Churn Prediction: Signals, Not Opinions
Churn prediction replaces the rep's gut call with signals already sitting in your CRM. Why a discount is the most expensive answer to a risk flag.
An account shows up as at risk in the quarterly review. The account executive says the customer is basically happy and mentions the five percent discount that went out last week. A six-figure annual contract has just been decided on the strength of one person's story.
In the same week, a recording sits in the company's own systems in which that same customer says plainly that they will leave if a particular feature does not ship. Nobody opened it. The company pays for visibility and runs on instinct.
Churn prediction—forecasting attrition from observable signals rather than from impressions—is what is missing here. Not as one more tool, but as an evidenced second opinion drawn from data the business already owns.
Key takeaways
- A contract is a promise to pay. It proves nothing about whether the customer got value last month.
- The reliable second opinion on churn risk already lives in call recordings, email threads and support tickets, not in the next piece of software.
- A discount buys another term and never answers the cause. It is the most expensive possible reaction to a signal nobody read.
- Recommendation: before your next churn review, take one at-risk account and put the actual conversation record next to the rep's assessment. The gap between them is your blind spot.
Why an active contract is not proof of value
An active contract looks like predictable revenue. What it actually proves is that somebody signed twelve months ago, and it says nothing about what the customer got out of it last month. Without that evidence, the cancellation arrives as a surprise when it never was one.
Manuel Hartmann, Founder and CEO of SalesPlaybook, puts the test in one line in his LinkedIn newsletter Diary Of A CRO: recurring revenue requires recurring, demonstrable impact. A business that cannot show it does not lose the account to a competitor. It loses the account to its own invisibility.
From the LinkedIn newsletter Diary Of A CRO by Manuel Hartmann, issue of 18 June 2026.
Our position: retention is a systems question, not an account-management question. We build revenue systems so that delivered value becomes visible in the same place as revenue, inside the CRM, every month, without anyone assembling a report by hand. A team that only tells the value story at renewal has not had one for eleven months.
Grant Coleman, CRO at Uberall, names the thing it turns on in the same newsletter: "The mechanism is less important - whether it be SaaS, AI, or a merger of the two. What business outcome do you provide? That's what matters." Whether the product is sold as software or as AI decides nothing. The demonstrable outcome decides.
The most expensive signal in a churn review is an opinion
In most revenue teams the risk call comes from the person with the strongest incentive to deliver good news. That is a role problem, not a character problem. It stays expensive either way, because a decision about an annual contract rests on a story nobody checked against anything.
Coleman shows how far the story and the record can drift apart, using a case from his own team: "The AE's read: 'I think they're pretty happy, gave them a 5% discount.' The transcript showed the customer saying explicitly: I will churn if this feature doesn't work." The fix turned out not to be a pricing question at all. It was shipping one thing, and that opened an upsell conversation.
Answering a risk flag with a discount before anyone has checked the cause. It costs margin this year, pushes the cancellation out by one contract term and makes the cause harder to see, because the numbers look healthy again afterwards. You can spot it when the review talks about price before anyone has opened the last conversation record.
Our position: an assessment does not improve because a room debates it. It improves when an independent source sits next to it. So every risk call we run starts with what the customer actually said, and the team's interpretation comes second.
Where the signals already sit, and why nobody sees them
The evidence for a churn risk is nearly always already inside the building. It just sits in formats no report can read, so it never reaches a dashboard. The result is both absurd and common: a team buys another tool to surface insight it already owns.
