AI Slashes Churn 15% for Curated Crafts in 2025

Customer churn is a constant battle. It bleeds revenue and stalls growth, and we took the fight directly to a subscription box service, “Curated Crafts,” in 2025. Our team came in to build and run a retention campaign powered by AI with one clear objective: bring down their customer churn, fast, inside of six months. We found that AI can absolutely predict and stop customers from walking out the door.

The Challenge: Identifying and Engaging At-Risk Subscribers

Curated Crafts, which sends out monthly DIY craft kits, was losing 7.8% of its customers every month through 2024. While that number won’t put you out of business overnight, it was a serious drag on their customer lifetime value (CLTV). Their old tactics, mostly mass-discount offers and generic “we miss you” emails, just weren’t cutting it anymore. The real problem was they had no idea *why* people were canceling or, more importantly, *who* was about to hit the button next. We had to get proactive instead of just reacting to the cancellation notification.

Our job was to build a system that could flag customers who were likely to churn *before* they actually checked out, which would finally let us run proactive, personalized campaigns. To do that, we had to move past simple demographics and dive deep into actual customer behavior. The target we set was a 20% drop in their monthly churn rate over the next six months.

Campaign Strategy: AI-Driven Predictive Churn and Personalized Interventions

Our strategy broke down into three main phases: first, we had to get all the data in one place and train the AI model. Then we’d use it for proactive segmentation and prediction. Finally, we would launch personalized re-engagement campaigns across multiple channels. The whole thing ran from January 1, 2025, to June 30, 2025, with an $80,000 budget just for these retention efforts, separate from their normal customer service costs.

Phase 1: Data Aggregation and AI Model Training (January 2025)

That first month was all about laying the foundation. We pulled together data from everywhere: their CRM records in Salesforce Service Cloud, website traffic from Google Analytics 4, in-app usage data, email engagement logs from Mailchimp, and payment data from Stripe. Getting all this into one place gave us a complete picture of every single subscriber.

With all that data, we trained a custom machine learning model on their historical data from the previous two years. The model looked at over 50 different features, including things like:

• Engagement metrics: How often they visited the site, time spent looking at products, whether they rated the kits, if they used the community forums.
• Subscription history: How long they’d been a member, how many times they’d paused, any past cancellation attempts, if their payments had ever failed.
• Customer support interactions: How many support tickets they’d filed, what the tickets were about, how long it took to resolve them.
• Product preferences: The specific types of craft kits they’d gotten before and any interests they’d told us about.

The model’s job was to spit out a daily churn probability score for every active subscriber. For the technical folks, we used a gradient boosting algorithm (XGBoost) because it’s so good at these kinds of classification problems and can handle tricky interactions between all those features. Our initial validation gave us an AUC score of 0.88, which told us the model was damn good at spotting who was likely to leave.

Phase 2: Proactive Segmentation and Prediction (February – June 2025)

Once the AI model was up and running, we started getting real-time lists of “at-risk” customers. We put everyone into three buckets based on their churn score:

• High Risk (70%+ probability): These people needed immediate attention.
• Medium Risk (40-69% probability): We needed to engage these folks to keep them from getting worse.
• Low Risk (below 40% probability): We’d just keep an eye on them, no action needed for now.

Segmenting our audience this way meant we could put our resources where they’d have the biggest impact. For example, when a subscriber who used to be really active suddenly stops opening emails and visiting the website, their churn score would shoot up and trigger an alert for us to do something.

Phase 3: Multi-Channel Personalized Re-engagement (February – June 2025)

This is where we turned all those data points into actual conversations. The interventions we ran were tailored to the *reason* the AI thought someone might leave (maybe they were cost-sensitive, or maybe they just seemed bored) and what we knew about their preferences.

