An ops screen started offering recommended changes. Most of them were sensible. The people on shift stopped taking them after one bad afternoon, because the history showed the model's suggestion and not their judgment. When it went wrong, the record made them look like they had merely clicked. When it went right, the model got the credit. They went back to changing things the old way, where at least the decision was visibly theirs.
The product had erased the owner in order to look automated.
show the suggestion and the choice
A recommendation should sit beside the control, not instead of it. What it wants to change. What it looked at. A clear accept, a clear decline, and a line for why, if why matters later. The audit log should name the person. The model is not an author of record.
If accepting is easier to defend than declining, people will accept things they do not believe. That is how a cautious team gets talked into a bad change.
Automation that hides the human decision also hides the human accountability.
let them override without a speech
Overrides are not defiance. They are the product working. Make them ordinary. Review the pattern of overrides the way you would review a defect: if everyone declines the same suggestion, the suggestion is wrong, or the screen has not earned trust.
Keep the person on the hook. Then give them a screen that tells the truth about that.