A power company's AI tool approves commercial energy-efficiency financing applications and was validated as fair across business sizes at launch. Eighteen months later, an internal audit finds approval rates diverging by business size. In the same period, new sales representatives have been sourcing leads from a narrower set of industries than before. What best describes the situation, and where should the committee focus?

Framing this as bias drift is correct: the model passed fair validation at launch, and the divergence emerged afterward alongside a documented change in the applicant population, pointing to a shift in inputs rather than a flaw baked in from the start, and the appropriate fix is stronger ongoing monitoring. Blaming the original training data jumps to a conclusion despite evidence that the model performed fairly at launch, so replacing that data would not address the actual, more recent cause. Calling it a labeling error mislabels this as a development-stage issue, but labeling problems would typically have shown up in the original validation, not eighteen months later alongside a change in lead sourcing. Concluding no changes are needed ignores the audit's findings entirely and treats a one-time launch certification as permanent, which is precisely the assumption that let the divergence go unnoticed. The general principle: when a system was previously validated as fair, a new disparity is a signal to investigate what changed in the environment or inputs, and to strengthen ongoing monitoring, rather than to assume the original build was defective.