Falling orders
Signals
“Central Gym has cut its orders by 40% in two months and raised two complaints about late deliveries.”
Alert
- Risk
- High
- Main reason
- Delivery complaints
Call from the account manager this week
Use case
AI customer churn prediction
A machine learning model spots which customers are most likely to leave and why, so the sales team can act before losing them, not after they have gone.
- A weekly list of at-risk customers, ranked by probability and value.
- Every alert explains the signals behind it.
- The team decides what to do; the model only flags.
The problem
Most customers do not warn you before they leave
They buy a little less each time, complain once or stop logging in to the portal. The signals are in the data, but nobody has time to look at them customer by customer.
Today
- Churn is discovered when the customer does not renew or stops ordering.
- Account managers focus on the large customers and mid-sized ones slip away.
- Retention efforts arrive late and treat everyone the same.
With AI
- Risk signals show up weeks before the customer leaves.
- Each customer has a probability and a main reason.
- Retention focuses on those with the highest risk and value.
How it works
From customer history to a follow-up call
-
Customer data
Purchases, contracts, complaints, service usage and sales contacts.
3 years · 4,200 customers
-
The model is trained
It learns the patterns of customers who left, using classical machine learning.
Validated against real churn
-
Weekly score
Every active customer gets a churn probability and the factors that weigh most.
High risk: 38 customers
-
Human review
The team acts
The account manager reviews the list, decides the action and logs it in the CRM.
Call, visit or offer
Examples
Three different signals, three different actions
Sample data from a fictitious supplies company.
Silence
Signals
“Riverside Clinic has placed no orders in six weeks, having usually ordered every fortnight.”
Alert
- Risk
- Medium
- Main reason
- Purchase frequency
Reminder with their usual order
Contract ending
Signals
“A maintenance customer has 60 days left on their contract and is using the portal less.”
Alert
- Risk
- Medium
- Main reason
- Upcoming renewal and lower usage
Review meeting before the renewal
Signals the model usually uses
Data already held in the ERP, CRM or support desk
- Purchase frequency and value
- Complaints and claims
- Service or portal usage
- Contract dates
- Sales contacts
- Price changes
Down to earth
What to know before you start
-
The history needs churned customers
The model learns from customers who left. With very few recorded, the first step is to define clearly what counts as churn.
-
It flags, it does not decide
The model gives probabilities, not certainties. Whoever knows the customer decides whether and how to act.
-
The effect is measured
We compare churn among flagged customers with a control group to find out whether retention is working.
Visual guide · Churn prediction
Customers
give warning
before they leave
They rarely say so. Their data does: they buy less, complain or go quiet. A model reads it every week.
01 · Your portfolio, this Monday
60 customers, three colours
- 4 High risk
- 8 Medium risk
- 48 No signals
The sales team does not have to look at all 60: it starts with the 12 in colour.
02 · What gives churn away
Three signals already in your data
Buying less
Orders fall month by month, often after a complaint.
Silence
It used to order every fortnight and has gone quiet for weeks.
Contract ending
A few months left on the contract and less use of the service.
03 · The Monday list
Who, why and what to do
- Central Gym Risk: High · Value: 3/3 Delivery complaints Call from the account manager
- Sunrise Academy Risk: High · Value: 2/3 Price increase Sales visit
- North Workshops Risk: Medium · Value: 3/3 Renewal soon and lower usage Review meeting
- Riverside Clinic Risk: Medium · Value: 1/3 Order frequency Reminder with their usual order
04 · Feet on the ground
It warns,
it does not decide
It gives probabilities, not certainties. Whoever knows the customer decides whether and how to act.
And you measure whether it works:
Flagged customers
The team acts
Control group
No alert
Comparing churn in both groups shows whether retention really works.
Loyalty
See which customers come back and which are fading away
As well as the weekly list, the dashboard shows how retention evolves by customer type, region or product, to spot problems that affect a whole segment rather than one customer.
- Retention by sign-up cohort.
- Segments where churn is rising.
- Effect of retention actions.
Method
First we check the model gets the past right
-
Define churn and review the data
We agree what churn means for your business and check what data is held for each customer.
Churn definition and data assessment
-
Model and validation
We train the model and measure whether it would have spotted, in time, the churn in months it has not seen.
Accuracy report on real churn
-
Pilot with the sales team
The list reaches the CRM every week. We measure churn avoided against a control group.
Weekly list and measurement of the effect
Related service
-
Custom AI development
We design and build artificial intelligence solutions tailored to your processes, data and tools when an off-the-shelf solution does not fit or falls short.
Related training
-
Machine learning with Python course
Real predictive models, from the data to evaluation and interpretation.
- Duration: 24 hours
- Level: Intermediate
-
AI for sales teams and B2B selling course
More time selling and less time preparing, with AI as your sales assistant.
- Duration: 6–8 hours
- Level: Beginner
Do you know which customers you could lose this quarter?
Tell us what kind of customers you have, how they buy and where their data is kept. We will tell you whether there is enough information for a churn prediction model.