Churn early warning for mid-sized companies: four signals you can spot in the data you already have

Customers don't leave overnight

The call comes on a Tuesday. A long-standing customer says that from next quarter it will be ordering elsewhere. To management, it feels like a bolt from the blue. The relationship was considered stable.

Look more closely afterwards, and the departure was no surprise. Orders had been coming in less often for weeks. The basket had become smaller, and some product groups were missing entirely. The buyer had not replied to two emails, and the last visit from field sales was months ago. All of this was in the ERP, in the CRM and in the inbox. Nobody had brought it together, though, and nobody was responsible for looking at it.

Churn shows up in the data long before it reaches management.

The principle behind this is backed by research. A study in the European Journal of Operational Research (2005) shows this using the example of a consumer goods retailer: even among loyal customers without a contract, creeping partial defection can be predicted from their past buying behaviour. The best predictors were three simple measures: how long ago the last purchase was, how often the customer buys and how much revenue that generates.

A field experiment in B2B wholesale (Industrial Marketing Management, 2022) goes one step further. It looked at customers without a framework agreement, who could therefore leave at any time without giving notice. A model predicted the churn risk for each customer. According to the study's abstract, targeting the customers with the highest risk reduced churn significantly more than contacting customers at random, and revenue developed positively.

So the signals are in your data, and it pays to react to them. You don't need new software or an analytics team for this. You need someone who reads the signals, and a fixed time to do it.

Why SaaS health scores don't fit mid-sized companies

If you research health scores, you almost always end up with software vendors. There, the health score is a value per customer fed by usage data: how often does the customer log in, which features do they use? If usage drops, the customer counts as at risk.

A wholesaler, a machine builder or a service provider with project business does not have this data. Its customers don't log in anywhere. They order, commission work, complain and call. On top of that, many mid-sized companies work without subscriptions or fixed terms. A customer then does not cancel. They simply order less. That does not show up on any cancellation list.

You still have the data you need. Your inventory management system already stores the timing, frequency and revenue of purchases. Where a software vendor has logins, you have order and contact data.

Four signals can be read from it: order rhythm, basket, service contacts and the time since the last real contact. One basic rule applies to all four: the yardstick is the past behaviour of this particular customer, including their seasonal patterns, rather than a flat threshold.

Signal 1: Order frequency changes

The first signal is in the order date. Your ERP stores which customer ordered when for every order. That covers two of the three measures from the retail study: how long ago the last purchase was and how often the customer buys.

Read off each customer's usual rhythm

A spreadsheet export is enough.

1. Pull the customer number and order date of all orders from the last 12 to 24 months, limited to your key customers.
2. For each customer, calculate the days between two consecutive orders.
3. Record the typical interval. Use the median, meaning the middle value of the sorted intervals, rather than the average. A single long break, for example because of a company shutdown, distorts the average heavily and the median hardly at all.
4. Note the longest break from periods in which the relationship was running normally.

At the end you have one line per customer: "Usually orders every X days, longest normal break Y days."

Put the current break next to it

Count the days since the last order and compare them with this rhythm. A simplified example: customer A orders roughly every ten days, customer B once a quarter. A 25-day break is a clear signal for A and completely normal for B. A flat rule such as "30 days without an order means risk" fires too late for A and wrongly for B.

Three test questions help:

• Is the current break longer than the longest normal break so far? That does not mean the customer is leaving yet, but it is a reason to call.
• Are the intervals getting gradually longer over several orders? A creeping trend often says more than a single long break. In B2B relationships the pattern is common: weekly orders become monthly, then quarterly.
• Is there a seasonal explanation? Factory holidays, project cycles or annual budgets explain some breaks. A comparison with the same period last year shows it.

Set a starting rule, check after a few weeks whether the flagged customers really were unusual, and refine it. Recalculate the rhythm about once a quarter.

One stumbling block in advance: if a customer has several customer numbers, for example one per site, or sometimes orders through a head office, you first have to merge these numbers. Otherwise an active customer looks like one who is leaving.

Signal 2: The basket shrinks or shifts

A customer can keep ordering at the usual pace and still be on the way out. Then the churn shows in the content of the orders. Three patterns:

• Quantities fall. The same items, smaller quantities. It is rare for demand to have fallen by the same amount. More often, the customer is covering part of it elsewhere.
• The range narrows. Five product groups before, three today. Whole product categories drop away, usually without a word.
• The high-margin items disappear. What remains is standard goods and urgent orders. The customer is evidently sourcing exactly the items you earn money on somewhere else.

