NexaWorks
nexaworks
Signal library

Signals in Call transcripts · 7 signals

The churn signals hiding in call transcripts.

Gong, Chorus, Fireflies — what people actually say on QBRs, cadence calls, and escalations.

Why this source matters

Calls are where customers think out loud. Tone shifts before language does; absence patterns appear before complaints. A transcript is the only source that captures both what was said and who stopped showing up to say it.

Call transcriptsTypical lead time: 30–90 daysFalse-positive risk: Medium

"Remind me what we pay for"

Questions about pricing, contract terms, and what's included spike in call transcripts. When customers start auditing the deal aloud, the renewal is already being litigated.

Detection logic

Phrase-pattern scan on transcripts: pricing, contract, renewal-terms, and cancellation-adjacent questions. Flag a rising trend over 3+ calls, not a single mention.

When it lies to you

New stakeholders legitimately need onboarding on the commercial relationship. Check attendee tenure before reading intent into it.

Call transcriptsTypical lead time: 60–120 daysFalse-positive risk: High

Sentiment drop across calls

The emotional tone of customer calls trends negative over a quarter — more frustration markers, fewer positive outcome statements. Tone shifts before language does.

Detection logic

Run sentiment scoring per call, then trend the account's average over 90 days. Flag sustained 20%+ declines; ignore single-call dips.

When it lies to you

One bad incident or a difficult personality can skew a quarter. Require the trend across multiple calls and multiple speakers.

Call transcriptsTypical lead time: 30–90 daysFalse-positive risk: Medium

Competitor named by the customer

The customer names a competitor unprompted — “we're looking at X”, “X offered us”. Unprompted mentions mean evaluation is already underway, not just beginning.

Detection logic

Maintain a competitor-name list; scan transcripts for mentions with surrounding context. Weight unprompted mentions far above answers to “who else are you considering?”.

When it lies to you

Classic discount leverage: naming a competitor to extract pricing. Look for depth — feature comparisons signal real evaluation, price talk signals negotiation.

Call transcriptsTypical lead time: 60–180 daysFalse-positive risk: Medium

Value language disappears

The customer stops talking about outcomes and results and talks only about tickets, features, and fixes. When the conversation becomes purely transactional, the relationship already is.

Detection logic

Track the ratio of outcome-words (results, ROI, goals, impact) to task-words (ticket, bug, feature, fix) per call. Flag a sustained inversion over a quarter.

When it lies to you

During active implementations the conversation is legitimately task-heavy. Compare against the account's own lifecycle stage, not a global benchmark.

Call transcriptsTypical lead time: 30–90 daysFalse-positive risk: Medium

Meeting no-show rate climbs

Scheduled calls get skipped, rescheduled, or attended by substitutes increasingly often. Calendar behavior is the most honest engagement metric you have.

Detection logic

Compute attended vs scheduled rate per account per month from calendar/call data. Flag a 25%+ drop from the account's baseline.

When it lies to you

Reorgs, holidays, and quarter-end chaos move calendars without meaning. Require the pattern across 6+ weeks.

Call transcriptsTypical lead time: 30–90 daysFalse-positive risk: High

New stakeholder asks fundamentals late

An unfamiliar attendee asks basic “what do you actually do” questions deep into the relationship. Either onboarding failed — or someone new is building the case to replace you.

Detection logic

Flag first-time attendees on established accounts whose questions map to vendor-evaluation patterns (capabilities, pricing, contract terms) rather than onboarding patterns.

When it lies to you

Genuine new hires need genuine onboarding. Check the attendee's role and tenure — evaluators ask about commercial terms, new users ask about workflows.

How to instrument it

Track attendee lists across recurring call series (the absence signal), maintain a short phrase-pattern library for commercial and competitor language, and trend sentiment per account over 90 days — never judge a single call. Pair transcript signals with calendar data for the no-show pattern.

Common pitfalls

Transcription mangles names and numbers, so verify entities before acting. Sentiment models need per-account calibration — a blunt New York buyer reads as “negative” to a generic model. And be transparent: customers should know calls are analyzed, ideally because you tell them it improves their service.

Signals in the other sources

CRMSupport ticketsProduct usageBilling & paymentsEmail & Slack
Get your free Account Risk Snapshot

We run signals like these against your 20 riskiest accounts — before you sign anything.