AI Agents in Customer Support: The Real Performance Numbers
What three years of production deployment tells us about deflection rates, escalations, and CSAT.
Customer support is the canonical 'AI replaces a workflow' story, and the numbers are now mature enough to be honest. Well-deployed agents resolve 35–60% of tier-one tickets fully, depending heavily on product category, knowledge-base quality, and the strictness of refund and account-change policies. The remaining tickets cost more to handle than they used to, because the easy ones never reach a human.
The implication is operational, not technical. Support teams now skew senior, handle harder cases, and need different metrics. Treating an AI deployment as a headcount reduction without restructuring the team underneath is the most common cause of failed projects.