In AI fraud detection for financial institutions, PMI-CPMAI-aligned practices place strong emphasis on data quality, completeness, and relevance as the foundation of model reliability and regulatory compliance.
Because the team has access to various internal and external data sources, the appropriate method is to perform a comprehensive data audit and cleansing process.
A data audit systematically examines each source for accuracy, consistency, timeliness, coverage of key fraud patterns, and alignment with business and regulatory needs. It checks for missing values, duplicates, inconsistencies across systems, and potential bias (e.g., underrepresentation of certain customer segments or regions). Cleansing then addresses identified issues through deduplication, normalization, imputations where appropriate, and removal of unusable or misleading records. This process ensures that the data used to train and operate the AI solution truly reflects real-world transactions and fraud behaviors, supporting trustworthy and explainable outcomes.
Limiting data to internal sources only (option B) may unnecessarily reduce coverage and predictive power, especially when reputable external data (e.g., watchlists, consortium data) can enhance detection. Integrating data "as is" (option C) violates good AI governance and greatly increases the risk of poor model performance and regulatory concerns. Using pretrained models without tailoring (option D) ignores the need for alignment with the institution's own data and fraud patterns. Therefore, the method that directly addresses the objectives is conducting a comprehensive data audit and cleansing process.

PMI-CPMAI aligns data preparation with executing data cleansing and enhancement activities so that datasets meet model and operational requirements. Inconsistent clinical coding is a data quality issue that threatens accuracy, fairness, and interpretability, because identical conditions may be represented differently across sources. The PMI-aligned response is to determine and apply the necessary transformation steps- standardizing codes to a controlled vocabulary, mapping local codes to a canonical schema, normalizing formats, and documenting rules and lineage so the process is auditable. Ignoring inconsistencies (B) increases noise and can embed systematic bias (e.g., certain clinics appearing "higher risk" due to coding artifacts).
Relying only on synthetic data (C) can reduce fidelity if the synthetic process fails to reflect true clinical distributions. Skipping validation (D) violates responsible delivery expectations because it undermines patient safety and data integrity. PMI's responsible and trustworthy framing supports disciplined data readiness work before model development proceeds.