You spot a customer who's leaving in time by watching changes in how they buy before they disappear entirely: they buy less than their average, space out their orders, stop ordering products that used to be fixtures. Those signals are in data you already have —your orders or invoices— long before the customer says anything. Building an alert that watches them doesn't need an expensive program: it needs deciding what's “odd” for each customer and having something look at it each week.
Honesty first: this doesn't predict the future or prevent every loss. A customer may leave for a reason your data can't see (a change of owner, a cheaper rival). What it does do is warn you about the ones who leave slowly —the majority— while you can still call and ask.
Why don't you notice until it's too late?
Because a silent loss makes no noise. There's no angry call or complaint: there's an absence, and absences don't appear in your inbox. The customer who ordered weekly moves to every fortnight, then monthly, and since each order on its own looks normal, nobody adds up the trend. By the time someone notices “we haven't heard from so-and-so in a while”, so-and-so is gone.
What signals warn that a customer is leaving?
None is conclusive alone; together they are:
- Frequency drops. They ordered weekly and now every three weeks. The change of rhythm says more than any single order.
- Amount drops. They still order, but half as much. Sometimes it's the step before leaving entirely.
- The basket changes. They stop ordering the main product and only order the extras. Usually it means they buy the main thing from someone else.
- The silence stretches. More time than usual for that customer with no order. “Usual” differs for each one: that's why you measure per customer, not in general.
How do you build the alert, step by step?
- Gather the order history. The date, customer and amount of each order is enough to start. It's in your invoicing program or a spreadsheet.
- Work out what's normal for each customer. Their average frequency between orders and their average amount. A weekly customer's normal isn't a quarterly one's: each has a pattern.
- Define what's odd for each. For example: more than twice their average gap without ordering, or last month's amount below half. Simple, explainable numbers, not magic.
- Have something look weekly. An automatic review that runs through customers and flags those crossing their threshold. Without it, the task gets done “when there's a moment”, which is never.
- Produce a short at-risk list. Not a hundred-page report: five or ten names worth calling this week. The list has to fit into a morning's phone calls.
- Call and ask. The alert retains no one; the call does. Sometimes it's a problem you didn't know; sometimes just that they feel looked after. The tool points; you save the relationship.
Simple rules or do you need AI?
To start, simple rules. Comparing each customer with their own average and warning when they stray covers most cases, is cheap and —most importantly— you understand it: you know why the alert fired. AI comes later, if you want to refine, when you have many customers and complicated patterns. But building a model before trying the rules is starting the house from the roof.
| Approach | When to use it |
|---|---|
| Simple rules (per-customer average) | Almost always to start. Cheap, explainable, enough. |
| A churn-probability model (AI) | Many customers and complex patterns, when rules fall short. |
When is it NOT worth it?
If you have few customers and know them by name, you already have the best possible alert: you remember them. Building a system for twenty customers you see every week is using a cannon on flies. This pays off when you have so many that they don't fit in anyone's head, which is why the slow leavers go unnoticed. It also loses meaning if your sales are one-off (no recurrence to watch).
What mistakes do people usually make?
- Measuring everyone the same. “A month without ordering” is alarming for a weekly customer and irrelevant for a quarterly one.
- Producing a report nobody reads. If the alert doesn't end in a call, it's an ornament. Make it a short, actionable list.
- Waiting for certainty. By the time you're sure they're leaving, they've left. The alert is for early suspicion, not the death certificate.
- Trusting it all to the machine. The system detects; retaining is a human conversation. Without the call, the best alert is useless.
Frequently asked questions
What data do I need to spot customers leaving?
Very little: the date, customer and amount of each order or invoice. With that you compute each customer's frequency and average amount and detect when they stray from their pattern. It's in your invoicing program or a spreadsheet; you need nothing else to start.
Do simple rules work or do I need AI?
To start, simple rules: compare each customer with their own average and warn when they move away. It's cheap, you understand it and it covers most cases. AI makes sense later, with many customers and complex patterns, but not before trying the simple version.
Does this stop customers leaving?
Not on its own. The alert warns you in time; retaining is a conversation. It mainly catches the slow leavers —the majority— while you can still call and ask. Those who leave suddenly for a reason your data can't see, it won't catch; no system does.
I have few customers, do I need it?
Probably not. If you know them by name and remember them, you already have the best alert. This pays off when you have so many customers that the ones buying less bit by bit go unnoticed.
The customer says it with their orders before their mouth
Almost no loss announces itself: the customer orders less, more spaced out, changes what they buy, and one day disappears. Each order on its own looks normal, so nobody adds up the trend. The signal is in your data —frequency and amount per customer— before the customer says a word.
Building the alert doesn't need AI to start: it needs comparing each customer with their own average and producing a short, actionable list each week. The system detects; retaining is a timely human call. Without that call, the best alert is an ornament.
Could you say which customers are buying less than three months ago?
If you have many customers, some are probably leaving slowly without you noticing. With your order data, a simple alert can surface each week the ones worth calling. Let's talk.
See the data analysis service Let's talk about your customers