incorrect. Experiment tracking tools log performance metrics and configurations, but they do not inherently manage model artifacts, versions, or deployment records. Governance demands artifact control, approval workflows, lineage, and version management beyond experiment logging.
    • E. Incorrect. A multi-armed bandit is a deployment strategy for testing, not a governance mechanism. It complicates control, compliance, and rollback because governance requires clear approval, version identification, and managed release processes.

    Subdomain 5.2: Manage AI solution deployment

    33.What is the most appropriate approach for managing user access to an AI system in a production environment?

    1. A.Grant all data scientists full administrative access to the production environment so they can quickly troubleshoot model issues.
    2. B.Implement role-based access control (RBAC) with least privilege, providing separate roles for training, deployment, and monitoring.
    3. C.Use a single service account for all system interactions to simplify credential management and ensure audit trails are centralized.
    4. D.Allow users to self-provision access to the AI system after completing a brief online training module to streamline onboarding.
    Show answer & explanation

    Correct answer: B — Implement role-based access control (RBAC) with least privilege, providing separate roles for training, deployment, and monitoring.

    • A. Incorrect. Granting full administrative access to all data scientists violates the principle of least privilege and increases the risk of accidental or malicious changes to production systems. Access should be tightly controlled and limited to only what is necessary for each role.
    • B. Correct. Implementing role-based access control (RBAC) with least privilege ensures that users have only the permissions necessary for their roles. Separate roles for training, deployment, and monitoring enforce segregation of duties, reduce security risks, and improve governance and accountability.
    • C. Incorrect. Using a single service account for all system interactions eliminates granularity in access control, making it difficult to track actions and increasing the risk of widespread compromise if the account is breached. It also undermines separation of duties and accountability.
    • D. Incorrect. Allowing users to self-provision access after a brief online training module weakens governance and can lead to unauthorized or inappropriate access. Access should be approved and assigned according to business need, role, and security policy rather than convenience.

    Subdomain 5.7: Oversee AI solution contingency plan

    34.A financial services firm's AI-powered credit scoring system becomes unavailable due to a critical infrastructure failure. The business continuity plan dictates shifting to a manual underwriting process. Which of the following must be ensured to maintain operational resilience during this transition?

    1. A.The manual process should rely on simplified criteria that mimic the AI model's logic.
    2. B.Underwriters must be trained on the AI model's decision boundaries for consistency.
    3. C.A fallback system using a rule-based engine should be tested and ready for deployment.
    4. D.All loan applications should be paused until the AI system is restored to avoid errors.
    Show answer & explanation

    Correct answer: C — A fallback system using a rule-based engine should be tested and ready for deployment.

    • A. Incorrect. Simplified criteria may assist but do not guarantee consistency or accuracy; the manual process should align with defined business rules, not necessarily mimic the AI model's logic.
    • B. Incorrect. Training on AI decision boundaries is not practical or necessary for manual underwriting; operational resilience depends on a validated alternative process, not on replicating model internals.
    • C. Correct. A tested rule-based fallback system provides a reliable immediate alternative when the primary AI fails, supporting continuity and reducing disruption, even if manual underwriting is part of the plan.
    • D. Incorrect. Pausing applications halts business operations and undermines resilience; continuity planning aims to keep services running through an alternate approved process.

    Subdomain 5.4: Oversee AI solution metrics (e.g., KPI, model performance)

    35.An AI manager needs to create a dashboard for both technical and business stakeholders to monitor an AI-powered inventory optimization system. How should the dashboard be structured to effectively serve both audiences?

    1. A.Provide a unified dashboard that lists all collected metrics without categorization, relying on stakeholders to apply filters to find relevant information.
    2. B.Create separate dashboard sections: a business tab (cost savings, inventory turnover) and a technical tab (model accuracy, latency).
    3. C.Focus exclusively on business KPIs like cost reduction and order fulfillment rates, omitting technical metrics as they are not relevant to business decision-makers.
    4. D.Avoid real-time dashboards; send automated weekly PDF reports with all metrics to stakeholders, reducing the need for interactive exploration.
    Show answer & explanation

    Correct answer: B — Create separate dashboard sections: a business tab (cost savings, inventory turnover) and a technical tab (model accuracy, latency).

    • A. Incorrect. A unified dashboard without categorization forces stakeholders to manually filter irrelevant data, reducing usability for both technical and business users. Role-based organization is important for clarity.
    • B. Correct. Separating the dashboard into business and technical sections allows each audience to quickly access relevant metrics. Business stakeholders track outcomes like cost savings and inventory turnover, while technical stakeholders monitor model accuracy and latency. This aligns with best practices for stakeholder-specific reporting.
    • C. Incorrect. Business KPIs are important, but omitting technical metrics leaves the team without visibility into model health and operational issues. Effective monitoring requires both business outcomes and technical performance measures.
    • D. Incorrect. Automated weekly PDF reports lack interactivity and timeliness, which are critical for timely decision-making in AI-driven systems. Dashboards should support near-real-time insights and interactive exploration.

    Want the full experience?

    These are just samples. Practice the full PMI Certified Professional in Managing AI (PMI-CPMAI) question bank in quiz mode — free, no signup, with domain practice and exam simulation.

    Practice all 335 questions in quiz mode