Most AI rollouts stall for the same reason: teams debate whether the model is good enough, when the real question is whether the people using it are ready. I started working on this problem as the sole UX researcher in MLIO, JPMorgan Chase's machine learning and AI product space, where I led research on the agent-assist pilot that became EVEE. Usage was low, and nobody could say whether the tool was the problem or the rollout was.

What came out of that work was a simple rule. Before an AI capability moves up a stage, measure three things at the stage it's on: do people know it's there, do they use it, and do they trust it the right amount? Each one gates the next.

I've since extended that rule into the five-stage model below. Each stage has explicit evidence gates, a measurement set, and governance requirements, grounded in human-automation research. The thresholds are working heuristics, calibrated per product and risk level.