A churn prediction is only worth something if it changes what a person does on Monday morning. Many teams buy or build a model that outputs a risk percentage for every customer, then discover that nobody knows what to do with 400 numbers. The better target is much smaller: a short, ranked list of accounts, with a reason beside each one, and a named person responsible for the next step.

Define churn before you predict it

It sounds obvious, and it is skipped constantly. Does churn mean a cancelled contract, a non-renewal, a downgrade, or a customer who has simply gone quiet? Each has different signals and a different response. Write one sentence for the definition and the time window: for example, “customer does not renew at the end of their current term.” Prediction models and dashboards built on a fuzzy definition give fuzzy answers.

Also decide which customers belong in scope. Brand-new accounts, accounts in a free trial and long-term enterprise contracts behave differently. Mixing them produces a list that is wrong for everyone.

Signals worth tracking

You do not need hundreds of inputs. A handful of signals, collected consistently, beats a sprawling set of unreliable ones. Typical categories:

• Engagement: Product usage trends, logins, activity with key features. Look at direction of change, not just level.
• Relationship: Whether the champion has changed roles, how recently someone spoke with the customer, whether you have more than one contact.
• Support: Volume and tone of tickets, unresolved issues, repeated escalations.
• Commercial: Payment delays, contract terms, usage versus plan, recent downgrades.
• Sentiment: Notes from calls and survey comments, where available.

Some of these sit outside the CRM, in product analytics, billing or support systems. Bringing them together is often the hardest part of the project, and it is where much of the real value comes from.

Rules first, models second

You can begin with a simple rule-based list before any machine learning. Something like: no contact in a defined period, usage down meaningfully from the previous period, an open high-priority ticket, and renewal inside the next quarter. Count how many flags each account has and rank by that.

A rules list is transparent, easy to explain and quick to build. It also teaches you which signals matter in your business. When you later add a predictive model, you can compare it with the rules and see whether it genuinely finds risk the rules miss.

Rank by risk and by value

Risk alone gives a misleading list. A high-risk account that is small and unprofitable deserves a different response from a moderate-risk account that represents a large share of revenue. Put risk and value side by side, and sort so that the accounts where intervention matters most appear first.

Keep the list short enough to work. If a customer success manager can reasonably have a meaningful conversation with eight accounts a week, a list of eighty is a backlog, not a tool.

What each row should contain

Column	Why it helps
Account and owner	Who acts
Renewal or term date	Urgency
Risk rating (high, medium, low)	Quick sort without false precision
Top two reasons	Gives the owner a starting point
Account value band	Prioritization
Last meaningful contact	Shows whether it is a neglected account
Suggested action	Concrete next step

Use bands rather than precise percentages. A score of 73 versus 71 implies precision the model does not have.

Turn the list into actions

For each flagged reason, pre-define a response. Low usage might lead to a training or onboarding session. A departed champion might call for outreach to find a new sponsor. Unresolved tickets call for a manager-level check-in. Payment trouble goes to finance with a gentle, human approach before any automated reminder.

The list is not a verdict. Some flagged accounts are fine, and some unflagged accounts will leave. Ask owners to record the outcome of each action and whether the flag turned out to be right. After a couple of quarters, that feedback is the most valuable data you have for improving the list.

Common pitfalls

Alert fatigue is the biggest one. If every account is high-risk, nobody reacts. Tune thresholds so the list stays selective.

Another is self-fulfilling intervention. If you only contact flagged customers, you cannot tell whether the model predicted the outcome or your outreach changed it. A small holdout group, handled normally, helps you measure whether the retention effort works at all.

Lastly, be careful with signals that could introduce unfair treatment, and with sensitive personal data. If customers are individuals and not businesses, privacy rules around profiling may apply. Check with a qualified professional where regulations are relevant.

A hypothetical walk-through

A software team with a few hundred business customers builds a first list in a spreadsheet exported from the CRM. They flag accounts whose renewal falls within ninety days and that show at least two of: no logged contact in two months, falling usage, an open escalated ticket, or a changed main contact. That yields about twenty accounts. The customer success lead assigns each to a manager, who contacts them within the week.

Three of the twenty turn out to be planning to leave over a missing integration, which the team could not fix in time but now understands. Two renew after training sessions. The rest were fine. The team keeps the notes, adjusts the rules, and goes again the next month.

Frequently Asked Questions

Do we need machine learning for churn prediction? No. Start with a rules list. Machine learning can help once you have clean historical outcomes and enough examples, but rules often capture most of the practical value.

How much history do we need? Enough renewals and cancellations to see patterns, which varies by business. If you have few churn events, rely on rules and qualitative review.

How often should the list refresh? Match it to how fast signals change and how fast your team can act. Weekly is common for many teams; refreshing daily when no one acts daily adds noise.

Who should own the list? Typically customer success or account management, with sales operations or analytics maintaining the data and logic.

What if the CRM does not hold usage data? Bring in an export or connection from product analytics and billing. A churn list built only from sales notes will miss the strongest signals.

Next Step

Write your churn definition in one sentence, pick three signals you can already measure, and build a first list of the ten accounts that matter most this quarter. Assign each an owner and a date, and record what happens.