AI Support Deflection Rate

Question

What support deflection rate does an AI agent need to beat current human ticket cost?

Quick answer

Formula: required_deflection_rate = (all_escalated_cost_per_ticket - human_cost_per_ticket) / (all_escalated_cost_per_ticket - ai_cost_per_deflected_ticket)

• Assumption: escalated tickets still carry AI first-pass cost before the human handoff cost is added.
• Assumption: the human baseline and escalated human handling cost are measured on the same resolved-ticket unit.
• Assumption: the required deflection rate is only meaningful if resolution quality stays acceptable after automation.

Example: if AI-only resolved tickets cost $0.12, fully escalated tickets cost $6.62, and the current human baseline is $4.50, the required deflection rate is about 32.7%.

How To Read the Threshold

• A lower threshold means AI resolves enough cheap tickets to beat the human baseline quickly.
• A threshold near 50% means blended economics depend heavily on real resolution quality and ticket mix.
• A threshold near 100% usually means the AI first pass is too expensive or escalations are too frequent.

Worked Threshold Example

Suppose AI resolves simple billing questions for $0.18 per ticket, but unresolved tickets still cost $0.06 of AI handling before a $5.40 human escalation. That makes all-escalated cost $5.46 per ticket against a current human baseline of $4.20.

The threshold is (5.46 - 4.20) / (5.46 - 0.18) = 23.9%. If fewer than roughly one in four tickets resolve fully in AI, the blended support path still costs more than staying fully human.

What Pushes the Threshold Up Fastest

• Too many AI turns before resolution or before handoff.
• Escalated tickets that still consume several AI retries before a human takes over.
• Cheap human ticket categories where automation savings are naturally small.

When the Target Is Operationally Unrealistic

A mathematically reachable threshold can still be a bad operating target. If the model needs a 70% deflection rate but your current ticket mix includes many policy exceptions, edge-case billing issues, or regulated workflows, the real answer is usually to narrow the AI scope instead of forcing the global target.

Related reads

• AI Support Agent Cost per Ticket
• How To Estimate Requests Per User/Month
• What Cache Hit Rate Means for RAG