AdAI
Definition

Churn Prediction is the use of machine learning to forecast which customers are about to leave. Trained on the patterns that preceded past cancellations (usage drops, support complaints, payment issues, lower engagement), a churn prediction model assigns each current customer a risk score. The customers at the top of that list get retention attention before they walk.

Key Takeaways

  • Churn prediction estimates which customers are likely to leave in the near future, so retention effort can be aimed at the right accounts before they cancel.
  • Inputs are usually usage data, support history, payment patterns, and lifecycle stage. The model learns which combinations preceded actual past churn, then flags current customers showing those patterns.
  • The economic case is strong. Bain's long-standing research finds a 5% increase in retention can lift profits by 25 to 95% depending on industry, because retained customers cost less, buy more, and refer others.
  • Most SMBs do not need to build a custom model. HubSpot, Vitally, ChurnZero, Gainsight, and Salesforce CRM Analytics all include some form of customer health or churn scoring. Acting consistently on a built-in score usually beats building from scratch.
  • Models typically hit 70 to 85% accuracy in production. The bottleneck is rarely the model. It is whether the team actually contacts the flagged accounts and intervenes fast enough.
25-95%
profit increase from a 5% increase in customer retention, according to Fred Reichheld's Bain research, with the higher end in industries with high customer lifetime value
Source: Fred Reichheld, Prescription for Cutting Costs, Bain & Company; The Loyalty Effect, 1996
5-25x
cost of acquiring a new customer compared to retaining an existing one, across most industries Reichheld and successor researchers have measured
Source: Fred Reichheld, Bain & Company research, widely cited 1990s-present
70-85%
typical accuracy range for production churn prediction models, depending on data quality and business type
Source: Industry benchmarks aggregated from Vitally, ChurnZero, HubSpot Service Hub product documentation, 2024-2025

In Simple Terms

Customers rarely leave overnight. There is usually a build-up: fewer logins, less engagement, a complaint that did not get resolved, a missed payment, a quieter Slack channel. By the time a cancellation hits, the warning signs have been visible for weeks or months. The problem is that no human has time to watch every customer for those signs.

Churn prediction is a model that watches for you. It is trained on the patterns of customers who actually left in the past, so it learns which combinations of behaviour precede a cancellation. Then it scores every current customer against those patterns. The output is a ranked list: this one is 87% likely to churn in the next 90 days; this one is 12%.

For an SMB, the value of that list comes from acting on it. A weekly check-in call from customer success. A targeted retention offer. An owner reaching out personally to the top three at-risk accounts. The model only matters as an input to a retention process that actually runs.

What Goes Into a Churn Prediction Model

The most common inputs across SMB businesses.

Engagement. Logins per week. Active features used. Time since last meaningful action. For a B2B SaaS, sometimes broken down to "active power users in the account" rather than just any login. Engagement signals are usually the strongest single predictor of churn.

Support history. Open tickets, time-to-resolution, whether issues were escalated, NPS or CSAT scores, complaint sentiment. A customer who has filed three angry tickets in the last month is in a different state than a customer who has filed none.

Payment and account changes. Failed payments, downgrades, contract renegotiations, reductions in seat count. Account changes in the wrong direction are direct signals.

Lifecycle stage and account characteristics. How long they have been a customer. Plan tier. Industry, company size (B2B). Onboarding completion (huge signal for new customers, fading for established ones).

External signals, where available. For B2B SaaS, news about the customer (layoffs, acquisitions, leadership changes) can predict churn even when usage looks fine.

Built-In vs Custom: What to Use

Use built-in scoring when

Your CRM, customer success tool, or subscription platform already has a health score. HubSpot Service Hub, Vitally, ChurnZero, Gainsight, Salesforce CRM Analytics, ChartMogul, ProfitWell, and many others. Turn it on, calibrate against your own past churn, build a process around acting on it. Most SMBs never need to go further than this.

Consider a custom model when

Built-in scores keep flagging the wrong accounts. Your business has long sales cycles or complex multi-product accounts the generic models do not understand well. You have your own data sources (usage logs, custom events) the off-the-shelf tools cannot see. In those cases, a custom model with OpenAI fine-tuning, Google Vertex AutoML, or a data person using Python tools (XGBoost is the typical workhorse) is usually a few weeks of work and produces meaningful lift.

“A 5% increase in customer retention correlates with at least a 25% increase in profit.”

Fred Reichheld, Bain Fellow, creator of the Net Promoter SystemPrescription for Cutting Costs, Bain & Company

Frequently Asked Questions

What inputs do churn prediction models use?
Usually a mix of usage data and account history. Login frequency, feature usage, support ticket volume, payment history, plan changes, time since last activity, and customer lifecycle stage are the most common. For a B2B SaaS, additions like account size, integrations active, and key user activity also feed in. The model learns which combinations of these features tend to precede a cancellation, then flags current customers showing those patterns.
How accurate is churn prediction in practice?
Typical production models hit 70 to 85% accuracy, depending on the business and how clean the data is. Higher when the signals are strong (a customer who has not logged in for 30 days is almost always at risk). Lower for surprise churn (a customer who looked engaged and then left because of a competing offer or a reorganisation on their side). Accuracy is rarely the bottleneck. Acting on the prediction is.
Do I need a custom model, or do my existing tools do this?
Many tools now have built-in churn or health scoring. HubSpot Service Hub, Vitally, ChurnZero, Gainsight, and Salesforce CRM Analytics all include some version. For most SMBs, turning on the built-in score and acting on it consistently is worth more than a custom model. Custom builds make sense when the built-in tool does not have enough signal from your specific use case (long sales cycles, complex products, multi-product accounts).
What should I do when a customer is flagged as high churn risk?
The intervention depends on the business. For B2B SaaS, a check-in call from customer success usually works. For consumer subscriptions, a targeted offer (discount, pause option, upgrade incentive) catches many. For services businesses, a personal call from the owner or account lead. The pattern across all of them: act fast (within days, not weeks), make it personal, and ask what is changing in the customer's situation rather than only pitching.
Will reducing churn really make that much difference?
Yes, and the math is more dramatic than most owners expect. Bain & Company's long-running research has found that a 5% increase in customer retention can lift profits by 25 to 95% depending on industry. The reason: retained customers cost less to serve, buy more over time, and refer others, while acquiring replacements is far more expensive.

Related Glossary Terms & Resources

Join 5,000+ SMB owners getting weekly AI agent insights

Subscribe Free