For an AI-based fraud detection system, PMI-CPMAI-aligned guidance on benefits realization and performance management stresses that success metrics must be quantified against a clear baseline and monitored continuously over time. To properly define and measure KPIs such as detection accuracy, reduced investigation time, and cost savings, the project team should first establish a baseline using historical data comparisons (D). That means analyzing historical fraud cases, prior detection rates, average investigation duration, and historical financial losses to understand "pre-AI" performance. This provides a reference point against which improvements can be measured in a verifiable way.
In addition, PMI-CPMAI emphasizes continuous performance monitoring (B) as part of AI lifecycle governance. Fraud patterns, transaction volumes, and user behavior evolve, so model performance relative to KPIs must be tracked on an ongoing basis using dashboards and periodic evaluations. This supports early detection of performance degradation, allows recalibration of thresholds, and validates that business benefits (e.g., decreased losses, reduced workload) are being sustained.
Relying only on qualitative feedback, random benchmarks, or purely theoretical targets does not meet PMI-CPMAI expectations for evidence-based measurement and governance. Therefore, the two appropriate approaches are: implementing a continuous performance monitoring system (B) and establishing a baseline using historical data comparisons (D).

In CPMAI and PMI-aligned AI lifecycles, integrating diverse data types from multiple operational systems is typically handled through robust data engineering pipelines, most commonly implemented as ETL (extract, transform, load) or closely related ELT patterns. For a logistics optimization use case, the AI system needs to bring together traffic patterns (often from external or sensor feeds), internal delivery schedules, and vehicle performance/telematics data into a consistent, analyzable structure.
An ETL pipeline is designed precisely for this: it extracts data from heterogeneous sources, transforms it into common formats and schemas (handling units, timestamps, geocodes, data quality rules), and loads it into a target store (data lake, warehouse, or feature store) that downstream AI components can consume. CPMAI emphasizes that this integration work is a core part of the Data Understanding and Data Preparation phases, because AI models depend on unified, high-quality inputs rather than fragmented, siloed feeds. While real- time frameworks, federated models, or warehouses may play additional roles, the primary method explicitly focused on integrating diverse data sources into a coherent whole is an ETL pipeline, making option B the best fit.

PMI-CPMAI describes a very practical early-stage activity: breaking down a solution into components or sub- functions and then deciding which components actually require AI and which do not. This is often referred to as a components-based analysis. The idea is to decompose the overall workflow or product into units such as data ingestion, preprocessing, prediction, rule-based decisioning, user interface, reporting, and integration layers.
For each component, the team asks:
Does this require cognitive capability (learning from data, pattern recognition, probabilistic reasoning)?
Or can it be handled by conventional software, rules, or existing systems?At the same time, they identify technology and data requirements: data sources, data quality, storage, pipelines, compute needs, and integration points for each AI-relevant component. PMI-CPMAI ties this directly into later tasks such as technical feasibility, architecture design, and MLOps planning.
Detailed data mapping (option A) is useful but focuses mainly on information flows, not necessarily on AI vs non-AI partitioning. Technical feasibility assessment (option B) evaluates whether a proposed AI approach is realistic but presumes that the AI portions are already identified. Only components-based analysis (option C) simultaneously answers "which parts need AI, which do not, and what are the tech/data needs for each?", which matches the scenario precisely.

PMI guidance for responsible and trustworthy AI stresses that ethical performance is not a one-time checkbox; it requires ongoing oversight, including transparency, accountability, and continuous controls. PMI- CPMAI's exam outline explicitly highlights maintaining audit trails for algorithmic decision-making, implementing compliance monitoring mechanisms, and managing accountability documentation- foundational practices that align directly with continuous monitoring and auditing. In high-stakes healthcare use cases like readmission prediction, model drift, data drift, and shifting patient populations can degrade performance and fairness over time, which can create patient safety risks. Continuous monitoring enables the team to detect deteriorating accuracy, emerging bias, and unexpected failure modes early; auditing supports traceability of decisions, data lineage, and adherence to governance requirements. PMI also emphasizes that ethical AI demands validation and transparency, noting that accountability and continuous monitoring are crucial to maintain ethical standards and minimize undesirable outcomes. Encryption (A) protects confidentiality, and explainability (B) supports transparency, but neither alone ensures sustained ethical compliance. Stakeholder impact analysis (D) is valuable during assessment, yet monitoring/auditing is the most direct operational method to ensure ethics remain intact after deployment.

PMI-CPMAI guidance on evaluating operational AI systems, especially in risk-sensitive domains like fraud detection, stresses that project managers must link model performance to business KPIs using multiple complementary evaluation methods, not a single metric. The material explains that fraud models have asymmetric costs (false positives vs. false negatives), evolving fraud patterns, and complex business impacts, so "no single measure is sufficient to characterize business value or risk." Instead, teams are encouraged to use a diverse set of validation techniques, such as holdout and cross-validation, backtesting on historical periods, confusion matrices, cost/benefit-weighted metrics, and A/B or champion-challenger tests in production-like environments.
PMI-CPMAI also notes that evaluation should combine technical metrics (precision, recall, ROC/AUC, F1, lift) with business-oriented indicators (fraud losses avoided, investigation workload, customer friction, and regulatory or compliance thresholds). Using multiple techniques allows the project manager to check consistency across views and avoid being misled by a single "good-looking" number that hides harmful side effects. Relying on quarterly financial reports or external experts alone does not provide the granular, model- specific insight required, and a single comprehensive metric contradicts PMI's emphasis on multidimensional evaluation. Therefore, to ensure an accurate and reliable assessment of the AI fraud system against business KPIs, the most effective method is utilizing a diverse set of validation techniques.