Deflected or resolved: the metrics gap in AI support

AI keeps over 45% of contacts away from human agents, but only around 14% reach full self-service resolution. The gap between those two metrics runs to 20–40 percentage points — and it’s what optimistic automation reporting conceals.

Key figures

• AI deflects over 45% of contacts from humans Industry AI support benchmarks, 2026
• Only around 14% reach full self-service resolution Industry AI support benchmarks, 2026
• Median self-service deflection — 22% (range 8–45%) Industry AI support benchmarks, 2026
• Tier-one median 41.2%, top quartile 58.7% Industry AI support benchmarks, 2026

What’s the difference between the two metrics?

What they count. Deflection is any conversation that never reached a human agent. Resolution is an issue closed from start to finish.

The first metric credits a customer who got an answer and a customer who gave up equally. The second only credits the first. That’s why deflection is convenient for reporting and nearly useless for understanding quality.

How large is the gap?

Substantial. Benchmarks show AI deflecting over 45% of contacts Industry AI support benchmarks, 2026 while only around 14% reach full self-service resolution Industry AI support benchmarks, 2026 , with the divergence between the two metrics across programmes running to 20–40 percentage points Industry AI support benchmarks, 2026 .

So in a typical programme most «deflected» contacts aren’t resolved. They either return later or become a lost customer the report never sees.

What levels count as normal?

It depends which tier you mean. Median self-service deflection is around 22% across a range of 8–45% Industry AI support benchmarks, 2026 , while for tier-one contacts the median is higher — 41.2% with a top quartile of 58.7% Industry AI support benchmarks, 2026 .

The spread here exceeds the difference between vendors, and that matters: deflection is driven primarily by contact mix and knowledge-base quality, not by platform choice.

What should you measure instead?

Repeat contacts. The most reliable indicator that a deflection was a genuine resolution is the absence of the same customer with the same question over the following days.

That metric is harder to collect and impossible to present as a large number in a report. But it’s what separates automation that removes work from automation that postpones it.

Where do these figures come from?

This report aggregates published 2026 AI support benchmarks. The metrics come from enterprise-level programmes; deflection and resolution are defined differently across vendors, which partly explains the spread in values. The ranges given describe inter-company variation rather than measurement uncertainty.