Predicting Churn 60 Days Early: AI Customer Intelligence for E-commerce

Industry: E-commerce (Home & Garden Retail) | Company size: 350 employees | Active customers: 280,000
Deployment: Knowlee AI Customer Intelligence | Timeline: 4 weeks to first predictive signals, 8 weeks to full automation

The Challenge

A mid-size direct-to-consumer e-commerce retailer selling home and garden products had built a substantial customer base over eight years. Their catalog ran to 4,200 SKUs. Their email list had 420,000 subscribers. Their monthly active buyer count hovered around 18,000.

The problem was not acquisition — the company's performance marketing was efficient and their brand had genuine organic following. The problem was retention. Across their active customer base, the average customer purchased 2.3 times over a 24-month window before effectively going dormant. Customers who should have been making four, five, or six purchases over that window were stopping at two.

The company's marketing team had a retention program: an email newsletter, periodic promotional campaigns, loyalty points accumulation, and a post-purchase survey sequence. None of it was targeted. The same email went to a customer who had purchased six times in the last year and a customer who had purchased once eight months ago and hadn't been seen since. The loyalty program accrued points for everyone, including customers who had effectively churned and would never redeem them.

The core analytical gap was that the team had no visibility into which customers were at risk of churning until they had already churned. The reactivation emails went out to the whole list on a schedule — not to the customers who needed them most, at the moment they needed them.

When the Head of Retention dug into the cohort data, the economics became stark. Retaining a customer for a third purchase increased their predicted lifetime value by 2.8x compared to a customer who stopped at two. But the company was letting those customers drift away without any targeted intervention because they couldn't identify who was drifting until it was too late.

"We have all the data," the Head of Retention noted. "We have transaction history, browse history, category affinity, email engagement, seasonal patterns. We just can't turn it into action fast enough, at the individual level."

The Approach

The company approached Knowlee with a specific question: can we predict churn at the individual customer level, far enough in advance to do something about it?

The initial data audit was encouraging. The company had 18 months of transaction data, three years of email engagement data, two years of browse behavior from their on-site analytics, and customer service interaction records going back five years. This was a rich dataset for building predictive models.

The deployment team identified a 60-day prediction window as the target: the goal was to identify customers who were likely to churn within the next 60 days — early enough that a targeted intervention could still change the trajectory.

Two weeks of exploratory analysis identified the behavioral signals that were most predictive of churn in this specific category:

• Purchase recency relative to the customer's personal baseline (a customer who normally buys monthly and hasn't purchased in 45 days is a different signal than an occasional buyer with the same gap)
• Decreasing email engagement — open rates and click rates falling below a customer's own historical average, even if they're still technically engaging
• Shift in browse-to-purchase conversion — customers who are still visiting the site but stopping short of buying
• Category drift — customers whose browse patterns are shifting away from their historical category preferences, which often indicates they're exploring alternatives
• Seasonal adjustment failures — customers who didn't return after a seasonal purchase even though comparable customers in the same cohort did

These signals, combined in a scoring model, produced a churn probability score for each customer that was updated daily. The model was trained on 18 months of historical data with known outcomes, achieving a 78% accuracy rate in predicting 60-day churn in validation testing — compared to the team's prior approach, which effectively had a 0% prediction capability (retrospective identification only).

The Solution: What Was Built

Component 1 — Daily Churn Risk Scoring

Every active customer receives an updated churn risk score each morning. The score is based on a weighted combination of the behavioral signals identified in the analysis phase, personalized to the customer's own baseline patterns rather than population averages. A customer is flagged as high-risk (>65% churn probability), medium-risk (35-65%), or low-risk (<35%).

The score is written to the customer's record in the CRM and is available to the email marketing platform, the customer service team, and the loyalty program management system.

Component 2 — Automated Retention Triggers

Different risk thresholds trigger different retention interventions, automatically:

High-risk customers receive a personalized retention sequence: a targeted email with a product recommendation in their highest-affinity category, followed 4 days later by a time-limited incentive (not a blanket discount — a category-specific offer, like free shipping on a garden tools order for a customer whose purchase history is concentrated in garden tools). If neither converts, the customer is flagged for a customer success outreach call if their historical LTV is above a threshold.

Medium-risk customers enter an early nurture sequence: two content emails designed to re-engage rather than discount (a buying guide, a seasonal tips piece, a "customers like you also bought" recommendation set). This sequence is designed to re-establish the purchase habit before it breaks rather than trying to recover it after.

Low-risk customers continue to receive the standard marketing program — no special treatment, no wasted resources.

Component 3 — Loyalty Score Personalization

The loyalty program was rebuilt around predicted LTV rather than cumulative points. Customers approaching a high-risk status receive accelerated point accrual notifications and targeted redemption prompts. Points that are about to expire trigger a personalized email highlighting what they can be used for, matched to the customer's category preferences.

High-value customers showing early churn signals receive proactive VIP outreach — a direct message from the customer success team, not a mass email. The system identifies which customers warrant this intervention based on historical LTV and predicted future value.

