Customers who are about to leave rarely say so directly on a support call. What they do is use specific language patterns -- combinations of words and phrases that, read in context, point unmistakably toward an account at risk. These patterns don't appear in satisfaction surveys. They appear in the transcript, usually around the three-minute mark, and they go unread in contact centers that review 2% of their calls.
Here are the five patterns that Voicemarrow flags consistently, what they sound like in practice, and why they're easy to miss in a manual QA workflow.
Signal 1: Repeated-Contact Language
The clearest early churn signal is when a customer explicitly references how many times they've called about the same issue. "This is the third time I've called about this." "I called last week and nothing got fixed." "I keep calling and nothing changes." These phrases indicate that the customer's tolerance threshold is close to exhausted.
What makes this signal distinctive is its specificity. The customer isn't expressing general frustration -- they're keeping count. A contact center that doesn't track first-contact resolution at the individual account level has no way to catch this pattern. The information lives in the transcript.
Signal 2: Competitor Mentions
Direct competitor references in a support call are a strong signal, not a vague one. "We're actually looking at switching to [competitor]." "A colleague uses [competitor] and says it's much easier." These aren't offhand comments -- a customer who brings up a competitor during a support interaction has done at least some evaluation of their options.
The challenge with competitor mentions is that they require keyword coverage of actual competitor names -- which changes over time as new entrants appear. A static word list maintained by a QA manager misses new names. A model that scores transcripts for company-name patterns alongside sentiment context catches current references.
Signal 3: Cancellation or Discontinuation Language
Explicit cancellation language is a late-stage signal but still actionable. "I want to cancel my account." "I'm thinking about cancelling." "How do I cancel?" A customer who says this on a support call is in the window between deciding to leave and leaving. That window is where retention interventions work.
What's notable about this signal is how often it appears not in escalation calls but in routine support calls. A customer calls about a billing question, the issue takes longer than expected, and somewhere in the call they mention cancellation in passing. In a manual review of 2% of calls, that call doesn't get pulled. At 100% transcript coverage, it's flagged within minutes of call completion.
Signal 4: Escalation Demand Patterns
Customers who demand to speak with a supervisor have already decided that the agent cannot solve their problem. "I need to speak to a manager." "Can you transfer me to someone senior?" The phrase itself isn't the signal -- the context is. An escalation demand in the first two minutes of a call suggests a customer who has been through this before and is skipping directly to what they believe is the only path to resolution.
The higher-value version of this signal is repeated escalation demand across multiple calls from the same account. A customer who has asked to speak to a manager twice in 30 days is not a routine support case. They are an account at serious risk, and the signal has been sitting in the transcript both times.
Signal 5: Sentiment Arc Collapse
Sentiment arc collapse is the hardest signal to catch manually because it requires reading the whole call, not flagging a phrase. A call that starts neutral or slightly positive and ends with a customer who has progressively shorter, more curt responses, longer silences, or declining engagement indicates a customer who came in willing to be helped and left feeling they weren't.
Sentiment arc analysis maps the emotional register of the call over time -- early call versus mid-call versus close. A collapse (positive opening, neutral middle, negative close) is a different pattern than a call that starts frustrated and gets resolved. The first is a churn signal. The second is a successful support interaction. Manual QA has difficulty distinguishing these at volume because the distinction requires reading the whole transcript, not scanning for keywords.
Catching These Before They Cost You an Account
The operational challenge is timing. Churn signal detection is only useful if it reaches the account team before the customer churns. A signal caught in a transcript reviewed three weeks later, in a manual QA sampling cycle, is not actionable retention intelligence -- it's a post-mortem data point.
At Voicemarrow, churn flag detection happens at the transcript level within minutes of call completion. A flagged account triggers a task in your CRM or a webhook to your account team -- the people who can actually act on it, before the customer completes their cancellation process or moves their contract to a competitor.
The five signals above don't require sophisticated inference. They require reading every call. The infrastructure to do that now exists. The gap is in connecting it to a response workflow fast enough to matter.