Say I complete a task 30% faster but what was the annoyance-cost of the model constantly getting it wrong and having to keep correcting it
That did the companys revenue / profit significantly change in some way?
Even at a large company if a person 100xd their productivity that would meaningfully move the needle.
You've added a step.
And make no mistake, it's entirely possible at some point they are storing outputs for being trained on to replace your Org's business too. Selling the inference agent is what's paying the bills in theory right now, that may change, and the next logical step is cannibalizing customers.