When an AI model underperforms against defined KPIs (70% accuracy vs required 89%), PMI-style AI evaluation guidance directs project managers to first investigate data-related issues, especially representativeness and quality of the training data, before focusing on algorithms or infrastructure. If the training data is not representative of real-world data (option A), the model may learn patterns that do not generalize to production conditions. For example, it might be overexposed to common, simple cases and underexposed to rare but critical scenarios, specific customer segments, geographies, or newer product types.

This mismatch is one of the most common causes of accuracy degradation between expected and actual performance. Ensuring representativeness involves checking that the data covers the full spectrum of operational scenarios, class distributions, time periods, and user demographics relevant to the use case. Inadequate compute (option B) more often affects training time than final accuracy, assuming the model trains to convergence. Failure to split datasets correctly (option C) leads to unreliable evaluation metrics, but the question already states an accuracy result and a KPI gap, pointing to performance, not just measurement. Algorithm selection (option D) is important but typically evaluated after confirming that the data foundation is sound. Thus, the first issue to investigate is whether training data is representative of real-world data.

The best answer is C. Utilize a diverse set of test cases . PMI-CPMAI’s model evaluation domain focuses on building comprehensive evaluation plans and formulating appropriate evaluation questions and criteria. Validation is not treated as a one-time technical check, but as a structured process designed to test model behavior across a range of relevant conditions, edge cases, and data contexts. Using a diverse set of test cases is the best way to assess whether predictions are accurate, robust, and dependable enough for a public-sector policy setting.

Option A is useful for software quality but does not validate predictive performance. Option B is weaker because a single validation exercise can miss important failure modes, bias, or context-specific weaknesses. Option D supports engineering discipline, but continuous integration testing focuses more on code and deployment workflow than on validating model prediction quality itself. PMI’s CPMAI framework emphasizes comprehensive evaluation design, iteration, and addressing performance issues such as drift and changing conditions. That makes broad and varied test coverage the most PMI-aligned approach to model validation. In practical terms, diverse test cases provide stronger evidence that the model will generalize beyond a narrow sample and support trustworthy decision-making.

In the PMI-CPMAI perspective on responsible AI and data governance, regulatory compliance starts with knowing exactly what data you have and how sensitive it is. Before you can design controls, encryption schemes, or risk plans, you must first perform a data audit and classification to identify personal, sensitive, and regulated data elements, as well as their sources, flows, and storage locations. This aligns with the guidance that early in the AI lifecycle, project teams should create a clear data inventory and mapping to understand which datasets fall under privacy regulations (such as health, financial, or personally identifiable information).

By conducting a thorough data audit to identify sensitive information, the project team can determine which regulations apply, what consent or legal basis is required, and where to apply specific safeguards (access controls, anonymization, retention limits, etc.). Encryption and broader risk management plans are important, but they are secondary steps that rely on the foundational insight gained from the audit. Verbal commitments from stakeholders have no formal regulatory standing. Therefore, in the initial stages of data collection and aggregation, the task that most directly supports regulatory compliance is a thorough data audit to identify sensitive information.

The best answer is D. Potential non-cognitive alternatives . In PMI-CPMAI, the early business assessment is not just about deciding whether AI can be used, but whether AI should be used at all for the problem. Under Identify Business Needs and Solutions , PMI’s official exam content outline explicitly states that initial AI feasibility includes comparing AI approaches against traditional solution alternatives . That means the project manager should first determine whether a simpler, rules-based, workflow, reporting, or conventional software solution could solve the problem without introducing unnecessary AI complexity, risk, cost, or governance burden.

This also aligns with your uploaded CPMAI-aligned playbook, which emphasizes that teams should avoid applying AI automatically and should choose governance and solution rigor proportionate to the actual need and risk. The playbook repeatedly stresses that the right decision starts with the business problem and whether AI is truly the appropriate approach, rather than assuming an AI solution by default.

Why the others are weaker: business objectives matter, cost matters, and interpretability may matter later, but the key question for deciding whether a cognitive approach is appropriate is whether viable non-cognitive alternatives already exist. That is the clearest PMI-CPMAI-aligned choice.

In PMI’s treatment of AI in customer-facing environments, responsible AI, privacy, and regulatory compliance are consistently framed as high-impact risk areas. For a telecommunications company using AI chatbots for customer service, any breach of customer data privacy is not just a technical issue but a legal, regulatory, and reputational threat. It may trigger regulatory investigations, fines, lawsuits, and loss of customer trust.

While scalability risks (such as the chatbot not handling volume) and integration risks (such as poor connection with existing platforms) may harm service quality, they are usually remediable through technical improvements, capacity upgrades, or refactoring. Conversely, PMI’s AI governance perspective emphasizes that violations of data protection laws can incur “non-recoverable” damage: sanctions, forced shutdown of systems, and long-term brand erosion. Therefore, the potential that “the solution might breach customer data privacy regulations, leading to legal consequences” is typically assessed as a higher-order risk than operational challenges.

PMI-CPMAI content stresses implementing privacy-by-design, strict access controls, encryption, and compliance checks early in the solution lifecycle. This means that, in a feasibility and risk assessment, data privacy and regulatory compliance represent the highest risk category, and thus option D is the most appropriate answer.

PMI’s AI management guidance explains that models deployed in dynamic domains—such as healthcare—are particularly vulnerable to data drift, where “the statistical properties of input data or underlying real-world processes change over time, leading to performance degradation if models are not monitored and updated.” In the scenario, the cognitive diagnostic system is continuously exposed to new patient data and evolving medical knowledge from multiple databases. PMI notes that in such cases, “AI models that are not periodically retrained, recalibrated, or revalidated against current data will show reduced accuracy, reliability, and clinical usefulness over time.”

The final report states that the model’s performance degraded over time, affecting reliability and effectiveness, which is the hallmark symptom of data drift rather than an initial validation issue. PMI-CPMAI content stresses setting up continuous monitoring, performance dashboards, and drift detection mechanisms specifically to track “the impact of data drift on model accuracy and business or clinical outcomes,” triggering model refresh or redesign when thresholds are exceeded. Changes in business model requirements could affect alignment of outputs to objectives but would not, by themselves, explain gradual technical degradation in predictions. Therefore, the most appropriate cause, as framed in PMI’s lifecycle and MLOps perspective, is the impact of data drift on model accuracy, requiring ongoing monitoring and retraining to restore performance.

PMI-CPMAI explicitly includes conducting AI go/no-go assessments as a gated decision mechanism to determine whether conditions are sufficient to proceed. In CPMAI-aligned practice, stakeholder alignment on objectives is necessary but not sufficient; readiness must also cover data availability, permissions, privacy/legal constraints, and the feasibility of meeting acceptable performance metrics. A go/no-go assessment brings these prerequisites into a structured review, allowing the project manager to document assumptions, identify critical gaps (e.g., data rights, retention limits, PII handling), and decide whether to proceed, pivot, or stop before incurring avoidable cost and rework. Starting model development prematurely (A) can create downstream rework if data access or compliance fails. Jumping to deployment planning (C) is even more premature when foundational data and legal feasibility are unknown. Buying compute (D) addresses capacity, not feasibility. The PMI-aligned action that enables responsible forward movement is the formal go/no-go gate using readiness criteria.