All three point to partial defection: the customer stays, and part of their business goes to the competition. With revenue per order, the third measure from the retail study comes into play here. Range breadth and margin share work just as well as early signals.

Why does the quiet decline come before the cancellation? Without a framework agreement, nobody has to cancel. And even with a contract, a buyer rarely switches all at once. More typically, they test a new supplier with part of the range, compare price and delivery reliability and then shift volume piece by piece. The cancellation, if it comes at all, is the formality at the end of a decision that was made long ago.

How to check it: for each key customer, evaluate three metrics, namely revenue per order, the number of product groups ordered and the share of high-margin items in revenue. Compare them with the same period last year for the same customer. Whatever deviates clearly and lastingly from the customer's own pattern stands out.

The signal is valuable because it comes early. As long as the customer is still ordering, there is something to win back.

Signal 3: Service contacts deteriorate

The first two signals come from order data. Service shows something else, namely the state of the relationship. The signal comes from two directions.

It gets louder. Complaints pile up, queries become sharper, and the customer asks for management more often. A single complaint is normal business. It becomes notable when complaints cluster within a short time or the same problem keeps coming back. A customer who escalates usually still wants to save the relationship. That is your window.

It goes quiet. This signal is trickier. A customer who used to call regularly, ask questions and sometimes grumble no longer gets in touch. Internally, that is often taken as a good sign. It is often the opposite: a customer who no longer complains sometimes no longer expects anything to change, and is already looking for alternatives.

The data is usually there, just scattered. In the ticketing system you will find the number and type of cases per customer, repeats and escalations. If there is no ticketing system, the service inbox is the most important source; a rough count per month and customer is enough. Then there are field sales reports with cancelled appointments or notes such as "purchasing holding back" or "competitor on site". These soft indications often get lost because they exist only in one employee's head.

The data is rarely the obstacle. Service sees the complaints, field sales the cancelled appointments, sales the orders. The full picture only emerges when these three views come together in one place.

Signal 4: Customers who stop getting in touch, and the contact clock

The fourth signal is the simplest, and yet it is the one most often overlooked. A key customer stops getting in touch: no call, no query, no appointment. Nobody notices, because everyone assumes a colleague is currently in contact.

The contact clock helps here. For each key customer, it measures how many days have passed since the last real contact. Nothing else. The idea carries the measure "time since the last purchase" from the retail study over from purchases to the relationship as a whole.

What counts as real contact

Only what shows a response from the customer or involves them directly counts: a conversation by phone or in person, an email the customer writes or answers themselves, a meeting that took place, an order.

Newsletters, automatic invoices, shipping confirmations and calls that go unanswered do not count. Many CRM systems are misleading here: the "last activity" field also updates with automatic emails. A customer then looks active although they have been silent for weeks.

The order counts, because a customer who orders regularly has not gone silent.

How to set up the clock

To start, a spreadsheet with your 20 most important customers is enough:

The order date comes from the ERP via a weekly export. The responsible person enters conversations, emails and meetings themselves on the same day; that takes seconds. If you work with a CRM, create a separate date field "last real contact" instead of trusting the automatic activity field.

Once a week, a second person looks at the list, for example the head of sales. This second look is necessary, because nobody likes to report of their own accord that their customer has gone quiet.

When the clock goes off

A customer who orders and calls every week is notable after three quiet weeks. A customer with two projects a year is not. Look at the history of the last year and set a starting rule, for example: flag once the gap reaches twice the usual interval. After a few weeks you will see whether the rule fires too often or too rarely.

Precision matters less here. The clock shows which important customer has just gone quiet.

From signal to a simple health score

Four signals only help if someone can read them at a glance. A health score is exactly this condensation: for each customer, an assessment of how much at risk the relationship currently is. A spreadsheet with one row per key customer is enough, in Excel or as an additional field in the CRM.

Four columns, one traffic light

Each column can have a third status: "unknown", if you have no reliable data for this signal. This gives you the traffic light:

• Green: no signal is notable, and at most one is unknown.
• Amber: exactly one signal is notable, or two or more are unknown.
• Red: two or more independent signals are notable.

"Independent" is the decisive point here. If two columns measure the same gap, they count as one signal. This affects two cases. The contact clock goes off and the last real contact was the last order: then the clock and the order rhythm overlap. Or the service contacts only stand out because they have gone quiet: then they overlap with the clock. That way no customer turns red just because a single fact was counted twice.

This rule is a starting point and has not been calibrated. After two or three months you will see whether the traffic light goes off too often or too rarely, and adjust.

