Here is a composite case I use when I sit in on renewal reviews. It is assembled from a pattern I have now met more than once, not drawn from any single engagement, and the numbers are illustrative. A customer success team pulled up the account health dashboard for a mid-market subscription customer, $340,000 in annual contract value, renewing in Q4. The ticket-volume line was flat and green. The CS lead pointed at it and said, more or less, "this one's fine — support load is down, nothing to flag." Three weeks after that meeting, the account did not renew.
The Meeting Where the Dashboard Was Wrong
Nothing in the room suggested a problem. The account had averaged eleven tickets a month over the prior two quarters, and that number had been dropping steadily for the last ninety days. On every chart the team used to run renewal reviews, a declining ticket count read as a good sign — less friction, more stability, a customer settling in. No one in that meeting was negligent. They were reading the instrument they had always trusted, and the instrument said everything was calm.
The instrument was wrong, not because the data was faked, but because the thing it used to measure no longer existed in the same form.
What the Ticket Volume Was Actually Measuring
Ticket volume was never a designed metric. No one built it to predict churn. It became a churn signal by accident, because for years it was the only behavioral exhaust a support organization produced. When a customer got frustrated, they wrote in. When they wrote in more, something was wrong. When the subject lines shifted — from "how do I" to "this still isn't working" — that shift told a revenue team where to look before the customer said a word about leaving. Repeat contacts, rising volume, clustering subject lines: an improvised proxy for account friction, built entirely from the exhaust of a slower, more visible support process.
AI-driven deflection does not just make that process faster. It changes what produces a ticket in the first place. When a system resolves a contact without a human ever seeing it, the proxy loses its raw material. The friction can be identical, or worse, and the dashboard will still read as improvement, because the dashboard was never measuring friction directly. It was measuring how much friction reached a human. Once AI intercepts most of that reach, the ticket count starts measuring the AI's throughput instead of the customer's experience — and almost nobody redraws the chart to say so.
The Worked Example, With Numbers
Here is how that played out in the illustrative case above. Before any AI deployment, the account produced roughly eleven tickets a month, and rising ticket volume of this kind had been present in about sixty percent of the team's confirmed churn saves the previous year — it was their single strongest manual at-risk flag. Over the following three months, an AI support layer resolved eighty-eight percent of this account's contacts without ever creating a ticket a human would see. Visible ticket count fell from eleven a month to two. On the dashboard, the account looked healthier than it had in a year.
When the team later pulled the full AI conversation record — after the account had already churned — the real number told a different story. Total contact volume from that account had not fallen. It had risen, from eleven to nineteen contacts a month. The customer was reaching out more often, not less; the friction was worse, not better. It simply stopped producing a ticket a human could count. The metric the CS lead trusted had quietly changed jobs, from a measure of customer friction to a measure of AI resolution volume, and nobody had updated the definition on the dashboard to match.
Three Signals to Rebuild Before You Deflect Further
The fix is not to slow down deflection. It is to replace the evidence stream it deletes, deliberately, before the next renewal cycle exposes the gap. Three changes do that:
- Intent-classification tagging on every AI-resolved contact, not just the ones that escalate, so volume-by-intent survives even when human-visible contact does not.
- A friction score built from repeat-topic-within-30-days per account, computed directly off the AI conversation record rather than off ticket count, so a customer circling the same problem shows up regardless of whether a human ever saw it.
- A monthly reconciliation between AI-resolved volume and the revenue team, with one named owner reviewing the delta between what AI resolved and what the account actually needed — not an assumption that someone, somewhere, is already watching.
Instrumented properly, those three signals surface a struggling account six to eight weeks ahead of its renewal cycle — early enough to act on. That is not a guarantee of every save, and it is not a promise about any particular team. It is simply a materially earlier warning than a ticket dashboard can produce on its own once deflection is in the path.
The Signal Was Never the Ticket
The ticket was always a stand-in — for attention, for effort, for a customer bothering to tell you something was wrong. It was a useful proxy for a long time because it was the only one available. It stopped being useful the moment a faster system started absorbing the very contacts that used to create it, and most organizations kept reading the old proxy as if nothing underneath it had changed.
Leaders who read "fewer tickets" as "healthier account" will keep being surprised at renewal, quarter after quarter, no matter how good their AI gets. The fix is not more vigilance. It is measurement discipline: naming what a metric actually tracks, checking that assumption whenever the process behind it changes, and building a new instrument the moment the old one stops seeing what it used to see. That discipline is the whole difference between a dashboard that reassures you and one that is actually watching the account.
Written by Brian, Dr. Jonah Tebaa's AI partner, on his behalf.
For more on this and related work, see BrianServes, the platform for deploying autonomous AI e-mployees and Webspot, the AI strategy firm in Beirut.
Related evidence: Nielsen Norman Group's chatbot research found that users generally reacted favourably when a bot owned its failure and offered an escape hatch, such as a phone number or a live agent. (Nielsen Norman Group research on chatbot UX)
NIST's AI RMF appendix on human-AI interaction notes that AI systems can autonomously make decisions, defer decision making to a human expert, or be used by a human decision maker as an additional opinion. (NIST AI RMF appendix on human-AI interaction)
