How it works
How our Churn Risk and Upsell Scores work
A model that learned from tens of thousands of SaaS users, tuned to your own data, scores every account daily and tells you why. Here is what it reads, how it learns, how accurate it is and where its limits are.
In short
Customerscore.io gives every account two scores: a Churn Risk Score from 1 to 5 and an Upsell Score from 1 to 3. Both come from a machine learning model, not from rules someone wrote by hand. The model learns which patterns came before churn and before an upsell, scores every account against those patterns and against its own history every day, and explains each score with drivers and brakes in plain language.
Two scores, one model
One score tells you who is about to leave, the other who is ready to grow. Both are refreshed daily.
Churn Risk Score: 1 to 5
1 means healthy, 5 means high risk. A score of 4 or 5 means the account needs attention now. Each step up the scale is a real step up in churn: in our data, customers at 5 churned 31 times more often than customers at 1.
Upsell Score: 1 to 3
1 means the lowest upsell potential, 3 means the account is most ready to grow. It is learned the same way, from the patterns that came before customers expanded, so your team knows which conversations to start.
How the model works, step by step
No weights to set and no thresholds to guess. The model finds what matters in your base and shows its reasoning.
- 1
It joins your customer data
Billing, product usage, CRM and support data for every account, mapped to one customer. It also reads signals about people: who still uses the product, whether the champion is still active, whether a key contact changed role or left.
- 2
It learns from customers who left and customers who stayed
The model compares the two groups and finds the patterns that came before a cancellation, and the ones that came before an upsell. It starts from what it learned on tens of thousands of SaaS users, then is tuned to your own history. A company with several products gets a separate score for each product, because each product is used differently.
- 3
It scores every account daily
Each account is compared with those patterns and with its own past behavior, so an expected seasonal dip is not flagged as risk, while the same drop in an account that never slows down is.
- 4
It takes a deep look every two weeks
Besides the daily scoring, the model goes through each customer's full history every two weeks and learns from the latest churn, so it keeps up as your product and customer base change.
- 5
It explains every score
Every score comes with drivers, the signals that push it up, and brakes, the signals that hold it back. Your team sees why an account is at risk, or ready to grow, before the first call.
What a score looks like
An illustrative account, not a real customer. This is the view a CSM gets before the call.
Example account
Northwind Analytics
Churn Risk
4
Upsell
1
Drivers: what pushes the risk up
- Usage dropped 60% in 30 days
- The champion has not logged in for 5 weeks
- Two seats removed at the last invoice
Brakes: what holds it back
- Annual contract with 9 months left
- Support tickets resolved quickly
- Core feature still used every week
What data it needs
You can start with billing alone and add sources as you go. Each source you add gives the model more to learn from.
| Data source | Needed? | What it adds |
|---|---|---|
| Billing | Required | Payments, plan and seat changes, renewals. Enough on its own for a first score. |
| Product usage | Recommended | Active users, usage trend, feature adoption. Makes the score clearly more accurate. |
| CRM | Optional | Renewal dates, ARR, owners, contacts and their roles. |
| Support and conversations | Optional | Tickets, escalations, meetings and emails. |
To calibrate the model to your base, it needs at least 2 months of history, ideally 100+ active and 40+ churned customers. Your data is processed in the EU; Customerscore.io is ISO 27001 certified and GDPR compliant.
How accurate it is
We checked the scores we gave B2B SaaS customers in 2026 against what happened in the following 60 days.
AI Churn Risk Score
Churn within 60 days, lowest vs. highest score
31 times more churn at score 5 than at score 1.
Rule-based health score
Churn within 60 days, best vs. worst band
12 times more churn in the worst band than in the best.
Both scores put a similar share of customers into their extreme groups. Methodology, every score level and the limits of the analysis are in the churn prediction accuracy study.
Where its limits are
No model is right about every account. These are the cases where you should know what the score can and cannot do.
Few churned customers means a less sharp score
The model learns your base from the customers who left. A young company or one with very little churn starts from the pre-trained model, and the score gets sharper as its own history grows.
It can only read what it can see
A budget cut mentioned in a meeting that nobody logged, or a product with no usage tracking, stays invisible. Connecting usage and CRM data closes most of that gap.
New customers have little history
In the first weeks an account has little behavior to compare against, so early scores lean more on patterns from similar customers than on the account itself.
It scores daily, not by the minute
A score reflects the latest daily data. For a churn decision that builds over weeks this is enough; it is not an alarm for something that happened an hour ago.
A score is a prediction, not a verdict
In our data about 1 in 6 customers at Churn Risk 5 churned within 60 days. Most did not, often because someone acted. The score tells your team where to look; the team decides what to do.
FAQs
No. Every score comes with its drivers, the signals that push the risk up, and its brakes, the signals that hold it back, written in plain language such as "Usage dropped 60% in 30 days". Your team sees why an account scored the way it did before the first call.
No, and that is on purpose. The model is pre-trained on tens of thousands of SaaS users, then tuned to your own data and history. Each account is scored against its own usage, billing and history, so the score reflects your customers, not an average.
Every account is scored daily against the latest data. On top of that, the model runs a deep-dive through each customer's full history every two weeks and learns from the latest churn.
Billing data alone is enough for a first score. To calibrate the model to your base, it needs at least 2 months of history, ideally 100+ active and 40+ churned customers. With fewer churned customers the score still works, it is just less sharp until your own history grows.
Yes. You can run a rule-based score with your own metrics and weights alongside the predictive Churn Risk Score and Upsell Score. When the two disagree, the explanation is often the most useful part.
No. The AI does the watching, your team acts. A rising score can send a Slack alert or start a playbook that creates tasks for the account's CS manager, and accounts that do not need a human touch can get automated emails. People stay in control.
The team behind the scoring engine sets it up with you on your data, and most teams are live in 1 to 3 weeks. Customerscore.io is EU-based and EU-hosted, ISO 27001 certified and GDPR compliant.
Keep reading
AI customer health score software
Both scores for every account, in the tools your team already uses.
Read moreChurn prediction
Catch silent churn while there is still time to change the outcome.
Read moreUpsell Score
Find the accounts that are ready to grow, with the reason why.
Read moreCustomer health score guide
Metrics, examples and how to build a health score yourself.
Read moreSee the scores and their reasons
We will walk you through the Churn Risk and Upsell Scores, the drivers and brakes behind them, and what setup on your data looks like. Most teams are live in 1 to 3 weeks.
Book a demo