Why four signals are enough

A customer can go quiet in four ways: they order less often, they order less, the relationship deteriorates in service, or they fall silent. Each column covers one of these routes. Further signals such as payment behaviour, a change of contact person or an approaching contract end come from the six early warning signals with a clear action behind them.

The wholesale experiment works with a statistical model. For getting started, the traffic light is enough, and it has one advantage: everyone on the team can see why a customer is amber.

Where the traffic light reaches its limits

The traffic light is only as good as the data behind it. If field sales does not document conversations, the contact clock goes off although the relationship is intact. Duplicate customer numbers make each individual strand look weaker than the whole. Scattered complaints leave signal 3 incomplete.

That is why the "unknown" status exists. Missing data does not count as green. A customer with two unknown signals deserves more attention than one with four cleanly green ones.

Who acts when the traffic light turns amber

A traffic light on its own does not retain a customer. It shows where you should look. Whether anything follows from it depends on two questions: who is responsible, and when does someone look? If both stay open, an amber customer sits in the spreadsheet for weeks until they are red or no longer ordering. If authority or budget is missing for this, an implementation mandate for the health score programme helps.

In the wholesale field experiment, the lever lay in selecting the customers. The prediction decides whom you call first. It can only have an effect if someone then picks up the phone.

One owner per key customer

Every customer on the list gets exactly one responsible person, named in the spreadsheet. Usually this is field sales or account management, meaning whoever knows the customer best anyway. This person does not have to solve every problem themselves. But they make sure that within one week it is clear what lies behind the amber signal.

Once a week, amber and red only

Set a short weekly meeting in which only amber and red customers are discussed. There are three questions for each one:

1. What has changed: order rhythm, basket, service contacts, contact clock? Or is data missing?
2. Who is taking care of it, and by when?
3. What came out of the last step?

The third question makes the difference. Without it, the meeting turns into a list of good intentions.

The first steps

Amber means: ask before the customer does. First, a call with no sales intent. Ask how things are going and whether anything has changed on the customer's side, such as a new contact person in purchasing. If the call stays evasive or the signals pile up, go and visit; a lot only becomes visible in a face-to-face conversation. If a concrete problem is on the table, such as a complaint, a pricing issue or a competitor's offer, bring in the responsible manager for a clarifying conversation.

The result goes into the same spreadsheet as the traffic light. That way you can see next week whether the customer is recovering or slipping further. If a customer turns red, they belong on management's desk.

Checklist: what you can check this week

For a first early warning, you don't need a project, just an honest look at the current state. Clarify these questions with sales, the sales back office, service and accounting:

1. Who are your 20 most important customers? By revenue, contribution margin or strategic importance. If management and sales do not have the same list in mind, that is already the first finding.
2. What data do you have on these customers? Order data from the ERP, invoice data from accounting, complaints from the ticketing system or inbox, visit notes from the CRM or field sales reports. You don't have to collect anything new, just bring together what already exists.
3. What do the four signals show today? Go through the list once: who is ordering less often than usual, who is missing product groups, where are complaints piling up or has it gone quiet, and when was the last real contact in each case? That is your first contact clock.
4. Does each of these customers have a named person responsible? "Sales" is not enough.
5. Where are you lacking transparency? Note where you got no answer or could only get at the data by hand. Typical cases are separate systems, contacts in individual employees' inboxes and field sales reports that are not stored centrally. These are your blind spots, the "unknown" status in the traffic light.

If you would like to put your results into context, talk to UNIT ONE. In a Reactivation Audit we look together at which signals you can already measure and where a sensible starting point lies.

Which of your 20 most important customers has gone longest without getting in touch, and do you know that today?

Frequently asked questions

How early do customers of mid-sized companies typically show signs of churn? If you trace back the history of customers who left, you usually find the first signals weeks to months before the departure: in less frequent orders, a shrinking basket, changed service contacts or contact that stops.

What is the contact clock and how do I set it up? It shows the days since the last real contact for each key customer, as a column in a spreadsheet or a date field in the CRM. You derive the alert threshold from the customer's previous contact rhythm.

Do I need a new CRM or a BI team for early churn detection? No. To start, an export from the ERP, a spreadsheet and one person who maintains it are enough.

Sources

• Customer base analysis: partial defection of behaviourally loyal clients in a non-contractual FMCG retail setting (2005). European Journal of Operational Research 164(1), pp. 252 to 268. ideas.repec.org
• Proactive customer retention management in a non-contractual B2B setting based on churn prediction with random forests (2022). Industrial Marketing Management 107, pp. 134 to 147. doi.org/10.1016/j.indmarman.2022.09.023