Turning a signal into something you can act on takes two things. First, a field on the customer record (in HubSpot, a property) that holds the signal as a value rather than as free text buried in a note. Second, an automated rule (a workflow) that reads the field and raises a task whenever the value changes. Skip the first and the rule has nothing to read.
| Signal | What it says about the risk | Countermeasure | Source |
|---|---|---|---|
| Recording of the last customer call | The customer's own words, timestamped: the only evidence nobody has interpreted yet | Read it before the review, not after | Diary Of A CRO, 18 June 2026 |
| History of support tickets | Repeat tickets on the same feature point to an unresolved blocker, not an unhappy person | Group tickets by theme and hand them to product | Diary Of A CRO, 18 June 2026 |
| Email thread with the decision maker | Response times and who is copied reveal whether the account still has an internal champion | Reach new participants deliberately, before the champion moves on | Diary Of A CRO, 18 June 2026 |
| Usage of the feature being paid for | An account that does not use what it bought has no value to lose | Tie usage to the promised outcome and show it monthly | Diary Of A CRO, 18 June 2026 |
| Growth among the largest existing customers | The counter-test: where value is evidenced, revenue in the base grows instead of eroding | Track Net Revenue Retention separately for top accounts | Case study cito |
Our position: we keep these signals as fields rather than notes, because only a field can carry a report, a rule and an alert. A note carries a reminder, and only for the person who wrote it.
From signal to action in four steps
Without a fixed sequence, every risk flag ends in a one-off decision. Two comparable accounts then get handled differently, and afterwards nobody can say what worked. Four steps are enough to turn a signal into an action you can trace.
- Detect the risk. The signal is stored as a value on the customer record the moment it appears, instead of being narrated in the review.
- Name the cause. The verbatim conversation record decides what this is about, not the rep's summary of it.
- Assign the countermeasure. Every cause has a pre-agreed response. Price is never the first one.
- Write the outcome back. After the countermeasure, record whether it worked. Skip this and the system learns nothing.
Our position: the fourth step is the one almost everyone drops, and it is the only one that turns a pile of signals into a prediction. Without the outcome written back, you start measuring from zero again next quarter.
Why RevOps is not a cost centre
Revenue operations is treated as administration in a lot of companies and staffed accordingly. The consequence is quiet and expensive: what one person figures out stays with that person, and the business pays the same learning curve once per employee.
Coleman frames the alternative as a build decision: "I'd rather my RevOps team build the right tooling - because they've got access to all the data, not just the data on one laptop. Then provide the outcomes as easy-to-use tools for the whole team." One person's find becomes shared infrastructure, and the leverage reaches everyone instead of one desk.
What that looks like in an existing customer base shows up in our own client work. At cito, the result was not better sentiment. It was a number:
5x
more deals and more than double the growth in the existing base, with HubSpot
Source: Case study Perspective
Both numbers come out of the same work: knowledge about the customer moved out of people's heads and into the system, where anyone can read it. SalesPlaybook is a HubSpot Diamond Partner, and HubSpot is where those signals converge here too, with the customer record, the ticket, the workflow and the report in one place.
Want to know which signals are already sitting in your CRM and going unread?
Free · 60 minutes · no pitch · a straight fit-or-no-fit answer.
What this article does not claim
A prediction does not replace a conversation. It tells you where to look, never what to do. Confuse the two and you have traded a gut feeling for a number with decimal places, which is not an improvement.
Three limits belong with this, and they come from practice rather than modesty.
No written-back outcome, no model
Signals without a recorded result are a collection, not a forecast. The fourth step is the precondition, not the polish.
A few large accounts beat any statistic
With twenty accounts in the base, no calculation holds. There the conversation record replaces the model, and that is the right order.
A signal is not a verdict
A recurring support ticket can also mean somebody is working hard in the product. Interpretation stays a human job.
One question to answer before the next tool arrives
Can your business say today, for its twenty largest customers, what outcome it delivered them last month? If not, you do not need a prediction model. You need a measurement point. The first step costs no software, only an hour: pick one at-risk account, read the actual conversation record and set it beside the rep's assessment.
Free · 60 minutes · no pitch · a straight fit-or-no-fit answer.
Frequently asked questions
What is churn prediction?
Which signals reveal churn risk earliest?
Why is a discount the wrong answer to a churn risk?
Does churn prediction require a dedicated tool?
How many customers do you need before a model pays off?
Customer proof
See how other revenue teams solved it.
Explore documented outcomes from comparable pipeline, CRM and sales execution projects.
View relevant client stories