Data Governance is Key to AI Trust, Not Just Model Performance

The Challenge of AI Trust

During a panel discussion on AI trust, transparency, and accountability, an operator emphasized that building trustworthy AI systems requires more than just evaluating investment opportunities or model viability. The core issue often lies within the data itself and how it is defined and governed.

The "Monthly Active User" Dilemma

A key example involved the definition of "monthly active users" (MAU) at a financial institution. While customers incur system-generated SMS charges and receive interest payments, these transactions do not reflect active engagement. However, if MAU is defined as "any customer with at least one transaction," these dormant customers are counted as active.

This discrepancy leads to different interpretations across departments: marketing reports high MAU, product sees low engagement due to depressed average transaction values, and finance focuses on revenue-generating activities, neither of which includes SMS charges or interest payments. Three teams, using the same underlying data, arrive at three different conclusions.

Impact on AI Systems

The problem extends to AI: if models for churn prediction, credit scoring, or personalization are built on this inconsistently defined data, the AI inherits the confusion. An AI model might treat a dormant customer with an SMS debit as if they were transacting daily. This is not a model problem, as the model performs as instructed; it is a governance problem stemming from a lack of agreement on fundamental data definitions, such as what "active" truly means.