85% of CX leaders say customers will drop a brand over a single unresolved issue, and more than 60% of customers now say they’ll defect after just one bad experience, up 22% from the year before (Zendesk, CX Trends 2026 / Customer Service Statistics). Patience for a rocky support interaction is shrinking fast, which means the window to catch a call before it escalates is shrinking with it.

Part of what makes escalations expensive is that customers experience them as compounding. 74% say it’s frustrating to repeat their story to a new agent every time a call gets passed up the chain (Zendesk, CX Trends 2026). Each handoff resets the customer’s patience clock to zero while doing nothing to reset the manager’s.

Escalations don’t start at the escalation

By the time a customer says “let me talk to your manager,” the actual decision point was usually several exchanges earlier, a misunderstood request that didn’t get clarified, a promise from a previous call that didn’t get honored, or a tone shift the agent didn’t register in time. None of that shows up in a ticket tag. It shows up in the conversation itself, in patterns that most QA processes never reach because they only sample a small fraction of calls after the fact.

Zendesk’s own data shows roughly 6% of chats get routed to escalation or a ticket once agents flag them as complex, a useful number, but a lagging one. It tells you an escalation happened. It doesn’t tell you it was coming.

What actually precedes a blow-up, in the data

Looking across escalated versus resolved calls, a handful of signals show up consistently well before the customer asks for a supervisor:

• Repetition without acknowledgment – the customer restates the same issue a second or third time, and the agent’s response doesn’t change
• Sentiment drops after a hold or transfer – tone shifts noticeably worse right after being put on hold or bounced to another queue
• Promise language that doesn’t get logged – “I’ll make sure this gets fixed” with no corresponding note or follow-up action
• Rising interruption rate – the customer starts talking over the agent, a strong precursor to asking for escalation outright

Any one of these, alone, isn’t a reliable predictor. Together, and tracked automatically across full call volume rather than a QA sample, they’re a workable early-warning system, the kind that lets a supervisor step in during the call that’s heading sideways, not read about it afterward in a CSAT survey.

Related reading

This builds on Closed-loop customer feedback: how conversation intelligence turns VoC into action, which covers routing conversation themes to CS, product, and support after the fact. This post is the earlier intervention point, catching the signal while the call is still live, before it becomes the feedback the post is about to close the loop on.

See it on your own calls

RepsMate flags escalation risk in real time: sentiment shifts, unacknowledged repetition, broken promises,  so supervisors can step in before the call ends badly, not after. Get a demo and see what it catches in your queue this week.