Every IT automation vendor pitch this year seems to come with a headline number — 60% deflection, 66% deflection, one benchmark report claiming AI automation resolves tickets sixteen times faster than manual handling. Some of those numbers are real. The problem is that "deflection rate" isn't a standardized metric, and different vendors measure genuinely different things while calling it the same name.
A useful distinction, laid out clearly in a recent analysis from Serval: deflection often just means the request never reached a human queue — which can include a user getting routed to an FAQ page or a chatbot that didn't actually solve anything, and simply giving up. Automation rate is a stricter number: the request was fully completed with zero human involvement. The gap between those two figures, at the same company, can be enormous, and it's rarely the one featured on the sales slide.
There's also a real ceiling on what's automatable in the first place, and it isn't evenly distributed. Password resets, access requests, and basic policy questions deflect at high rates almost immediately. Complex troubleshooting — the tickets that actually eat the most staff time — deflects far more slowly, if at all, because it usually depends on institutional knowledge that was never written down anywhere the AI can read it.
If you're evaluating one of these tools, the question worth asking isn't "what's your deflection rate." It's "what percentage of tickets in each category were completed with zero human touch, broken out by category, on our own ticket history rather than your best customer's." Most reputable vendors can run that analysis before you sign anything. If they can't or won't, that's the actual answer to whether the number will hold up once it's your data instead of theirs.