I time the old ritual with a phone stopwatch and a sample of 20. I time the new ritual the same way. I do not extrapolate from my laptop.
If label type-in drops from 70 seconds to 12, that is the slide. If visual lookup removes two radio calls an hour, that is the slide. Model accuracy is a footnote.
Include the failure tax
A tool that is fast when it works and catastrophic when it does not can lose. I count retries and escalations in the same study.
Re-measure after the novelty dies
Week one is theater. Week five is truth. I put a reminder to measure again after the champions stop hovering.
