Changes to production AI models—including retraining, parameter updates, and architecture modifications—can alter model behavior in ways that introduce new biases, reduce accuracy, or create regulatory compliance issues. Validation before deploying changes is the most critical safeguard.

Why C is Correct: According to ISACA AAIR change management guidance for AI systems, rigorous validation to assess changes' effects on predictive accuracy and model bias is the most important change management activity. Production AI models make real-world decisions affecting people and business outcomes. Unvalidated changes may degrade performance, introduce discriminatory patterns, or create regulatory violations that are difficult to detect and remediate after deployment.

Why A is Wrong: Allowing operational teams to adjust configuration parameters in real time bypasses change control processes and creates untracked, unvalidated changes to model behavior. This represents a governance risk, not an acceptable change management practice.

Why B is Wrong: Access controls for new model functionalities are a security and authorization concern. While important for access governance, they do not address the technical risk that model changes may degrade performance or introduce bias.

Why D is Wrong: Expediting production rollouts to minimize downtime prioritizes availability over quality assurance. Rushing changes without adequate validation trades one operational risk (downtime) for a potentially more severe risk (biased or inaccurate outputs affecting critical decisions).