AI Intelligence Layer

Models, predictions, and explainable reasoning

Every decision Customer360 makes is grounded in versioned models with confidence, drift, and feature attribution.

Churn Predictor

Classificationv4.2.1

Predict 30-day churn likelihood per customer

Confidence
91%
Precision
88%
Recall
84%
Drift
7%
Top features
Reasoning summary

Engagement decay across email + app, combined with two unresolved support tickets and a falling NPS slope, are the strongest churn signals this week.

Last trained: 2 days ago

Propensity to Buy

Classificationv2.7.0

Score likelihood to purchase a product line in next 14d

Confidence
86%
Precision
82%
Recall
79%
Drift
4%
Top features
Reasoning summary

Customers showing repeat browsing of premium SKUs with at least one wishlist add convert at 4.1× the baseline rate.

Last trained: 8 hours ago

Next-Best-Offer

Recommendationv3.1.4

Recommend the most profitable, accepted offer per customer

Confidence
83%
Precision
78%
Recall
81%
Drift
9%
Top features
Reasoning summary

Offer ranking optimises for blended margin × acceptance probability, suppressing offers a customer has rejected twice in 60 days.

Last trained: Yesterday

Behavioural Segmentation

Clusteringv5.0.2

Cluster customers into actionable behavioural segments

Confidence
79%
Precision
74%
Recall
76%
Drift
12%
Top features
Reasoning summary

12 stable behavioural clusters detected. Two are emerging (luxury-seekers, value-hunters) and warrant new playbooks.

Last trained: 3 days ago

Customer Health Score

Regressionv1.9.3

Composite 0–100 score reflecting relationship health

Confidence
88%
Precision
85%
Recall
83%
Drift
5%
Top features
Reasoning summary

LLM-derived sentiment from chat + reviews now contributes 20% of the score, surfacing dissatisfaction before it becomes attrition.

Last trained: 12 hours ago

Profile Narrator (LLM)

LLMv1.2.0

Generate plain-language summaries and explanations

Confidence
94%
Precision
92%
Recall
90%
Drift
2%
Top features
Reasoning summary

Grounded on canonical profile + last 90 days of events. Refuses to speculate beyond evidence; cites sources inline.

Last trained: Continuously fine-tuned