Component 4 — Win-Back Automation

For customers who have already churned — no purchase in 90+ days — the system runs a weekly win-back evaluation. Customers with historically high LTV are flagged for a win-back campaign that includes a substantive incentive. Customers who had one or two transactions are evaluated based on their browse history after the last purchase — if they continued browsing but not buying, that signals a different intervention than a customer who went fully dark.

Component 5 — Attribution and Learning

Every intervention is tracked to outcome. The system records which retention actions were taken for each customer and whether the customer made a subsequent purchase. This feedback loop continuously updates the model — interventions that are working get reinforced; those that aren't are deprioritized. The model improves month over month with accumulating outcome data.

The Results

35% churn reduction. Customer lifetime value doubled. Discount spend cut by 43% while revenue from retention tripled.

The discount spend reduction was an unexpected benefit. The prior program was applying the same promotional incentive to the entire customer base on a schedule. The AI-targeted approach concentrated incentive spend on high-risk customers who genuinely needed a reason to return — and used non-discount content engagement for medium-risk customers. The result was lower total discount spend and higher effectiveness.

The win-back rate improvement — from 6% to 18% — reflected the precision of the win-back targeting. Rather than blasting all dormant customers with a recovery offer, the system identified the subset most likely to reactivate based on their post-churn browse behavior and historical affinity patterns.

Before / After: Retention Program Mechanics

Key Takeaways

1. Churn prediction is only valuable if the prediction window is actionable.
Knowing a customer churned after they stopped buying is retrospective and useless. Knowing a customer is likely to churn 60 days from now — when they're still engaging, still browsing, still in the consideration zone — creates a genuinely actionable opportunity. The prediction window is as important as the prediction accuracy.

2. Personalization to the individual baseline outperforms population benchmarks.
A customer who normally buys monthly and misses one month is a different risk profile than a customer who typically buys quarterly and is two months out from their last purchase. Using each customer's own behavioral baseline to set the reference point produces far more precise risk signals than comparing everyone to the population average.

3. Less discount spend, better outcomes.
The prior intuition — that retention requires incentives — was partially correct. High-risk customers who needed a reason to return did respond better to targeted incentives. But medium-risk customers responded better to relevant content than to discounts. Applying the same discount universally was inefficient and was training customers to expect a promotional price rather than rewarding genuine loyalty.

4. The model improves with time.
The 78% prediction accuracy at deployment improved to 84% after six months of outcome data was incorporated into the model training. Organizations that deploy predictive systems should expect a ramp period — the model gets better as it learns from the specific outcomes of the specific interventions in the specific customer context.

5. Retention intelligence changes acquisition strategy.
Once the company had a clear picture of which customer cohorts retained well and which churned quickly, they used that data to adjust their acquisition targeting. Channel-specific customer profiles were analyzed for retention performance. Channels that were acquiring high-churn customers were deprioritized; channels producing retained customers received higher budget. This feedback loop between retention intelligence and acquisition decisions was not anticipated at the outset but became one of the highest-value applications of the data.

FAQ

How much historical data is needed to build an effective churn prediction model?
In this case, 18 months of transaction data with known outcomes was sufficient to build a model with 78% accuracy. The minimum practical requirement is 12 months of data that includes enough churn events to train the model. Companies with shorter histories or lower transaction volumes may need longer training periods or supplemental approaches.

Does the system work for subscription businesses, or only for transactional e-commerce?
The core prediction logic — identifying behavioral signals that precede disengagement — applies to subscription businesses as well, with some differences in how "churn" is defined. For subscription models, the signals look at usage patterns, feature engagement, and support interaction history rather than purchase frequency. Knowlee has adapted this framework for subscription contexts.

How does the personalization work technically — is every email unique?
Every customer receives a communication that is assembled from modular components: a product recommendation block that is AI-selected for their category affinity, an offer block that reflects their risk level and value tier, and content blocks that match their browse history. The assembly is unique to each customer; the underlying content assets are created once and reused intelligently.

What happens when the prediction is wrong — a customer is flagged as high-risk but doesn't churn?
False positives in this context are relatively benign. A high-risk customer who receives a personalized retention email and doesn't churn has received a positive brand interaction at no meaningful cost. The system calibrates thresholds to minimize false negatives (missing at-risk customers) more aggressively than false positives, which is the correct priority in a retention context.

Was there any concern about customers feeling surveilled by the personalization?
The personalization is designed to feel like attentive service, not surveillance. Customers receive relevant recommendations and timely offers — experiences that feel like a retailer paying attention, not tracking. Customer satisfaction scores have not declined post-deployment; in fact, the net promoter score improved 11 points over the 12 months following the AI program launch.

See How Knowlee Can Deliver Similar Results for Your Team

Churn prediction and retention automation work in any customer business with transaction data — e-commerce, SaaS, financial services, subscription media, and more.

Talk to a Knowlee specialist about your customer intelligence program — or explore our AI Customer Intelligence Platform overview.

Related reading:

• AI Customer Intelligence Platform: A Complete Guide
• AI Loyalty Score Measurement
• AI Marketing Analytics and Attribution