Predicting Customer Churn From CRM Data

In B2B, customers do not churn overnight — they leave a trail in your CRM data 4 to 7 months before the contract ends. Three signals are enough to build an early-warning system: widening order intervals, decaying activity, and a rising count of unresolved requests. Turn those three into a simple score, publish it as a weekly list, and you find out about churn while it can still be reversed, instead of when an invoice fails to appear.

Here is which data counts as a signal, how to set thresholds, and how to turn an alert into an action.

Why is churn always noticed too late?

Because most companies measure churn as an event: a contract is not renewed, an order does not arrive, the customer says they have moved on. Those are outcomes. The real sequence looks like this:

1. The customer gets frustrated about something and does not say it.
2. They place a small trial order with an alternative supplier.
3. They quietly reduce volume with you.
4. They take longer to respond to meeting requests.
5. At renewal, they cite price and leave.

By step five, nothing can be done. Steps one and three are visible in CRM data — they get missed because nobody is looking.

Which signals should you watch?

You do not need a machine-learning model. Five fields are enough:

• Order interval drift: If a customer normally orders every 45 days and the last order was 70 days ago, that is a signal. What matters is not the absolute gap but the deviation from that customer's own rhythm.
• Order size contraction: If the average of the last three orders is below 70% of the previous three, volume is slipping.
• Activity decay: If logged calls, visits and emails in the last 90 days are under half the historical average, the relationship is cooling.
• Unresolved requests: A support or delivery issue open for more than 30 days. This is the strongest leading indicator of silent churn.
• Contact turnover: If the buyer on the customer's side changed, the relationship reset. No introduction within 30 days means high risk.

How do you build the risk score?

Keep it simple. Assign points to each signal and add them up:

• Order interval stretched by more than 50%: 3 points
• Order value down more than 30%: 3 points
• Activity halved over 90 days: 2 points
• Request open longer than 30 days: 2 points
• Contact changed with no introduction: 2 points

Reading: 0–2 is normal, 3–5 goes on watch, 6 and above gets called this week. Calibrate against your own history: look at the scores your churned accounts would have had six months before they left. If fewer than 70% of them would have scored 6 or higher, lower your thresholds.

What happens when an alert fires?

An alert with no action becomes a tab nobody opens. Build three levels of response:

1. Watch (3–5): The rep makes contact every 10 days and logs the note. Not a sales call — a status call.
2. At risk (6–8): Request a face-to-face or video meeting with one clear question: "Has anything gone wrong on our side lately?"
3. Critical (9+): The manager steps in. Open requests get closed, and a commercial gesture is made if needed. The goal here is not a sale — it is getting the relationship back on the table.

Is price really the reason you lose customers?

"Price" is the most common entry in the loss-reason field, and usually the easiest one to type. Test it: count how many unresolved requests your churned customers had in the six months before they left. If that number is clearly higher than for retained customers, the cause is not price — it is operational problems nobody closed. Price is the polite sentence customers use on the way out.

What does a churn actually cost?

This is the business case for the whole system. Losing a customer worth $120,000 a year does not cost you $120,000:

• Lost revenue across the remaining relationship lifetime. At an average life of four years, the real number is closer to $480,000.
• Replacement cost: Winning a new customer of the same size typically costs 5–7 times more than retaining the existing one.
• Time cost: With a 68-day sales cycle, a replacement starts producing that revenue at least a quarter later.
• Reference loss: An unhappy departure talks in a small market. Unmeasurable and the most expensive line of all.

How do you measure whether the system works?

• Catch rate: What share of churned customers appeared on the list before they left? Target above 70%.
• Save rate: Share of at-risk accounts that stayed active.
• False alarm rate: If most of the list turns out healthy, the team stops trusting it. Keep it under 40%.
• Response time: Days between alert and first contact. Past seven days the system stops meaning anything.

Four mistakes to avoid

1. One threshold for every customer. A monthly buyer and a twice-a-year buyer cannot share a rule.
2. Watching revenue only. Activity drops before revenue does; the early signal lives in activity.
3. Keeping alerts on a manager dashboard instead of putting them in front of the rep.
4. Calling an at-risk customer with a discount. Offering a price cut to an unhappy customer admits the problem without fixing it.

All of this has one prerequisite: activities and requests actually being recorded. No record, no signal. Closync pulls those out of the conversation itself, so early warning works without asking reps for one extra field.