Solutions / Churn prediction

Churn prediction model

Every customer is scored weekly on how likely they are to leave in the next 30 to 90 days, with the reasons behind the score and an owner attached — so the list can be worked instead of read.

Retention Customer service ML
100% of base
scored, every week
30–90 days
of warning before the exit
Reasons given
not just a risk number
Who it is for

Subscription and recurring-revenue businesses with a few hundred accounts or more, where churn is currently discovered at renewal.

Short answer

Your entire customer base is scored every week for the probability of leaving in the next 30 to 90 days, using your own usage, billing and support history. Each name on the list carries the factors that drove the score and the person who owns the account, because a risk score nobody acts on is a report, not a retention programme.

The problem

You find out someone is leaving on the day they tell you, which is thirty days after they decided.

01

Churn is discovered at renewal

The decision was made weeks earlier — usually after a bad month of support, a champion leaving, or usage quietly dropping to nothing. By renewal day the conversation is a formality.

02

The signals exist, in five systems

Logins fell in the product, invoices started being paid late, three tickets escalated. No single system shows all three, so nobody joins them into a customer who is about to go.

03

Attention goes to whoever complains loudest

The quiet accounts leave. Success teams work the tickets in front of them, and the customer who simply stopped logging in never generates a ticket to work.

How it works

From trigger to result, step by step.

01

Learn from the churn you already had

We train on your own history — who left, when, and what their data looked like in the months before. If you do not have enough churn history yet, we say so and start with rules; a model on forty examples is a story, not a model.

02

Join the signals into one view

Product usage, billing and payment behaviour, support volume and tone, contact changes, and commercial history. The joining is most of the work and it is where the accuracy actually comes from.

03

Score every account, every week

The whole base is scored on the same schedule, so a risk that appears on Tuesday is visible on the list that week rather than at the next quarterly review.

04

Explain each score

Every flagged account shows the factors that moved it: usage down 60% over six weeks, two escalations, invoice paid 40 days late. A score without reasons gets ignored by the third week, and it deserves to be.

05

Put the list where the work happens

It lands in your CRM and in the success team's channel, assigned to owners, with a suggested play per risk type. Outcomes come back in — saved, lost, false alarm — and the model is retrained on them.

Before / after

What changes on the ground.

Today, by hand
×Churn is discovered on the renewal call
×Signals sit unjoined across five systems
×Quiet accounts leave without a conversation
×Nobody can say why an account was at risk
With the automation running
›30–90 days of warning before the exit
›One weekly list, covering 100% of the base
›Each name has reasons and an owner
›Saves and losses feed back into the model
What you get

Delivered, not demoed.

›A churn model trained on your own history, with honest accuracy reporting
›Signals joined from product, billing, support and CRM into one account view
›A weekly scored list with per-account reasons and assigned owners
›Delivery into your CRM and chat, with outcome feedback and retraining
›Documentation and a handover session — the model is yours
Built with

We build in your stack rather than moving you onto ours. The list below is what this solution most often connects to — other systems are a scoping question, not a blocker.

Python scikit-learn XGBoost PostgreSQL BigQuery dbt HubSpot Slack Airflow
Time to production6–8 weeks
Build priceFixed quote
First stepFree mini-audit
Honest limits

When this is not the right solution.

·If you have fewer than a couple of hundred customers, or under a year of churn history, there is not enough signal to learn from. Rules built with your success team will beat a model here, and cost less.
·If nobody owns the follow-up, the list will be read and not worked. We would rather scope the retention play with you first than deliver a dashboard that quietly dies in month two.
·If churn is driven by one known cause — a missing feature, a pricing tier, an onboarding gap — fixing that is worth more than predicting it. The model will only keep pointing at the same thing.

Questions we get about this one

That depends on your data, and anyone quoting a number before seeing it is guessing. We report performance against a baseline — how much better than your current guess — on a held-out period, and we show it before you rely on it.

Bring us the process that hurts.

The mini-audit is free: we take your version of this process apart and tell you plainly whether automating it pays. If it does, you get a scope and a fixed price.