Fuselab: an AI test won on clicks and lost on task completion

Fuselab's case studies make a simple point with unusually specific evidence: AI optimizes exactly the metric you give it, not the outcome you actually want. Their Fiserv example is the clearest, an AI test lifted click-through on a primary action by 22% while task completion dropped 8%, because the winning button placement hid a confirmation step. Elsewhere Fuselab reports a design-system audit shrinking from two days to 20 minutes plus an hour of review. The lesson isn't that AI fails, it's that teams stop specifying what actually matters once a dashboard shows a number improving.