{"slug": "matas-used-ai-to-find-subscribers-going-quiet-could-it-make-them-stay", "title": "Matas Used AI to Find Subscribers Going Quiet. Could It Make Them Stay?", "summary": "Matas reported a 29 percentage point retention lift after using Subsets' explainable AI to target dormant Club Matas Plus subscribers, according to a customer story published by Subsets, the Copenhagen vendor that supplied the prediction and experimentation software. The Danish retailer identified members who showed moderate activity over the prior 60 days followed by 30 days of inactivity, sent a personalised re-engagement sequence to a treatment group, and withheld it from a control group, with the groups differing on retention, website activity and orders after 30 days. Matas Head of Loyalty Julie Marie Hauerberg has separately confirmed the retailer uses Subsets to identify early signs of disengagement among Club Matas Plus members.", "body_md": "7 min read\n\nMatas held a campaign back from quiet subscribers. The vendor reports a 29 percentage point lift.\n\nEvery subscription business has members who have gone quiet. Matas had that retention problem hiding inside a large loyalty base. Some Club Matas Plus subscribers had stopped behaving like active members, and the Danish retailer had no easy way to tell whom to prioritise or whether a rescue campaign changed their odds of staying.\n\nThe team tested a narrow answer. An explainable AI system, meaning one that shows the behaviour behind each prediction, identified subscribers who had once been active and had since gone dormant. Matas sent a personalised re-engagement sequence to the treatment group and held it back from a control group. Thirty days later, the groups differed on retention, website activity and orders.\n\nThe details come from a [customer story published by Subsets](https://www.subsets.com/customer-stories/how-matas-increased-retention-by-29-with-subsets), the Copenhagen company that supplied the prediction and experimentation software. That makes it a vendor-published account. It still names the customer, the audience logic, the comparison, the duration and the business outcomes. Matas Head of Loyalty [Julie Marie Hauerberg](https://dk.linkedin.com/in/julie-marie-hauerberg-1301353) has separately confirmed that the retailer uses Subsets to identify early signs of disengagement among Club Matas Plus members.\n\n## Matas could not see who needed a retention intervention\n\n[Matas Group](https://matasgroup.com/group/what-we-do/) runs the Matas and KICKS beauty businesses across Denmark, Sweden, Norway and Finland. Club Matas Plus is the paid extension of its Danish loyalty programme, with benefits including extra points and free delivery.\n\nSubsets describes the paid service as having a six-figure subscriber base. That scale created two linked constraints for the loyalty team. It lacked a clear way to identify the subscribers most worth contacting. It also lacked an efficient way to test whether a retention treatment changed behaviour, given that many recipients would have stayed anyway.\n\nThe first problem calls for prediction. The second calls for a counterfactual, which asks what would have happened without the treatment. Matas combined them in one AI retention experiment and treated the model score as a starting point for proving commercial value.\n\n## The model found auditable signs of dormancy\n\nSubsets says its platform uses traditional machine learning and interpretable AI to surface lifecycle audiences. For this test, it identified people described as dormant subscribers who used to be active. The explanation attached to the segment was concrete. Members had shown moderate activity during the previous 60 days, followed by 30 days of inactivity.\n\nThat explanation gave the commercial team something more useful than a black-box churn score. The team could inspect the behaviour behind the classification and decide whether a re-engagement message made sense. AI handled predictive churn selection, while the team kept ownership of the treatment.\n\nMatas then built a personalised sequence with what the case study calls valuable content. Subsets keeps the messages, channel mix, incentive and number of touches private. Its platform passed only the treatment group into the loyalty flow through Matas’s existing customer relationship system. The control group received no flow.\n\n## The control group separated targeting from impact\n\nThis is the part worth copying. A predictive model can find people who look likely to churn. Only a comparison can show whether a campaign prevented the churn. Control group testing supplies that comparison by withholding the treatment from comparable eligible subscribers. LinkedIn applied similar logic when it [asked which people its ads could actually persuade](https://industrycontents.com/linkedin-causal-incremental-marketing-targeting-experiment/).\n\nAfter 30 days, Subsets reports that the treated group had a retention rate 29 percentage points higher than the control group. It also reports 39 percent higher web engagement and 20 percent more orders. These are business metrics a loyalty or growth team owns, which sets them apart from model accuracy measures.\n\nThe wording matters. The source reports a 29 percentage point lift, which differs from a 29 percent relative increase. It withholds the retention rates in both arms, so nobody can calculate the relative lift.