Last March, our team sat in a conference room staring at a chart none of us wanted to see. Meridian's smart suggestion feature — the one we'd spent four months building — had a 12% acceptance rate. Users were clicking the suggestion, seeing what the algorithm proposed, and then manually dragging events to different slots anyway. We weren't saving them time. We were adding a step.
The quiet cost of probabilistic scheduling
Every calendar app that offers “smart” scheduling runs into the same wall: context. A Tuesday 2 PM slot looks identical to an algorithm whether you're deep in a design sprint or half-reading messages. The system sees open time. You know that 2 PM is when your focus starts to dip. No model we tested could reliably distinguish between “available” and “actually good for deep work.”
We spent three weeks A/B testing different suggestion models against a manual baseline. The best one still had a 19% acceptance rate. The worst felt almost random.
So we made a decision that felt radical at the time: we deleted the entire feature. In its place, we built a keyboard-first command flow. Press ⌘K, pick a slot with arrow keys, and move on with zero guesswork.