SolutionsCustomer AnalyticsProDataAI Lens Services Case StudiesAbout Insights AI CheckContact
Get in Touch
Customer Analytics · Churn Prediction

Not just who is leaving. Why.

A probability of leaving for every customer, the reasons behind it, and groups a campaign can actually be built for. Refreshed every night, in your own cloud.

A score ranks your customers. It does not tell your team what to put in the email. This module delivers both, and the control group that says what it was worth.

Get in touch Watch the 1-minute video
In one minute

Churn Prediction in one minute

Why churn is noticed too late, where new customers really disappear, and why two customers with the same probability need opposite actions.

Voice and music are AI-generated. Customers, reasons and groups shown are illustrative. Plays via YouTube. Clicking play loads the video from YouTube (Google), which may set cookies. See our privacy policy.
Read the transcript

Most companies notice churn after ninety quiet days. By then the decision is weeks old, and the money is gone.

The expensive part is the start: a new customer buys once and is gone inside four weeks, the acquisition already paid.

So every customer carries a probability, recalculated each night: a watchlist, those at risk, and those to save now.

A score only ranks people. These two carry the same probability for opposite reasons: one never came back after the first order, the other is waiting on an open service case. Send both a discount, and only one answers.

Customers who share a reason are grouped, each with the action that fits. Two groups get nothing: those who were never leaving, and those who were never profitable.

Every action runs against a control group, so you report retained margin, not a response rate.

Churn Prediction, from Customer Analytics by ProDataAI. Know who is leaving, and why. Get in touch at prodata.ai.

Where churn is decided

Most of the loss happens before anyone calls it churn

Retention programmes usually start when a customer has been quiet for ninety days. By then the decision is weeks old and the money has gone somewhere else. Two things actually move the number: reaching existing customers while the pattern is still changing, and the first weeks of a new customer, where most of what you paid to acquire them is lost again.

  • Late churn, the kind everyone measures

    Orders stop, a contract runs out, a subscription lapses, a price letter comes back as a cancellation. It is in the monthly report, and by the time it is in the report there is little left to do about it.

  • Early churn, the kind nobody budgets for

    A new customer buys once, or signs up, and is gone again inside about four weeks. The acquisition is paid for, the first weeks decided the rest, and the customer never appeared in a churn report because they were never counted as a customer.

Your market

Churn does not arrive the same way in every business

The model is the same. What it watches, and how long you have to act, is not. These are the three patterns we meet most often, and the assessment says which one you are actually in.

  • Retail and e-commerce

    What it looks like

    Silence, and nobody announces it. The second order does not come, and before that the customer stops browsing, comparing and filling baskets.

    When you can still act

    The weeks right after the first order. Most second orders that ever happen arrive early, so a win-back at ninety days is a rebuy campaign and not a save.

    What works

    Help the research along while it is still running, then one second-order offer with a deadline. For the customers who were bought with a discount, nothing.

  • Subscription and telecom

    What it looks like

    Change rather than level: usage falling against that customer's own normal, service contacts stacking up, a ticket still open, a bill that jumped, a payment that failed.

    When you can still act

    The first weeks after activation, when the first bill and the first fault decide the relationship, and then again as the contract end comes into view.

    What works

    Fix the onboarding problem while it is still small. Near the contract end, rank by the revenue actually at risk, not by the score alone, and put a person on the customers worth a person.

  • Energy and utilities

    What it looks like

    An event, not a drift. A price letter gives the customer a right to cancel on the spot, and switching portals and switching services make acting on it a single click.

    When you can still act

    Before the letter goes out, not after. Once it is in the post, the whole book is exposed on the same day.

    What works

    Differentiate the increase by predicted risk instead of sending one percentage to everyone, and keep the price guarantee and the contract anniversary on the calendar as the other two moments.

What you get

Four things, every night, in the systems your teams already use

There is no new dashboard to log into. The model writes where the work happens.

  • A probability per customer

    Not an index out of a hundred that nobody can interpret: the probability that this customer leaves inside a window you choose, 90 days in retail, the renewal on a contract. What counts is the change against that customer's own normal, not a league table of heavy and light buyers. New customers are scored from their first week, on a model of their own.

  • The reasons behind each risk

    Not a number on its own: the handful of things that pushed this customer up, in the order of how much each one contributed.

  • Groups that share a reason

    Customers whose reasons look alike, gathered into a handful of named groups with a size and an action that fits all of them. A campaign manager can work with a group. Nobody can work with ten thousand individual scores.

  • Three bands, three budgets

    Watchlist, At risk and Save now. Each band carries the actions it allows and the money it may spend, so a campaign manager can work straight from it. You decide where the lines sit.

