Deflection counts tickets that never arrived. Resolution counts problems that got solved. They are not the same metric, and only one of them survives contact with an angry customer.
Support automation gets reported with two numbers that sound interchangeable and are not. Deflection rate measures contacts that did not reach a human. Resolution rate measures problems that got solved. A tool can post an excellent deflection rate purely by being hard to escalate from.
| Metric | Formula | What inflates it artificially |
|---|---|---|
| Deflection rate | contacts avoided ÷ total potential contacts | A hard-to-find contact button, a self-service page nobody found useful, or a customer who simply gave up. All three reduce contacts without solving anything. |
| Resolution rate | conversations resolved ÷ conversations attempted | Counting silent abandonment as success, and never reversing a resolution when the customer comes back with the same question. |
| Automated resolution share | conversations resolved with zero human replies ÷ all conversations | Counting conversations where an agent replied once. This is why attribution has to be checked per message author, not per conversation outcome. |
Take a month with 1,000 support conversations. The AI answers 700 of them without a human. Of those 700, 120 customers come back within three days with the same question and get a human reply.
That gap is precisely why our own definition applies a settling window rather than counting at the moment the AI stops typing — the detail is on how we measure AI resolutions.
Deflection is a reasonable capacity signal: it tells you roughly how much human time a self-service investment freed up, which is what you need for staffing decisions. It is a poor quality signal and a poor billing basis, because nothing in the formula distinguishes a customer who was helped from a customer who left.
Deflection rate is the number of contacts avoided divided by the total number of potential contacts, expressed as a percentage. Because 'avoided' includes customers who gave up as well as customers who were helped, a high deflection rate is not by itself evidence that self-service worked.
AI resolution rate is the number of conversations the AI resolved divided by the number it attempted. A strict version counts only conversations with zero human replies, excludes conversations the customer never engaged with, and subtracts resolutions that are reversed when the customer reopens the conversation.
It is a useful capacity metric and a poor quality metric. Deflection tells you how much human handling time was avoided, which is genuinely useful for staffing, but it cannot distinguish a customer who was helped from one who abandoned the conversation, so it should not be used to judge answer quality or to calculate a bill.
It depends almost entirely on how much of your ticket volume is answerable from documented knowledge, so a single benchmark is misleading. The useful comparison is against your own baseline: measure the share resolved without a human in your first weeks, then track whether it climbs as knowledge-base coverage grows.
2026-08-09 · 10 min read
Two billing shapes, three worked scenarios and the exact ratio where each one starts to win. Real arithmetic on published rates, including what a peak month does to a variable bill.
2026-08-04 · 9 min read
Intercom, Zendesk, Gorgias, Freshdesk and Tidio all bill for AI outcomes — and all define them differently. A sourced, side-by-side look at what each one actually charges you for.
2026-07-17 · 5 min read
Per-resolution AI pricing is becoming the industry standard — but the definition of a 'resolution' decides whether it's fair. Here's how to evaluate it, whatever tool you choose.
See how honestly-measured AI resolutions cut your support load — start on the Free plan, no credit card, no sales call to get started.