Deflection vs. resolution: measure the right thing

Here's a distinction that determines whether your support AI actually helps customers or just looks like it does on a dashboard: deflection versus resolution. Measuring the wrong one leads teams to celebrate AI that's frustrating their customers — a critical trap to avoid.

Deflection = the customer didn't reach a human agent. It counts any ticket the AI "handled" without escalation.

Resolution = the customer's problem was actually solved.

Why the difference is dangerous: deflection looks good (fewer tickets to humans = cost savings!) but it can hide failure. A customer who gives up in frustration, or who gets a wrong answer and goes away (temporarily), counts as "deflected" — the metric goes up while the customer is unhelped and angry. Optimizing for deflection alone incentivizes AI that blocks customers from help rather than solving their problems. You can hit a great deflection number while destroying customer satisfaction — which is exactly how the AI-support backlash happens.

The honest numbers: vendor marketing touts sky-high "resolution" rates (90%+), but those are best-case ceilings for deeply-integrated, well-guardrailed deployments. Realistic figures are more sobering — deflection varies widely (many teams land in the ~20–40% range, with weaker deployments near the bottom of that band), and genuine end-to-end resolution of complex or regulated tickets typically runs lower (30–60% for most teams), reaching high rates only with deep integration, permission to act, and tight guardrails. Treat vendor percentages as directional ceilings, model your own, and — crucially — measure resolution and satisfaction, not just deflection.

What to measure instead:

• Resolution rate — did the AI actually solve the problem? (Not just "didn't escalate.")
• Customer satisfaction (CSAT) on AI-handled interactions — are customers actually happy with the outcome? AI often runs measurably below humans on complex work — watch this gap.
• **Escalation rate and *why*** — how often and why the AI hands off (a healthy escalation is good — it's the AI knowing its limits, not a failure).
• Repeat-contact rate — did the customer come back with the same issue? (A sign the "resolution" wasn't real.)
• **Cost per *resolution*** (not just cost per contact) — the honest efficiency metric.

The principle: measure whether customers' problems are actually solved and whether they're satisfied — not just whether they were kept away from a human. Deflection without resolution is a false economy that trades short-term cost savings for long-term trust and churn. A support AI that genuinely resolves 40% of issues and keeps customers satisfied is far more valuable than one that "deflects" 70% by frustrating people into giving up.

How this shapes deployment: design your AI to actually resolve (which means grounding it in accurate knowledge and giving it the ability to take real actions — Module 2), and to escalate gracefully when it can't (rather than trapping customers). And measure resolution and CSAT, so you catch the difference between AI that helps and AI that just deflects.

The takeaway: don't be fooled by deflection. The goal of support AI isn't to keep customers away from humans — it's to solve their problems well. Measure resolution and satisfaction, treat healthy escalation as success not failure, and you'll deploy AI that genuinely improves support rather than AI that games a dashboard while eroding customer trust. This distinction underlies every good decision in the rest of the course.