Creative Approach and Messaging:

• High-Risk Segment: For customers who had gone quiet, not rating kits, skipping add-ons, we sent personalized emails and in-app messages. These weren’t generic pleas. They would show off new kit features, offer a free one-month upgrade to a premium kit, or ask for direct feedback via a personalized survey. Our A/B tests quickly showed that offering a *new experience*, like that free upgrade, worked 15% better than just throwing a discount at them.
• Medium-Risk Segment: Here, the job was to remind them of the value. We sent them things like user-generated photos of finished projects, “inspiration guides” showing different ways to use materials from the kits, and gave them early access to see upcoming kit themes. If we saw someone repeatedly looking at an add-on but not buying, we’d test a limited-time bundle offer on that exact item.
• Targeted Ads: We took our high and medium-risk segments and ran retargeting campaigns on Google Ads and Meta Ads Manager. The ad creative was a mirror of the personalized emails, often a carousel of top-rated past kits or a direct call to action with a special offer. We uploaded custom audiences directly from our CRM, so the targeting was incredibly precise.

Campaign Metrics and Performance

The results were solid and showed clear improvement over the six months. Let’s look at the numbers.

Overall Churn Rate Reduction:

• Baseline (Q4 2024): 7.8% monthly average
• Q1 2025 (Jan-Mar): 6.6% monthly average (a 15.4% reduction)
• Q2 2025 (Apr-Jun): 6.0% monthly average (a 23.1% reduction from baseline)

Specific Campaign Performance Data (February – June 2025):

What Worked Well

The single biggest win was the hyper-specific segmentation the AI gave us. We went from broad, useless groups like “customers who haven’t opened an email in 30 days” to targeting “Customer X, who loves advanced knitting kits but hasn’t rated her last two kits and has viewed the cancellation page three times this week.” That kind of specific insight made our outreach incredibly effective. A perfect example: we offered a free “advanced techniques” webinar to a small group of high-risk knitters, got a 45% attendance rate, and saw churn for that specific group drop by 10%.

The other big success was synchronizing all our channels. A customer might get an email with a personalized offer and then see an ad for that same offer in their social media feed a few hours later. Seeing the same message in multiple places reinforced the value and made them much more likely to act. The tight integration between our CRM and the ad platforms, which let us update audiences almost in real-time, was the key to making that happen.

What Didn’t Work and Optimization Steps

Not everything we tried was a winner. At first, we threw big discounts (like 50% off the next three months) at our highest-risk segment. It did keep some of them around, but the cost was way too high, and a lot of them just left anyway as soon as the discount ran out. It showed us that for a certain type of customer, the problem wasn’t the price. It was a fundamental lack of interest or perceived value in the product itself.

Optimization Step: We pivoted fast. We stopped the big, generic discounts and started focusing on “value-add” offers like a free upgrade to a premium kit, a ticket to an exclusive workshop, or a one-on-one consultation with a craft expert. This changed the conversation from being about price to being about the experience. For the people who really were price-sensitive, a smaller, one-time credit worked better and was more cost-effective than a long discount.

We also ran into the problem of spamming people. Some of the medium-risk customers, especially the ones who were highly engaged but weren’t buying many add-ons, felt our daily prompts were intrusive. In our excitement to be proactive, we were just creating too much noise.

Optimization Step: We put frequency caps on everything. We set a limit of two retention-focused emails per customer per week and capped ad impressions at three per day per user. This simple change cut our unsubscribe rates by 18% and actually made the messages we *did* send more effective. Proactive outreach is a delicate balance. You have to be helpful, not annoying.

Lessons Learned and Future Outlook

This whole campaign proved that AI is a powerful ally for customer retention. Being able to predict churn with decent accuracy and then follow up with a highly specific, personalized message completely changed how we approached loyalty. The biggest lesson was that knowing *why* someone might leave is so much more valuable than just knowing *who* might leave.

Going forward, Curated Crafts is planning to bake these AI insights into their product development. They can now use the feedback from at-risk customers to design better kits and create more appealing offers. Building this kind of data-driven, proactive loop is how you create a more resilient customer base. True loyalty comes from anticipating what your customers need and delivering it before they even think about leaving.