Responsible Ai In The Enterprise Heather Dawe Pdf May 2026

| Stage | Responsible AI Action | |--------|----------------------| | Problem framing | Does using postal code as a feature proxy for socioeconomic status create redlining risk? Impact assessment conducted. | | Data collection | Test for proxy discrimination (e.g., zip code correlated with race). Apply reweighting to reduce bias. | | Model training | Adversarial debiasing to ensure protected attributes do not drive predictions. | | Pre-deployment | Model card documents disparate impact ratio. Human approval required for any premium differential >10%. | | Post-deployment | Monthly fairness audit. Drift detection on population changes. | | Incident response | If bias emerges, model is automatically switched to fallback rule-based system. |