Last year 29 of these 80 accounts left, and $5.5 million in revenue left with them. All 29 had written a warning into a support ticket before they went. Almost everyone complains in support, so volume alone is noise. Two specific sentences told you which complaints were real. Here is the whole chain, in one figure.
Sixty-seven of the 80 accounts grumbled in support about something. Listening for unhappiness, on its own, would have flagged most of the book. The accounts that actually left were saying something more specific.
When a customer wrote that they were weighing another vendor, or that something they were promised never arrived, the account left far more often than one that never used those words. The gap is not subtle, and it is measured against what really happened at renewal, not a sentiment score.
Put the two strongest together and they reach deep into the book. Thirty-nine accounts, carrying $6.65 million in yearly revenue, wrote at least one of the two sentences, and 72 percent of them left. That is the population a save team should be reading for, every week. One signal is worth setting aside: nearly every departing account had also written that the same problem failed again, but so had many that renewed, so on its own it flags too many to act on. It is a screen, not a separator.
This is the part the current early-warning system gets backwards. For 25 of the 29 accounts that left, the earliest warning came in a support ticket, not a survey. The surveys were first in only 4 cases.
And for 9 of the 29, the survey and health programs never produced a warning at all, even though the account left and the support text had already turned negative. The playbook waits on the survey, and for a third of these accounts the survey simply never spoke.
The reason is structural, not a failing of any one team. A support ticket is written the moment something goes wrong. A survey goes out on a schedule, and a health score only moves once its inputs move. The earliest, most specific account of the problem is sitting in the support text, and almost all of it lives in one channel, shown below.
It would be easy, and wrong, to say the dashboard missed these accounts. It did not: it marked 28 of the 29 departing accounts at risk before they left.
The problem is two sided. It marked them later than the support text did, and it marked too much. It also called at risk on 38 of the 51 accounts that went on to renew. A warning that fires on three quarters of your healthy accounts is hard to act on. The support phrasing is both earlier and more selective, which is why it belongs at the front of the early-warning system, not behind the score.
The same reading that surfaces the leavers surfaces the loyalists, and just as clearly. All 51 accounts that renewed named a concrete, measured result they credited to the platform.
They did not simply sound pleased. They pointed at agents handling more, work taken off the team, a number that moved. That specific, results based praise is the asset worth protecting, and the place to ask for a reference or an expansion.
The signal already exists, unread, in the support text. Three moves turn it into retained revenue. Sizing is deliberately conservative: it assumes you save only one account in seven that you reach in time.
Flag any account whose support text says it is weighing another vendor, or that a promise was not kept. These two phrases reach 39 accounts and $6.65 million in revenue, and 72 percent of the accounts that wrote them left. Saving one in seven of that group is worth roughly a million dollars a year in kept revenue.
For these accounts the first support warning landed a median of seven months before they left, and for nearly a third the surveys never warned at all. Route a flagged account to a save owner the day the phrase lands, while there is still time to change the outcome, rather than waiting on a survey that may never come.
The score flags three in four accounts that go on to renew, so it cannot stand alone as the trigger. Use it as background, and let the specific support phrasing decide who gets a save motion. Keep the "same problem failed again" pattern as a watch list, not an alarm.
This brief reads 782 customer messages from 80 enterprise accounts across eight feedback systems (support tickets and chat, the NPS, CSAT, renewal and onboarding surveys, the quarterly business reviews, the expansion notes, and the product feedback), from October 2024 to March 2025. Only messages written by the customer were read; replies from Zendesk staff were not. Every "more likely to leave" figure compares the share of accounts that left among those who wrote a given sentence with the share among those who did not, measured against the real renewal outcome on each account, with a test of statistical significance. The renewal outcome is fact, not a prediction. The timing figures are account level: for each of the 29 accounts that left, the earliest support warning (a competitor or alternative named, a promise unmet, a problem fixed then broken again, a high-effort unresolved issue, or an escalation) was compared to the renewal date and to the earliest survey or health warning.
What to hold lightly. The "same problem failed again" phrase appears in nearly every departing account but also in many that renewed, so it screens rather than separates. Exactly one departing account ($54,000) was rated healthy by the score and still left; that single account is noted, not leaned on, and the larger gap is that the surveys often produced no reading at all. The retained-base result categories overlap, so they are read as the distinct-account view, not as a sum. Account revenue figures come from the account records; the public figures on ownership, pricing, the market, and retention benchmarks are cited from public reporting and are never mixed with the figures drawn from these messages.