Case Studies

Built on real data.
Validated by the numbers.

Prediction models built from real customer activity.

Churn Prediction Featured · Validated on industry benchmark (IBM Telco, 7,043 records)

From losing your best customers to knowing exactly who to call first

A telecom company with 7,043 customers is losing more than one in four. Management assumed it was price competition. The data told a different story entirely. We built a binary classification model on SageMaker Canvas using the IBM Telco Customer Churn dataset — a real-world benchmark used across the industry.

The model identified contract type as the dominant churn signal — not price, not service issues. Month-to-month customers churned at dramatically higher rates than customers on annual or two-year agreements. New customers in their first year on month-to-month contracts represented the single highest-risk profile in the entire customer base.

Key takeaway: The signal that mattered wasn't price or service quality. It was contract type. Retention teams that work month-to-month customers in their first year are working the right segment.

Feature importance
SageMaker Canvas feature importance — Contract ranked #1
Model metrics
SageMaker Canvas advanced metrics — AUC 0.850
77.6%
Accuracy on this dataset*
0.850
AUC-ROC score
Contract Type
Top churn signal
26.5%
Baseline churn rate
Lead Scoring Also available · Validated on EdTech dataset (9,200 records)

Which leads to call first — at 95% accuracy on this dataset*

Behavioral signals (time on site, email engagement, CRM tagging) predict conversion far better than demographics. Top driver: CRM tagging at 25.6% of model impact. Rep capacity recovered: ~31%.

Step one: surface the problem
The first deliverable is always the same — show you what's already there.
Cohort retention · anonymized client engagement
Cohort retention heatmap
Raw multi-year CSV, untouched. The red is the point. Before we predict anything, we surface what the data has been hiding.
Example output
Your Monday morning at-risk report
Not raw data. Not a spreadsheet to decode. Scheduled reports showing which customers need follow-up, why they were flagged, and what action to take.
PrediktSales Customer Risk Report
Monday · 8:00 AM
3 customers flagged
Peak Music Lessons
Jon C
Attendance down 40% · no lesson booked in 28 days
87% risk
Suggested action: Personal outreach before the next billing cycle.
Riverside Gym
Michael R
Check-ins down sharply · payment retried twice
72% risk
Suggested action: Trainer outreach and membership review call.
Horizon Martial Arts
Sarah L
Attendance down 25% · missed last 2 classes
58% risk
Suggested action: Front-desk follow-up and renewal reminder.

Get a free consult

Tell us what you're working with and we'll tell you exactly what's possible — and what it would cost. No obligation.

Get a free consult
* Prediction accuracy depends heavily on the quality and consistency of the underlying business data. Results shown reflect performance on the specific datasets referenced and will vary based on CRM structure, historical records, workflow consistency, and available inputs.
PrediktSales Assistant