Losing a good customer rarely happens in one day. Orders get smaller, payments slow down, a key contact stops answering, and by the time the account shows up as lost in a quarterly report, the decision was made months ago.
Customer churn prediction uses the history already in your systems to spot those signs early, so your team can call while there is still something to save. Here is what it takes, what data you need, and the myths that make it sound harder than it is.
Why keeping customers is worth the effort
The economics are well known. As Harvard Business Review summarized, acquiring a new customer is anywhere from five to 25 times more expensive than retaining an existing one, and research by Frederick Reichheld of Bain & Company found that increasing customer retention rates by 5% increases profits by 25% to 95%.
Your own figures will differ. The direction rarely does: a customer who stays is cheaper to serve and usually buys more over time. The hard part is knowing which customers to focus on before they leave.
What customer churn prediction actually produces
A churn model reads past customer behavior and learns which patterns came before customers left. It then scores every current account on how closely it matches those patterns.
The useful output is simple: a short list for your account managers each week, with a risk level and the reasons behind each name. For example: "Order value down for three months in a row, two invoices paid late, no reply to the last two quotes."
The list advises; a person decides whether to call, visit, offer help or do nothing.
First, define what churn means for your business
This sounds obvious and is often skipped. The definition changes everything that follows.
Contract businesses (software, maintenance agreements, subscriptions): churn is a cancellation or a non-renewal. The date is clear.
Repeat-order businesses (distribution, manufacturing, wholesale): customers rarely cancel. They drift. You might define churn as "no order in 90 days for a customer who used to order monthly", or as a fall in spend below a set share of their usual level.
Partial churn: a customer who moves half their volume to a competitor never leaves on paper, but it costs you just as much. Decide whether you want to catch this too.
Agree the definition with sales and finance before any model is built. If they disagree on who counts as lost, they will disagree on the list.
The data you need for churn prediction
You probably already hold most of it. The table shows the usual sources and what each one tells a model.
Data | Where it usually lives | What it can signal |
|---|---|---|
Order history (dates, values, products) | ERP or order system | Falling order size, longer gaps between orders, fewer product lines |
Payments and credit | Finance or accounting system | Slower payments, disputes, credit holds |
Support tickets and complaints | Helpdesk, shared inbox | Rising complaints, unresolved issues, a change in tone |
Contracts and renewal dates | CRM or contract files | Renewals approaching without contact, downgrades |
Product or portal usage | Your software, customer portal | Fewer logins, fewer active users, features abandoned |
Account activity | CRM, email | Unanswered quotes, a key contact leaving, fewer meetings |
Pricing and discounts | ERP, quotes | Pushback on price, requests for bigger discounts |
Two things matter more than the number of sources:
History that includes customers who left. The model learns from past departures, so it needs enough of them, ideally across a couple of years so seasonal patterns show.
One customer ID across systems. If "Acme Ltd" in the ERP is "ACME Limited" in the CRM, someone has to match them first.
Five myths about churn prediction
Myth: You need huge amounts of data
Fact: a few years of orders, invoices and support records from a mid-size firm is often enough to start. What matters is that the history is reasonably clean and includes customers who left.
Myth: You should start with the most advanced model
Fact: simple first. Google's own Rules of Machine Learning open with "Don't be afraid to launch a product without machine learning" and advise keeping the first model simple to get a baseline you can measure against. A clear rule, such as "flag any customer whose 90-day spend fell by a third", may get you a long way before a model adds more.
Myth: High accuracy means the model works
Fact: churn data is lopsided. If 5 of every 100 customers leave in a year, a model that predicts "nobody leaves" is right 95% of the time and useless. Google's guidance on class-imbalanced datasets explains why this needs special handling. Ask instead: of the customers flagged, how many were really at risk, and how many leavers did the list miss?
Myth: The model tells you why customers leave
Fact: it shows which signals came before departures, such as late payments or fewer orders. The real reason, a new competitor or a bad delivery, comes from the conversation your account manager has.
Myth: Prediction reduces churn on its own
Fact: a list nobody acts on changes nothing. Decide in advance who receives the list, what they do for each risk level and how you record the outcome.
Are you ready? A quick checklist
You have agreed, in writing, what counts as a lost customer.
Orders and invoices go back at least two years in a system you can export from.
You can match customers across your ERP, CRM and finance system.
You know roughly how many customers left in each of the last two years.
A named person will own the weekly list and the follow-up.
You have a way to record what happened after each call.
Four or more ticks means you can start. Fewer than that, and the first project is tidying the data, which pays off for every other report you run.
How a churn project usually runs
Agree the definition and the action. Who is lost, who gets the list, what they do.
Pull and match the history. Orders, payments, tickets and contacts, joined by customer.
Build a baseline. Simple rules first, so you know what a model has to beat.
Train and test a model on past data. Check it would have flagged customers who really left, before it touches live accounts.
Run it alongside your team. A few weeks of weekly lists, with account managers saying which flags make sense.
Measure and adjust. Track saved accounts and missed leavers, and tune the thresholds.
Churn warnings are one of the forecasts we build as part of our predictive analytics work, alongside demand and cash forecasts. Each warning shows its reasons, you set the thresholds, and the model reads from your own ERP, CRM and finance records without the data leaving your systems. When the list should also trigger a follow-up task or a draft email for an account manager to approve, that is where AI automation comes in.
Frequently asked questions
How far ahead can churn prediction warn us?
It depends on how quickly your customers' behavior changes before they leave. In repeat-order businesses the warning signs often build over months, which gives a useful lead time. The test on past data tells you what to expect for your own customers.
Do we need a data science team?
No. You need someone who knows your customers to agree the definition, a partner to build and maintain the model, and account managers who act on the list.
Is it only for subscription businesses?
No. Distributors, manufacturers and service firms with repeat customers can use it too. The definition of churn changes, the method stays much the same.
Will customers know they have been scored?
The score is an internal prompt for your team. Any use of personal data should still follow your privacy policy and local data protection law, and a person, not the model, decides what happens next.
If you want to know whether your data is ready for churn prediction, our free 5-day audit looks at one department's records and tells you what a first forecast would need and cost. It starts with an NDA, and the data stays in your systems.
