If you've sat through an AI service-desk vendor pitch recently, you've heard a deflection number somewhere in the 50–75% range. Aisera has cited 65% at Cisco. Moveworks has claimed a 75% average across customers. Those numbers are real — under the definition each vendor is using.
Under the stricter, Gartner-comparable definition — full AI resolution, zero human escalation, and no re-open within 72 hours — the honest 2026 industry average is 20–30% deflection. Best-in-class, mature deployments reach 40–60%, and the realistic ceiling after 18–24 months of sustained optimization sits around 55–65%. First-year deployments typically land at 20–35%.
Reconciling vendor-published numbers against the strict standard generally means applying something like a 10-point compression — a claimed 65% converts to roughly 55% once you control for looser definitions of what counts as “resolved.”
The gap between year-one performance and mature-deployment performance comes down to five things, in rough order of impact: how clean your knowledge base actually is (the single biggest lever), how broad your intent library is, whether the platform can take real actions like password resets and provisioning or only answer questions, how well the rollout was change-managed with end users, and simply how much runway the deployment has had — most of the real gains show up between month six and month eighteen, not week one.
None of this is a reason to skip AI at the service desk. It's a reason to build your board deck on the strict number, not the vendor's, so the second-year results land as a win instead of a disappointment nobody saw coming.