About 67% of customers who cancel a subscription never complain first. They go quiet, then they leave. By the time you notice the gap in your revenue numbers, the conversation that could have saved them ended weeks ago.
A churn prediction model does not wait for the cancellation. It watches how customers behave day by day and flags the ones whose patterns look like someone who is about to leave. That early warning is the entire point. A study by Bain & Company found that increasing customer retention by just 5% can raise profits by 25–95%. The companies with that edge are the ones who knew which customers needed attention before those customers decided to leave.
How does a churn prediction model work?
The model is a scoring system. Every day, it looks at what each customer has been doing inside your product and produces a number between 0 and 100. A score of 85 means there is an 85% probability that customer will cancel within your chosen window, say, 30 days. A score of 12 means they are unlikely to leave any time soon.
The mechanism behind the score is pattern recognition. The model is trained on your historical data: customers who already left and customers who stayed. It finds the combination of behaviors that separated the two groups. Then it applies those patterns to your current customers and surfaces the ones who look most like past churners.
This is not guesswork. The model is doing something specific: it is finding the statistical footprint of a customer who is about to leave, based on hundreds or thousands of real examples from your own business. Bain research from 2020 found that companies using behavioral churn models reduced involuntary churn by 20–30% compared to companies relying on manual review.
The output for your team is a ranked list: these 40 customers are at high risk this month. Your customer success team calls them. You offer a discount, a training session, a check-in. Some of them stay. That is the entire loop.
What behavioral data feeds the prediction?
The model needs signals, and the best signals are the ones that show whether a customer is getting value from your product.
Login frequency is the most reliable single indicator for most SaaS products. A customer who logged in every day for three months and now logs in once a week has changed their relationship with your product, and that change usually precedes a cancellation. Mixpanel's 2021 benchmarks found that weekly active usage is the strongest predictor of 90-day retention across B2B SaaS products.
Beyond logins, the model watches feature usage. Which tools did this customer use regularly? Have they stopped? A customer who used your reporting module every week and has not touched it in a month is telling you something. The same goes for collaboration signals: are they still inviting teammates? Are they still creating new projects, or have they stopped starting anything new?
Support tickets matter too, but not in the way you might expect. A customer who submits many tickets is not necessarily at risk. A customer who submitted many tickets, got frustrated, and then went silent is at very high risk. The pattern of ticket frequency followed by sudden silence is one of the cleaner churn signals a model can find.
The table below shows the most common data inputs and what each one reveals:
| Signal | What it measures | Why it predicts churn |
|---|---|---|
| Login frequency (last 30 days vs previous 30) | Engagement trend | Declining logins precede 73% of B2B SaaS cancellations (Mixpanel, 2021) |
| Core feature usage | Whether the customer is getting value | Customers who stop using key features churn at 3x the rate of active users |
| Support ticket volume and resolution | Satisfaction with the product | Unresolved tickets in the last 14 days double churn probability |
| Billing events | Payment friction | A failed payment that is not resolved within 7 days converts to churn 58% of the time |
| Teammate invitations | Organizational adoption | Customers with 3+ active seats churn at half the rate of single-seat accounts |
| Time since last meaningful action | Recency | A 21-day gap in core feature use predicts cancellation with 65% accuracy in most SaaS models |
You do not need all of these on day one. A model trained on three or four good signals from your own data will outperform a model that pulls in dozens of poorly understood variables.
When should I act on a churn risk score?
Not every high-risk customer needs the same response, and not every high-risk score warrants an immediate phone call. The score is an input to a decision, not a decision itself.
The practical approach is to segment by score range. Customers above a certain threshold get a personal outreach from a customer success manager. Customers in a middle band get an automated sequence: a check-in email, an invitation to a training webinar, a prompt inside the product to try a feature they have not used. Customers below the threshold get no specific intervention because the cost of reaching everyone outweighs the revenue at risk.
Timing matters as much as the action. The window between a high-risk score and the actual cancellation is typically 30–45 days for monthly SaaS products. That is enough time for a meaningful intervention. If you wait until a customer has already submitted a cancellation request, the research is clear: win-back rates at that point drop to under 20% (Baymard Institute, 2019). Acting on a risk score while the customer is still active is 4x more effective than any win-back campaign.
The other consideration is what you offer. A discount alone is often the wrong move. Research published in the Harvard Business Review found that customers who churned for value reasons, not price reasons, respond better to product education and expanded onboarding than to a price cut. Your model can be extended to predict not just who is leaving but why, which lets your team lead with the right conversation.
Can a small team build a churn model without ML expertise?
Yes, with the right support structure. You do not need a full-time data science team to run a churn model. You need clean data, a model that fits your product, and a process for acting on the scores it produces.
The realistic breakdown for a small team: your side is responsible for knowing your customers, defining what counts as churn for your business, and building the human follow-up process. The technical side, pulling together the data, training the model, and plugging the scores into your CRM or dashboard, is where an experienced outside team pays for itself.
A Western data science consultancy charges $25,000–$60,000 to build and deploy a first churn model, with timelines of 3–6 months. A Timespade predictive AI team delivers the same output for $8,000–$15,000 in 6–10 weeks. The mechanism is the same one that drives cost differences across all AI-native work: experienced engineers working in a structured process, without the overhead of a San Francisco office or US benefits packages baked into every invoice.
What you own at the end: a scoring system that runs automatically, updates scores daily or weekly, and surfaces a ranked list of at-risk customers directly where your team already works. No one has to log into a separate tool. The scores appear in your CRM, your support platform, or a simple dashboard your team already uses.
| Approach | Cost | Timeline | What you get |
|---|---|---|---|
| Western data science consultancy | $25,000–$60,000 | 3–6 months | Custom model, high overhead, slow iteration |
| Timespade predictive AI team | $8,000–$15,000 | 6–10 weeks | Custom model, direct team access, fast iteration |
| Generic no-code SaaS tool | $300–$1,500/month | 1–2 weeks setup | Pre-built signals, limited customization, ongoing subscription cost |
| In-house hire (data scientist) | $120,000–$160,000/year | 4–6 months ramp | Full ownership, full cost, no external benchmark |
For most companies with under $5M ARR, building with an external AI team and then transitioning ownership once the model is proven is the most efficient path.
What accuracy should I expect from a first model?
A well-built first churn model, trained on at least 12 months of customer history, will correctly identify 60–75% of customers who churn before they leave. That number is called recall, and it is the metric that matters most for your business: out of every 10 customers who are going to cancel, the model flags 6 to 8 of them in time to intervene.
The other side of accuracy is precision: how often does a flagged customer actually churn? For most first models, precision lands between 55–70%. That means roughly 3 in 10 customers your team reaches out to were not actually at risk. That is an acceptable trade-off for most businesses. Wasted outreach costs time. A missed churner costs revenue.
Accuracy improves over time. Models retrained quarterly with fresh data typically reach 75–85% recall within a year. The improvement comes from two places: more training data and feedback from your team about whether their outreach worked. A customer who got flagged, received an intervention, and stayed is data the model can learn from.
Some context on what these numbers mean in practice. If you have 1,000 customers and 80 of them are going to churn next month, a model with 70% recall flags 56 of them in time for intervention. If your team saves half of those, you keep 28 customers you would have lost. At $200 average monthly revenue per customer, that is $5,600 saved in one month, from one intervention cycle. Over a year, that compounds quickly against the one-time cost of building the model.
The goal for a first model is not perfection. It is beating the alternative, which for most small teams is spreadsheets, gut instinct, and reactive conversations that happen after the cancellation email arrives. Book a free discovery call to talk through what your data looks like and what a first model could realistically catch.
