The model proposes. The human decides.

AI dashboards keep failing in one of two directions. Either the system auto-applies its recommendations — and loses the room the first time it misfires — or it dumps model scores into a table that operators learn to scroll past. Both failures share a root cause: neither treats the recommendation as what it actually is. A decision, waiting for an owner.

Make the recommendation an object

The pattern that works is promoting each suggestion to a first-class object with four mandatory fields: what the model wants to do, why it thinks so, what impact it expects, and how confident it is. Then give the human two equal-weight actions — accept and veto. Not "dismiss", not a tiny ✕ in the corner. A veto button with the same visual dignity as accept, because saying no to a model is a legitimate professional judgment, not an edge case.

Veto rate is a product metric

Once vetoes are explicit, something interesting happens: you can count them. Veto rate per recommendation type becomes the honest health metric for your AI feature — far more honest than usage. A rising veto rate on pricing suggestions tells the data science team exactly where the model is losing the room, weeks before anyone writes an angry ticket.

Keep cause and effect visible

Accepted recommendations should not vanish into a log. In the dashboards I design, accepted items animate into the trend line they affect — the operator literally watches their decision join the curve. When outcomes stay attached to decisions, operators build a calibrated feel for when to trust the model, which is the entire point.

Human-in-the-loop is not friction to be optimized away. It is the mechanism that makes the loop trustworthy.