Resolution rate is the share of conversations the AI actually solved. Deflection rate is the share that never reached a human. Every resolved conversation is deflected, but not every deflected conversation is resolved, and the gap between the two is where unhappy customers hide.
| Metric | What it counts | How it misleads |
|---|---|---|
| Resolution rate | Conversations solved by the AI, by the vendor's definition. | Depends entirely on what counts as a resolution. |
| Deflection rate | Conversations that did not create a human ticket. | Counts customers who gave up, called instead, or came back later. |
| Containment rate | Conversations that stayed inside the bot. | The same problem as deflection, under another name. |
| Automation rate | Share of all contacts touched or handled by automation. | Often includes simple routing and form-filling. |
The evidence that the gap is real
One customer case published by Forethought shows it clearly: deflection reached 60% quickly, while customer satisfaction sat at 3.1 out of 5. Only after months of tuning did both numbers become healthy. A high deflection figure on its own told the team nothing about whether customers were being helped.
Vendor averages and named-customer results also diverge. Ada advertises 84%, while one of the customers on its own site reports 45%. Intercom Fin advertises an average of 76%, and a named customer reports 50.8%. Neither vendor is necessarily wrong: averages hide a wide spread, and your result depends on your own content and ticket mix.
What to ask for
- The resolution rate and the customer satisfaction score for AI-handled conversations, side by side.
- The repeat-contact rate: how many "resolved" customers came back about the same issue within seven days.
- The rate as a share of all conversations, not only those the AI attempted.
- Results from two or three customers with a ticket mix like yours, not the average.
- The wrong-answer rate, and how the vendor measures it.
Frequently asked questions
What is a good resolution rate for an AI support agent?
Named customers in the case studies we read report anything from about 30% to 80%. Simple, repetitive, well-documented questions push the number up; account-specific and emotional issues pull it down. Judge a vendor on results from companies like yours, then measure your own in a pilot.
Can a deflection rate be too high?
Yes. If deflection rises while satisfaction falls or repeat contacts climb, the bot is blocking customers rather than helping them. Always read the two together.
Keep reading
- What counts as a resolution?guide
- Per-resolution vs per-seat pricingguide
- How to run an AI agent pilotguide
- Agentic customer service, explainedguide
- The best AI customer service agents in 2026best list
This page was researched and drafted with AI assistance from the sources listed on it. We have not run hands-on tests of these products. Method: How we review