It arrives as a table in your warehouse, lists in your CRM and a daily file for your campaign tools. Nothing new to buy.

Score plus reason

A score ranks your customers. It does not tell your team what to send.

Two customers can carry the same probability for opposite reasons. One is drifting because nothing has arrived since the first order. The other because a complaint is still open. The same discount goes to both, and only one of them was ever going to answer a discount. So every risk comes with its reasons, and customers who share a reason are grouped, each group with the action that fits it.

  • First order, then silence

    Act
    What drives it

    No second order, no visits after the first week, arrived through a price comparison or a discount.

    What we do

    One second-order offer with a deadline, then stop. If that does not land, this customer was bought and not won.

  • Valuable and drifting

    Act
    What drives it

    Order frequency falling for weeks, value well above average, still reading the email.

    What we do

    A person, not a campaign. Service or sales calls, with something to offer that is not a discount.

  • Something went wrong

    Act
    What drives it

    An open service case, a late delivery, a complaint, a failed payment.

    What we do

    Close the case first. An offer on top of an unresolved complaint makes it worse, and the reason code is the only way the campaign tool knows.

  • Quiet by nature

    No spend
    What drives it

    Buys twice a year, always has, nothing in the pattern has changed.

    What we do

    Leave alone. A reactivation campaign here spends money on customers who were never leaving.

  • Gone, and not worth buying back

    No spend
    What drives it

    One discounted order, no margin on it, every signal pointing at the next discount somewhere else.

    What we do

    No spend. Recognise them at acquisition instead, so the next thousand like them cost less.

Illustration: example groups. Yours come out of your own data, carry your own names and are usually five to eight.

Two of these five groups get no money at all. That is the part of churn work that pays for itself first, and the reason this module sits next to Coupon Optimisation and Customer Lifetime Value: who is at risk, who is worth keeping, and what an incentive actually changes.

The first weeks

A customer who arrived last week needs a model of their own

In the first weeks you have almost nothing to work with: how they arrived, how large the first step was, and what they did around it. It is not much, and it is enough to separate a customer who will be back from one who will not. A customer with two years of history is a different question and gets a different model. In retail the clock runs on the second order; in a contract business it runs on the first bill, the first fault and then the contract end.

Where new customers disappear
Illustration · retention over time

Illustration. The shape repeats across retail, subscription and energy: steep at the start, flat later. Where the steep part ends, and what pushes customers off it, is different in every company, and measuring that on your own customers is the first thing the assessment does.

  1. Week 1

    How they arrived, what the first order was, and what happened around it: what they compared, what they read, how the delivery went.

    A band before the welcome message goes out, so the first thing you send is not the same for everyone.

  2. Weeks 2 and 3

    Whether they are still researching with you, browsing, comparing, reading reviews, filling a basket, or whether it went quiet after the first parcel.

    The journey branches. Someone still deciding gets help deciding. Someone who went quiet gets a reason to come back instead of a brand newsletter.

  3. Week 4

    The second order has not come, and this is the window in which most second orders that ever happen arrive.

    The last cheap moment: one offer with a deadline, or a service call if something went wrong. After this, winning them back costs a different budget.

  4. From week six

    Enough behaviour for the main model to take over.

    The customer moves onto the existing-customer risk and stays there, scored every night.

This needs history: a few months of new customers, including the ones who never came back. If your data does not carry that yet, you hear it in the assessment instead of getting a model built on nothing.

The agent layer

What runs as a job, and where an agent earns its place

Most of this should not be an agent. Scoring every customer at night, assigning the bands and pushing the lists into your CRM is an ordinary scheduled job, and a job is cheaper to run and easier to audit. Judgement is the part worth an agent.

  • The job

    Scores, reasons, groups and lists, every night, the same way every time. Nothing here decides anything, it predicts.

  • The caseworker

    One valuable customer at a time. It opens a case, reads that customer's own history, picks the action the reason calls for and spends inside a budget it cannot exceed. Above the threshold you set, a person approves.

  • The auditor

    Holds the control group, compares what was spent against what was retained and switches off a play that does not pay, including one it proposed itself.

Every agent runs under three rules

  • A budget it cannot exceed.
  • A control group for every kind of action.
  • A log naming the customer, the model version, the inputs and the approver.

The scores, the reasons and the groups run today. The caseworker is live for customer lifetime value and for segments; for churn it is built with the first client who wants it. Until then the actions run as campaigns your team approves, on the same scores and the same reasons.

Proof

The number comes before the campaign

Retention is the easiest place in marketing to claim a win, because a customer who was never leaving still counts as saved. So every action runs against a control group from the first week, and the report is in retained margin, not in response rate.

