I think the better framing might be "directly observable impact on performance"?
If Alice in sales lands a client by showing off a shiny new dashboard which Bob the programmer built, obviously they both had to perform. But it's fairly easy to run the counterfactual Alice: if Carol didn't land a similar client with the same dashboard, Alice is adding more value. (At least over large sample sizes.) It's very hard to run the counterfactual for Bob: would a worse dashboard have lost Alice's client? Would a better one have landed Carol's? Would another programmer have built something better or worse, in more time or less? Since you only want one of that dashboard, those things are exceedingly hard to answer.
Notably, some companies with consulting/deployed engineers do pay bonuses, because like a physical craftsman they're building many client-specific products. If one person's dashboards consistently land contract extensions than another's they're probably performing better.
(Of course, nothing is ever this clear or easy. If Alice just gets better clients, Glengarry Glen Ross style, Alice didn't actually sell better. If Alice promises a dashboard that doesn't exist yet, leaving Bob too busy to fulfill Carol's client's requests, that's not better either. And at Wall Street is infamously asymmetric, rewarding risk taking rather than actual performance.)