How a DTC Brand Recovered 22% of Lapsed Customers Using AI-Powered Win-Back Campaigns

The Brand, the Problem, and What Was at Stake

The brand in this case study is a mid-market DTC skincare company — let's call them Lumis — that had been operating for six years and had built a customer list of roughly 180,000 buyers. Their hero products sat in the $45–$90 range, with an average order value of $72 and an average purchase frequency of 2.4 orders per year for active customers.

By early 2025, their marketing team had identified a growing problem: their lapsed segment — customers who had made at least one purchase but had placed no orders in 12 months or more — had ballooned to 61,000 contacts, representing 34% of their total customer base. At their average order value, even recovering 10% of that segment would generate over $430,000 in revenue from contacts who were otherwise generating zero.

The previous approach to win-back had been a manual, batch-and-blast three-email sequence sent to every lapsed customer at the same time, with the same creative, and a flat 20% discount. That campaign had historically recovered around 5.8% of lapsed customers — a number the team suspected was far below what was possible. Worse, the blanket discounting was training their most price-sensitive customers to wait for coupons, eroding margin across the board.

"We were essentially treating 61,000 different people as one person. The lapsed customer who bought twice in six months and the one who bought fourteen times over three years were getting the exact same email. It was leaving an enormous amount of money on the table."

The stakes were significant. Customer acquisition costs had risen sharply in the preceding 18 months, making reactivation economics far more attractive than acquisition. Industry observations consistently suggest that winning back a lapsed customer costs 3–5x less than acquiring a new one with equivalent lifetime value. For Lumis, fixing their win-back program was effectively a lower-cost growth lever than paid media.

Strategy and Approach: What They Decided — and What They Deliberately Skipped

The team's core decision was to replace intuition-based segmentation with predictive churn prevention e-commerce modeling. Rather than defining "lapsed" as a binary on/off switch at the 12-month mark, they wanted to score every customer on a spectrum of recoverability — essentially predicting who was likely to respond to a win-back offer and who was effectively gone for good.

This distinction mattered enormously for budget allocation. Sending high-value win-back sequences (including potential free-product offers or aggressive discounts) to customers with a near-zero recovery probability was pure waste. Conversely, sending a weak, low-effort email to a customer with a high recovery probability meant leaving reactivation on the table.

The strategy rested on four pillars:

• Predictive scoring: Use an AI model trained on purchase history, email engagement, time-since-last-order, category affinity, and seasonal patterns to assign each lapsed contact a recovery probability score from 0–100.
• Tiered segmentation: Divide lapsed customers into three tiers (high, medium, and low recovery probability) and design distinct campaign tracks for each.
• AI-generated personalization: Use generative AI to produce subject lines, body copy, and product recommendations tailored to each customer's purchase history and browsing behavior — not just their name.
• Multi-channel sequencing: Layer SMS into the sequence for high-probability customers only, rather than blasting SMS to the entire lapsed list (a move that historically drove unsubscribes).

What they deliberately did not do: they did not immediately offer a discount to everyone. Discounting was reserved for the medium-probability tier and specific high-probability sub-segments where purchase history indicated price sensitivity. High-probability customers with a history of full-price purchasing were targeted with value and newness messaging rather than coupons.

Implementation: Steps, Timeline, and Tools

The full implementation ran over 14 weeks, broken into three distinct phases. The team used a combination of their existing email service provider (ESP), a predictive analytics layer bolted on via API, and their SMS platform — all coordinated through a customer data platform (CDP) that served as the single source of truth for scoring data.

The predictive model was trained on 36 months of purchase and engagement data. Key signals included: days since last purchase, number of lifetime orders, average order value, product category purchased, email open rate over the prior 12 months, and whether the customer had ever responded to a previous promotional email. The model was validated against a holdout set of customers whose behavior was already known, achieving a predictive accuracy of 81% on recovery probability.

Segmentation results from scoring: 9,200 contacts landed in the high-probability tier (score 70–100), 21,400 in the medium tier (score 40–69), and 30,400 in the low-probability tier (score 0–39). The low-probability tier received a single "we miss you" email only — no SMS, no aggressive offers, and no further follow-up if they didn't open. This decision alone saved significant SMS costs and protected sender reputation.

