The acceptance rate of an AI suggestion is a UX metric before it's a model metric.

The acceptance rate of an AI suggestion is a UX metric before it's a model metric. Users reject correct suggestions offered at the wrong moment far more often than they accept wrong ones offered well. Most AI features fail at completeness, not accuracy — surface everything the model found and you've handed someone a cleanup task instead of an answer.

That distinction decides where the money goes. A single acceptance number routes the work to the model by default, because model quality is the half that's legible and fundable, and you can spend a year improving something that was never the constraint.

I now make teams say which of the two they're solving before the work gets scoped. Not because the instrumentation is difficult, but because the metric quietly makes a staffing decision that nobody remembers making.