Predicting churn to protect a compliance SaaS business

A proof-of-concept churn model built for a health and safety compliance SaaS and accreditation business correctly identified 76% of churn instances in test data. Modelling on top of that result showed that contacting just the highest-risk half of customers due for renewal was projected to capture 87% of all churners, giving the business a way to concentrate a limited retention effort on the accounts most likely to actually leave.

The client sells accreditation, certification and SaaS products that help SME customers manage health and safety compliance, and it holds substantial customer data through its subscription model. Its measurement stack, though, stopped at business intelligence dashboards and KPI tracking: useful for showing what had already happened, but no help in deciding who to prioritise ahead of a renewal date. There was no data-driven way to identify at-risk customers or to prioritise renewals outreach. That mattered strategically because the business needed to lift retention on its flagship accreditation product back above a 90% target, and undirected outreach to every renewing customer is costly and dilutes attention away from the accounts that are actually at risk.

QuantSpark split the engagement into two workstreams. The first was a two-week Data Diagnostic: 13 workshops with 24 stakeholders spanning the C-suite, sales, marketing, finance and management information functions, alongside a review of eight internal documents and six datasets. This produced a capabilities and process audit and a prioritised, scoped Opportunity Menu, the shortlist of where predictive and analytical work would add most value. Churn on the flagship accreditation product was the opportunity carried forward into the second workstream.

The second workstream carried that priority into a working proof of concept. The team first hypothesised the likely drivers of churn, a discipline that keeps a model grounded in plausible business mechanisms rather than pure correlation-hunting, then matched each hypothesis against the client's internal datasets to see which could actually be evidenced. Only then was a Random Forest classifier trained, scoring churn risk at the level of the individual renewal rather than at the portfolio level, so that outreach could in principle be targeted customer by customer.

In practice the work ran as a single pipeline rather than two separate projects: diagnostic workshops and document review fed the process audit; the audit fed the scoped Opportunity Menu; the churn opportunity identified on that menu was carried into hypothesis generation and data matching; and that fed the model build and its validation against a held-out test set. The technical core throughout was standard supervised machine learning, a Random Forest classifier, trained on the client's own subscription, renewals and customer data rather than any proprietary or third-party system.

Two figures matter for the value case, and they are not the same kind of figure. The 76% churn-detection rate is a measured result, drawn from testing the model against data the model had not seen during training. The second figure, that contacting the highest-risk half of the renewal base was projected to capture 87% of all churners, is a modelled projection of what targeted outreach could achieve, not a retention uplift that has actually been realised and evidenced in the delivery record. The diagnostic work also surfaced that somewhere between 17% and 33% of churn is effectively unavoidable, customers leaving for reasons outside the business's control, which means the addressable, avoidable share of churn the model is aimed at is roughly 83% of the total.

Put together, the logic for the client is that a small, prioritised slice of the renewal book, rather than the whole book, can be targeted with retention effort and still catch most of the churn that is realistically preventable. That is the return-on-investment argument in outline: the cost of retention outreach, calls, incentives, account-management time, scales with how many customers you contact, while the benefit scales with how much churn is actually prevented among those contacted. A model that concentrates outreach on the highest-risk half while still catching the large majority of preventable churners lets a fixed retention budget go further than blanket coverage would. No pound-figure return has been calculated here: doing so would require assumptions about customer lifetime value, outreach cost and conversion rates that are not present in the source record, so none has been invented for this case study. What is evidenced is a tested predictive capability, not yet a measured business outcome.