For a deeper look at how the underlying tooling fits together, the brand's approach mirrors what's described in this AI retention marketing stack guide, which covers CDP integration, predictive scoring layers, and lifecycle automation architecture in detail.

The AI personalization engine generated unique subject lines based on each customer's most recently purchased product category and their historical engagement patterns. In testing, AI-generated subject lines outperformed the control (human-written) subject lines by an average of 18% on open rate across 12 tested variants.

Results: Before and After Metrics

The campaign ran its full 90-day active phase from launch. Results were tracked against the previous batch-and-blast win-back benchmark of 5.8% reactivation across all lapsed contacts.

In total, 13,461 lapsed customers were reactivated. Critically, 46% of those reactivated customers placed a second order within 60 days of their win-back purchase, suggesting that the personalized re-engagement approach also improved the quality of recovery — not just the quantity. The reactivated cohort's 90-day retention rate was 31% higher than customers reactivated through the previous generic campaign.

Key Learnings: What Worked, What Failed, and What Surprised Them

What worked beyond expectations: The tiered scoring approach was the single highest-leverage decision. Concentrating resources on the high- and medium-probability tiers — and dramatically scaling back effort on the low-probability tier — improved both results and economics simultaneously. The team had expected segmentation to improve results; they had not expected it to nearly eliminate unsubscribes from the lapsed list.

What failed: The SMS component underperformed in the medium-probability tier. Open rates were strong (SMS open rates were above 90%, as expected), but conversion rates from SMS were only marginally better than email alone for this tier, and the incremental revenue did not justify the channel cost for customers scoring between 40 and 55. In future campaigns, SMS in the medium tier will be restricted to customers scoring 55 and above.

What surprised them most: High-probability customers who received value-and-newness messaging (no discount) converted at a 38.9% rate, compared to 44.1% for those who received a discount. The gap was smaller than expected — and the margin difference was substantial. Many practitioners assume that lapsed customers universally require a price incentive to return; this data suggests that for high-affinity customers, product relevance and personalization can do most of the heavy lifting.

"The customers most likely to come back didn't need a discount — they needed to feel like we remembered them. The AI-generated recommendations referencing their actual purchase history did more work than any coupon code we've ever sent."

The unexpected operational finding: Updating recovery scores in real time (weekly model refreshes during the campaign) made a measurable difference. Customers who improved their score mid-campaign due to email engagement were moved to higher-tier sequences, and those contacts converted at a 29% rate — significantly above the original tier they were assigned to. Static scoring would have missed this entirely.

How to Replicate This: An Actionable Checklist

The following steps translate this case study into a replicable process for DTC brands with an established customer base and at least 12 months of purchase history data.

• Audit your lapsed segment first. Define "lapsed" based on your purchase cycle — for a brand with a 90-day replenishment window, lapsed might be 180 days. For Lumis, 12 months made sense. Get an accurate count and segment by lifetime order count.
• Build or license a predictive churn model. You don't need to build from scratch — several CDPs and ESP integrations offer predictive scoring out of the box. The minimum viable signals are: days since last order, lifetime order count, average order value, and prior email engagement rate.
• Tier your lapsed list into at least three buckets. High (top 15%), medium (next 35%), low (remaining 50%) is a useful starting split. Adjust based on your score distribution.
• Design distinct content tracks — not just distinct offers. High-probability customers often respond to relevance and product storytelling. Reserve discounts for mid-tier and price-sensitive sub-segments identified in the data.
• Use AI for subject line generation and product recommendations. Even basic AI tooling can generate and A/B test subject line variants at a scale no human copywriter can match. Product recommendations should pull from actual purchase history, not just bestsellers.
• Limit SMS to high-probability tiers initially. Expand to mid-tier only after validating that conversion lift justifies channel cost in your specific economics.
• Refresh your scoring at least weekly during the campaign. Customers who engage with early emails should be eligible for tier upgrades. Static scoring leaves reactivation opportunity on the table.
• Track second-order rate as your primary quality metric. Reactivation rate tells you who came back; second-order rate tells you who actually re-engaged with the brand. Optimize for the latter.

This approach scales up or down depending on list size and tech stack sophistication. A brand with 10,000 lapsed customers can run a version of this with native tools in most modern ESPs. The underlying logic — score, tier, personalize, sequence — remains constant regardless of list size or platform.

Frequently Asked Questions