The best answer is A. Using synthetic data generation . PMI's CPMAI exam outline explicitly includes supervising data augmentation and synthetic data generation as part of managing AI data preparation, and it also highlights the need to address bias, validate data preprocessing results, and ensure the data is suitable before and during model development. When predictions are biased toward a particular class, that usually points to an imbalance or under-representation problem in the training data. Synthetic data generation is an effective mitigation technique because it can increase representation for the weaker class and improve model learning across the full population.
Option B, stratified sampling, is useful for preserving class proportions in train-test splits and for evaluation discipline, but it does not directly correct a class imbalance problem as effectively as targeted synthetic augmentation. Option C affects optimization efficiency, not fairness or class representation. Option D may tune performance, but hyperparameter changes do not address the root issue if the data itself is skewed. PMI's materials also note that trustworthy AI requires active management of bias, risk, and compliance gaps , which supports selecting a data-centric mitigation approach rather than relying only on model tuning.

PMI-CPMAI emphasizes that AI projects typically involve uncertainty, experimentation, and evolving requirements. Data can change, model behavior must be tuned, and stakeholders may refine success criteria as they see early results. Because of this, PMI frames AI work as well-suited to adaptive/agile approaches that support short iterations, continuous learning, and rapid feedback loops.
In an adaptive/agile approach, the team plans in smaller increments, regularly reprioritizes the backlog, and refines scope based on empirical evidence from model experiments and pilots. This allows them to update features, retrain models, and adjust data or architecture as new insights are gained. PMI-CPMAI links this directly to AI lifecycles, where experimentation, evaluation, and deployment are repeated cycles rather than one-off phases.
Predictive approaches are more rigid and assume stable, knowable requirements upfront, which is rarely realistic for AI behavior and data-driven insights. Incremental and hybrid can add some flexibility, but adaptive/agile is the explicit choice in PMI's guidance when iterative improvement and changing requirements are primary concerns. Therefore, the most effective approach for an AI solution deployment in this context is adaptive/agile.

PMI describes Predictive analytics & decision support as the AI pattern/capability that uses data-driven learning to anticipate outcomes and inform decisions, including "optimizing resource allocation." Route optimization for pizza delivery is fundamentally a decision-support problem: the organization is using historical and real-time signals (orders, traffic, distance, time windows) to recommend an improved routing plan that minimizes time, cost, or late deliveries. PMI also notes that dynamic route optimization is a common example of "goal-driven systems," often associated with reinforcement learning. However, since "goal-driven systems" is not one of the available answer choices, the closest PMI-aligned option among those provided is Predictive analytics, because it directly supports operational decisions under uncertainty and can continuously improve recommendations as more data becomes available. In CPMAI terms, the project manager should ensure the chosen capability matches the business need (faster deliveries, fewer miles, improved SLA performance) and define measurable success criteria for route recommendations and on-time delivery performance.

In PMI-CP/CPMAI-aligned practice, compliance requirements such as data protection regulations (e.g., privacy laws, data residency) and explainability standards are treated as business and regulatory constraints, not as late technical details. They must therefore be identified and incorporated during the business understanding phase. At this stage, the project manager and stakeholders clarify the problem statement, success criteria, risk appetite, and constraints under which the AI solution must operate. That includes explicitly stating: which regulations apply, what level of transparency or explainability is required, which stakeholders must be able to understand model outputs, and which decisions must remain under human control.
By capturing these requirements early, they directly influence the choice of AI pattern, model families, data sources, architecture, and governance mechanisms. If these constraints are postponed until data preparation or final testing, the team risks discovering that the chosen models are too opaque, the data cannot legally be used as collected, or additional documentation and controls are needed that fundamentally change scope and timeline. CPMAI stresses that responsible AI and regulatory compliance are "built in from the beginning," so the correct point to approach these requirements is the business understanding phase.

For an AI-based customer support chatbot, PMI-CPMAI-aligned lifecycle guidance stresses that defining required data starts from real, historical interactions that reflect actual customer needs and behaviors.
Gathering historical customer interaction logs for training data (option B) is the method that best meets this requirement. These logs typically include customer questions, intents, issues, resolutions, and escalation paths, providing a rich, labeled or label-ready corpus that is highly representative of real-world use.
By analyzing these logs, the team can identify the most frequent intents, common phrasing, edge cases, and areas where customers are confused or dissatisfied. This directly informs data schema design, labeling strategies, and coverage requirements for the chatbot. It also helps define performance metrics (such as resolution rate for top intents) and guardrails. Synthetic data (option A) may supplement coverage but should not be the primary basis for defining required data, as it risks encoding designer assumptions instead of reality. Feedback from beta customers (option C) is valuable later in the evaluation and improvement phases.
Developing scripts based on anticipated queries (option D) aids dialogue design but does not truly define the underlying data required for robust training. Therefore, gathering and leveraging historical customer interaction logs is the most appropriate method to define required data for an effective support chatbot.