£16M
per year in additional revenue preserved
2
percentage point churn reduction over two years
Multiple zero-uplift campaigns identified and stopped
A/B and geo-holdout test designs in production
Uplift modelling and rigorous test design preserved £16M per year in revenue for a major UK insurer by reducing churn by 2 percentage points over two years. The programme replaced broad segment-based campaigns with individual-level targeting, embedded A/B and geo-holdout measurement into every campaign, and eliminated several long-running campaigns that had zero incremental effect.
The Problem
A large UK insurance provider was running dozens of retention, pricing, and sales campaigns every quarter. Some worked. Many did not. And the business had no reliable way to tell which was which.
The Solution
We scoped, built, and operationalised a suite of analytical capabilities across the Retention, Pricing, Sales, and Customer Experience functions, spanning three areas.
Targeting and personalisation: we built uplift and response models predicting which customers would respond to a specific intervention, trained on features engineered from policy data, contact-centre interactions, and digital customer journeys. Unlike standard response models (which predict who will respond), uplift models estimate the incremental effect at the individual level, distinguishing "persuadables" from customers who would renew regardless. Targeting shifted towards the persuadable segment, where spend actually changes outcomes.
Measurement infrastructure: A/B and geo-holdout test designs were embedded into the campaign workflow so that every campaign produced a clean incremental estimate of its own effect. For digital campaigns, randomised A/B holdouts ran at the customer level. For campaigns with geographic spillover (such as regional pricing changes), geo-holdout designs assigned clusters to treatment and control groups. Each test was specified to detect a meaningful effect size given the expected sample and variance.
Operational dashboarding: lift curves, ROI tracking, and model performance monitors (calibration plots, drift detection) were surfaced in dashboards the campaign and analytics teams use day to day. When calibration diverged beyond a threshold, models were flagged for retraining.
Results and Impact
| Metric | Value |
|---|---|
| Churn reduction | 2 percentage points over two years |
| Revenue preserved | An additional £16M per year |
| Underperforming campaigns removed | Multiple campaigns stopped after measurement showed zero or negative uplift |
| Functions supported | Retention, Pricing, Sales, Customer Experience |
| Data sources integrated | Policy, contact centre, digital journeys, claims, pricing |
| Test methodology | A/B (customer-level) and geo-holdout designs in production |
Key Takeaways
-
Cutting bad campaigns delivered faster results than perfecting good ones. The fastest path to impact was identifying and stopping campaigns with zero or negative measured uplift. Several long-running campaigns had continued on the strength of raw response rates that masked the absence of any incremental effect.
-
Measurement infrastructure is a prerequisite, not a follow-on. Without A/B and geo-holdout designs built into the campaign process from the start, there is no way to learn what works. The test infrastructure paid for itself many times over by preventing continued investment in ineffective campaigns.
-
Agreeing on the definition of success mattered more than the modelling. The single most impactful step was getting stakeholders to align on a shared metric: incremental churn reduction over a defined measurement window. Once the goal was fixed, priorities became obvious.