\n\nSubsets says its platform flagged the result as statistically significant and recommended turning the journey into an always-on flow. The public case leaves out the sample size, allocation ratio, confidence interval, significance threshold and exact definition of retention. It also leaves open whether Matas accepted the recommendation for this specific journey.\n\n## Matas extended the method beyond one flow\n\nHauerberg says Matas launched ten predictive retention flows during the first six months and could compare its AI-driven flows with control groups throughout. The customer quote says those flows produced higher retention than their controls, although no combined effect size is disclosed.\n\nThat follow-on tells us more than a claim that the first campaign simply worked. It suggests the team made the measurement pattern reusable across lifecycle automation. AI refreshed the eligible audience, the customer system delivered the treatment and an untreated group stayed available to estimate incremental impact.\n\nThe experiment also shows why selection and treatment belong in separate parts of the data model. A churn model decides who is eligible. In a rigorous implementation, random assignment decides who receives the intervention. Outcome measurement decides whether the intervention deserves to continue. Blending those jobs into one opaque score would make it hard to tell whether better subscription retention came from sharper targeting, a better message or a biased comparison.\n\n## The public evidence stops before the profit line\n\nThe 20 percent order increase gives the experiment a revenue-adjacent result. The case leaves out order value, discounts, campaign cost and contribution margin, so nobody can say whether the flow increased profit or simply generated more transactions.\n\nThe control also lacks description. The case calls it a control group yet leaves open whether assignment was random, whether the groups were balanced before the test and whether subscribers could enter other overlapping campaigns. Those gaps limit the strength of any causal claim.\n\nA commercial conflict deserves visibility. Subsets supplied the software, measured the test and published the result. The named Matas practitioner confirms the predictive-churn use case and control-group practice, but nobody has independently audited the detailed figures.\n\n## Copy the measurement pattern before the model\n\nA subscription business can try the most transferable lever without a machine learning team. It can define one auditable risk state with existing event data, such as activity in the prior 60 days followed by 30 days of inactivity. It can then split eligible subscribers into treatment and control groups before sending a re-engagement flow.\n\nThe minimum scoreboard should include retention after a full renewal window, orders or revenue per eligible subscriber, and unsubscribe or complaint rates. Report outcomes across everyone assigned to each group, including people who never opened or clicked. Keep the control intact long enough to catch delayed cancellations.\n\nOnly after that loop works does AI add leverage. A model can refresh the risk segment as behaviour changes, explain why each cohort was selected and help the team test more than one lifecycle moment. The model sharpens prioritisation. The holdout establishes whether the action changed the result.\n\nMatas’s test is useful because the growth mechanism is visible. Predictive churn found a plausible audience, a human team designed the intervention and a control group tested the business effect. Faster lifecycle automation matters only if that separation survives once the journey goes always on.\n\n### Sources\n\n- [Matas Group](https://matasgroup.com/group/what-we-do/)\n- [Subsets customer story on the Matas experiment](https://www.subsets.com/customer-stories/how-matas-increased-retention-by-29-with-subsets)\n- [Julie Marie Hauerberg on LinkedIn](https://dk.linkedin.com/in/julie-marie-hauerberg-1301353)\n- [Subsets company information](https://www.subsets.com/careers)\n\nCompare the evidence and operating lessons from more AI experiments in our [Growth Signal Index](https://industrycontents.com/ic-lab/growth-signal-index/), or browse benchmark tooling in the [Industry Contents Lab](https://industrycontents.com/ic-lab/).", "url": "https://wpnews.pro/news/matas-used-ai-to-find-subscribers-going-quiet-could-it-make-them-stay", "canonical_source": "https://industrycontents.com/ai-retention-experiment-matas/", "published_at": "2026-10-06 04:00:00+00:00", "updated_at": "2026-10-06 04:16:09.723454+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "ai-products"], "entities": ["Matas", "Subsets", "Club Matas Plus", "Julie Marie Hauerberg", "Matas Group", "KICKS"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/matas-used-ai-to-find-subscribers-going-quiet-could-it-make-them-stay", "markdown": "https://wpnews.pro/news/matas-used-ai-to-find-subscribers-going-quiet-could-it-make-them-stay.md", "text": "https://wpnews.pro/news/matas-used-ai-to-find-subscribers-going-quiet-could-it-make-them-stay.txt", "jsonld": "https://wpnews.pro/news/matas-used-ai-to-find-subscribers-going-quiet-could-it-make-them-stay.jsonld"}}