  • A control group for every action

    A share of each band gets nothing. It is the only way to know whether the campaign kept anybody, and it costs a few percent of the upside to find out.

  • Measured in money

    Retained margin against what the retention spend cost, by band and by group. Response rate is not the number.

  • The customers who stay anyway

    Part of every saved group would have stayed without the offer. The control group separates them, and they stop receiving incentives.

  • Monthly, on one page

    What each band and each group cost, what it retained and what was switched off. Written so the number survives a finance review.

Put to work

In the tools you already use

  • CRM and newsletter

    Lists per band and per group, refreshed nightly, with the reason on the record so the message can match it.

  • Service and sales

    The at-risk customers worth a person's time, on a call list, with what is driving the risk and the open case if there is one.

  • Shop and app

    An at-risk customer sees something different from a loyal one, without your team building a second site.

  • Google and Meta

    Retention budget away from the customers who stay anyway, and look-alikes built on the ones who stayed.

  • Coupons

    An incentive only where the reason says an incentive is what is missing.

  • Reporting

    One monthly view of risk, action and retained value, for marketing and for the board.

How we start

From your own churn to the first measured save, within weeks

A fixed-fee assessment, a pilot on time and material with an estimate up front, then a monthly run fee. You never commit to more than the next step.

  1. 1 2 to 3 weeks

    Assessment

    On your own data: how much of your churn is early and how much is late, what it costs you a year, the first risks and reasons, and the groups we would expect to find. If the data will not carry a model yet, you hear that in week two.

  2. 2 8 to 12 weeks

    Pilot

    One part of your customer base, two or three actions, control groups from day one, lists running into your CRM. At the end you know what was retained, measured.

  3. 3 monthly

    Run

    Nightly risks, groups kept current, actions added and switched off on what the control groups say.

Trust and data

Your customers stay yours

  • Runs in your cloud

    Risks and reasons are calculated where your data already is. Your customer data is not sent to us to be processed.

  • No training on your data

    Your customers do not train anyone's model, ours included.

  • Every risk explainable

    A risk, its reasons, its inputs and the model version that produced it stay on record. When a customer or an auditor asks why, the answer exists.

  • Your team decides

    The bands, the budgets and the actions are yours. Above a threshold you set, nothing goes out unapproved.

More on how we handle data: Trust & Data

FAQ

Frequently asked questions

How far ahead does it see?
Far enough to act, and how far that is depends on how your customers buy. In retail it is the next few weeks; on a contract it is the renewal, and in energy it is the day the price letter goes out. The window is set in the assessment, because a prediction nobody has time to act on is worth nothing.
We sell contracts, not baskets. Does this still fit?
Yes, and the shape of the work changes. In a contract business the risk is driven by usage falling against that customer's own normal, by service contacts and unresolved tickets, by billing surprises and by the contract end coming into view, and the save is ranked by the revenue actually at risk rather than by the score alone.
We already have a churn score in our CRM.
Good, then the question is what your team does with it. Most scores arrive without a reason and without groups, so they end up driving one discount for everyone. The reasons and the groups can be built on the score you already have.
Does this just mean sending more discounts?
No, and it usually means sending fewer. Two groups in five normally get no money at all, either because they were not leaving or because they were never profitable. The reason decides the action, and often the action is a phone call, a fixed service case or nothing.
What data do you need?
Orders or contracts, the customer record, activity if you have it (email, app, shop, service contacts) and the acquisition channel. History matters more than tooling: about a year, and it has to include the customers who actually left.
How accurate is it?
On your customers, nobody knows until it has been measured, and a number from another company's data would not help you. The assessment measures it on your own history and says what it finds, including whether the early window is predictable at all in your business.
Can we see why one single customer is at risk?
Yes, per customer: the reasons in the order of how much each one contributed. That is the part service and sales actually use.
What about GDPR?
The model runs in your own environment under a data processing agreement (Art. 28 GDPR), on data you already hold for the customer relationship. No special categories of personal data, nothing bought in, and a customer's risk is explainable on request.
Who runs the campaigns?
Your team, in your tools. We deliver the risks, the reasons, the groups and the actions that fit them, and we sit with your campaign manager for the first round.
What does it cost to start?
The assessment has a fixed fee. The pilot runs on time and material with an estimate at the end of the assessment, because preparing the data is usually the longest part. Once live, there is a monthly fee.

Find out how much of your churn is already decided.

Thirty minutes on a call: tell us how customers come in, how they leave and what you send them today. We will tell you where your churn actually sits, early or late, and what it would take to see it before it happens.

Get in touch
  • Free first call
  • Data stays in your environment
  • You own the model and the scores

or email ai@prodata.ai