Resolution rate is the share of customer issues that end with the customer's problem actually solved, whether in one contact or several. It is the support measure closest to the customer's own view of success. Its weakness is that someone has to decide what "resolved" means, and an agent should not mark its own work.
Why does resolution rate matter once agents act?
Most automated support metrics measure activity: conversations handled, contacts avoided, minutes saved. Resolution rate measures the outcome. That makes it the measure to hold an AI agent to, and the one most worth protecting from inflation.
The inflation risk is specific. If the agent's own judgement that it resolved something feeds the metric, an agent that gives a confident wrong answer counts as a success. So does one that closes a conversation when the customer stops replying. Resolution needs evidence from outside the conversation: the issue did not recur, the task was completed, the customer confirmed it.
How is resolution rate measured in practice?
Strong measurement combines several signals for each issue:
| Signal | What it shows | Weakness |
|---|---|---|
| No repeat contact on the same issue within a window | The customer did not need to come back | Customers who gave up look the same |
| Task completed in a system of record | The address changed, the refund posted | Only works for issues with a system outcome |
| Customer confirmation | The customer says it is solved | Low response, skewed sample |
| Quality review of a sample | The resolution was correct | Small sample, effort to run |
Report resolution by intent and by who resolved it: agent alone, agent then person, person alone. Our guide on measuring a support agent beyond resolution rate sets this alongside other measures.
An example from a telecoms provider
Take a telecoms provider whose AI agent reports a high resolution rate for router faults, based on customers ending the chat after troubleshooting steps. Linking chat to later contacts, the team finds many of those customers booked an engineer visit by phone within days.
The agent's steps were correct but often insufficient, and it was not allowed to book engineers. Resolution, properly measured, was far lower. The provider adds a handover when troubleshooting fails, and later certifies the agent to book standard engineer slots. Measured from the system of record, resolution rises for real.
How is resolution different from deflection?
Deflection counts contacts that never reached a staffed channel, typically because self-service intervened. It measures avoidance. Resolution counts problems solved. A deflected customer may or may not have been helped; there is no way to tell from the deflection itself. That is the reason deflection is not resolution, and why projections built on deflection should be checked against resolution.
Containment has the same gap: a contained conversation ended without a person, which is not evidence that anything was solved.
Which terms sit next to resolution rate?
- First-contact resolution: the share resolved in the first interaction, a stricter subset.
- Time to resolution: how long resolution took.
- Escalation rate: handovers to a person, which often lead to resolution rather than failing it.
Where Pryme fits
In Pryme, resolution is not something an agent asserts about itself. Actions an agent takes, such as a posted refund or a changed address, are filed with the system-of-record outcome, so resolution can be measured from evidence. A support agent that knows when to say no hands over when it cannot finish.
Questions people ask
Does a handover to a person lower the agent's resolution rate?
It depends how you report it. Count the issue's resolution regardless of who finished it, and separately track which issues the agent finished alone. A well-judged handover that leads to resolution is a good outcome.
What is a reasonable resolution rate for an AI agent?
There is no general benchmark worth quoting. It depends on which intents the agent handles and what authority it has. Compare by intent against your own human baseline instead.