Deflection Rate Is the Wrong Way to Measure AI Support — Pylon's Own Numbers Prove It

Pylon’s co-founders used their SaaStr AI Day session to attack the metric most support teams are optimizing for: deflection rate. The company’s argument, walking through an agentic support product it shipped the week before the talk: a 5,000-person customer with roughly 1,000 support staff deployed a fully automated resolution agent that deflected about 50% of tickets, and headcount didn’t move. CEO Marty Kausas’s explanation: ticket count isn’t work volume. Agents clear the easiest, fastest tickets first; everything genuinely hard still escalates to a human, so a 50% deflection number can coexist with an unchanged headcount need. Two examples from the session showed the more useful shift — tickets arriving at human agents already pre-investigated, including one that replaced what used to be a manual AWS log query. Beta customers reported 70% fewer escalations at one company and 64.5% faster first response at another.

Kausas drew the comparison to coding directly: AI spend has “meaningfully taken off” there, and he uses Claude and Codex daily without either coming close to automating him out of the job — his case for treating support the same way, as human-plus-AI augmentation rather than a replacement play measured by how many humans you can remove.

Gartner’s own forecast backs up why “deflection rate” is the wrong number to chase. It predicts AI agents will outnumber human sellers 10 to 1 by 2028, yet fewer than 40% of sellers will say agents actually improved their productivity — and VP Analyst Dan Gottlieb’s explanation reads like it was written for Pylon’s argument: “sales productivity does not stall because reps forget how to sell; it stalls because the system quietly caps them.” Deflection rate measures whether the system removed a ticket. It says nothing about whether the system removed the work.