Predictive Churn Prevention: How AI Identifies At-Risk Customers 90 Days Before They Leave

RFM segmentation identifies customers who have already churned. A recency score of 1 means the damage is done — the customer stopped buying months ago, and whatever caused them to leave has had time to solidify into habit. Predictive churn prevention operates upstream of that signal, identifying the customers most likely to churn before their behaviour makes the prediction obvious.

The difference is significant. A customer flagged as at-risk 90 days before their last purchase can be reached with a retention campaign while the relationship is still warm. A customer flagged after 6 months of silence requires a win-back campaign that starts from near-zero engagement — a fundamentally harder and more expensive problem.

What Churn Actually Looks Like in B2C Data

Churn in B2C businesses rarely happens as a discrete event. Customers do not cancel an account or send a resignation letter. They drift — visit less, spend less, engage less, and eventually stop altogether. The signals are visible in the data before the final exit.

Declining email and WhatsApp engagement: A customer who previously opened every message and now opens one in five is not just busy. Their interest is fading. Engagement velocity — the rate at which open and click rates are declining over time — is a stronger churn signal than the current open rate in isolation.

Shrinking basket size: A customer spending €85 per transaction who gradually moves to €40 transactions is signalling a change in relationship. It may be budget-related, or it may be that they are now splitting spend with a competitor. Either way, the trend line matters more than the individual data point.

Channel drift: A customer who shifts from in-store to online to no channel at all follows a predictable disengagement arc. Tracking the channel mix over time — not just the last touchpoint — reveals movement patterns that precede churn.

Reduced visit cadence without an explanatory event: A customer who visits every 10 days and suddenly visits every 25 days has not adopted a new schedule. Something changed in their relationship with the brand.

The Intervention Window and What to Do With It

Detecting a at-risk customer 90 days before their probable last purchase creates a window for intervention. What happens in that window determines whether the customer is retained.

Weeks 1–4 after detection: The intervention should feel like a natural touchpoint, not a retention alarm. A WhatsApp message referencing something specific about the customer’s history — not a generic re-engagement offer — is the right tone. “We noticed we haven’t seen you since the autumn collection — the new arrivals this week have a few pieces that match what you loved last time.” The customer does not know they were flagged as at-risk. They receive what feels like a thoughtful personal message.

Weeks 5–8: If no engagement from the first message, escalate the specificity. A message from a named person on the team (the manager, the personal shopper, the account manager — whoever is relevant) that is even more specific: “I wanted to reach out personally — is there anything we could do better for you?” This works because it invites a response rather than demanding a purchase.

Weeks 9–12: The final intervention window. If the previous two touchpoints produced no engagement, this is the moment for a meaningful value offer — not a promotional discount, but a genuine reason to return: an exclusive event, early access to something specific to their history, or a personal invitation that treats the non-response as an oversight rather than a rejection.

Automating the Detection and Response

Caramel’s AI agent monitors the signal stack continuously across every customer profile. When a customer crosses the combined-signal threshold, the agent automatically initiates the appropriate intervention sequence — calibrated by the customer’s segment, history, and communication preferences.

The natural language analytics layer allows marketers to audit the churn pipeline at any time: “How many customers are currently in the at-risk pool?” or “Which segment has the highest churn velocity this month?” — without requiring a data analyst to build the query.

For the RFM foundation that powers churn segment identification, see The RFM Model Every B2C Business Should Build Before Running Another Campaign. For lifecycle trigger campaigns that fire automatically on churn signals, see Trigger-Based Marketing: The 7 Customer Signals That Should Launch Automatic Campaigns.