Every support team measures the same four things: first-response time, full-resolution time, CSAT, and ticket volume. We ran a regression on 14 months of customer-outcome data against churn at the workspace level. Two of those four correlate strongly with retention. Two correlate weakly. And two we never measured correlate the strongest.
The methodology
14 months of monthly snapshots from 4,200 paid workspaces. Outcome variable: churn within 6 months of the snapshot. We ran a multivariable regression with workspace size, industry, plan, and seven support metrics as independent variables.
The two metrics that did NOT predict retention much
(1) Median first-response time. p-value 0.31. Customers tolerate slower responses if the eventual response is good. (2) Ticket volume per agent. p-value 0.42. Loaded teams who do great work retain just as well as light teams who do mediocre work.
The two metrics that predicted strongly
“The p95 resolution time is more predictive than the median. One bad multi-day saga can churn a customer who otherwise has 50 good interactions.”
(1) Full resolution time at the 95th percentile. p-value 0.001. Customers who have any single bad experience that drags on for days churn at 3.2x the base rate. (2) Repeat-contact rate (same customer, same intent, within 30 days). p-value 0.004. Tells you whether the problem was actually solved or just deflected.
The two metrics we now publish that we never used to
(1) Confidence-corrected CSAT. We weight survey scores by the customer's lifetime value and recency. A 5-star from a 6-month enterprise customer counts more than a 5-star from a one-time visitor. (2) Escalation-from-AI quality: when the AI escalates to a human, what fraction of those escalations were 'good catches' (a human-only issue) vs the AI giving up. Ours is 87%. Public 12 months after we shipped it.
What to do with this
Stop optimizing for median first-response time. Optimize for p95 full-resolution and for repeat-contact rate. The bad-day-saga is what kills retention; the slow-but-helpful average reply does not.