AI capabilities are accelerating, yet many organizations face a dilemma: they need intelligent systems, but what do you do when direct access to user data is restricted by regulation, contractual commitments, or internal principles? This session presents approaches for developing and deploying AI when raw customer content must remain inaccessible.

We’ll walk through a complete lifecycle: designing privacy constraints at the start; creating training and evaluation loops that rely on internal data, redacted samples, and synthetic augmentation; spotting distribution gaps without leaking sensitive information; and improving models when debugging cannot involve raw inputs. The talk also explores the role of cloud secure enclaves and on-device inference, outlining how each option supports different risk profiles and performance goals.