Building a real-time propensity model that matches incoming customers to the right sales agent — a 12% conversion lift that changed how our entire call center operates.
Picture this: 1,000 incoming customer calls per hour, 200 available agents. Some agents are natural upsellers; others excel at converting hesitant first-time buyers. Some customers are ready to purchase; others need convincing.
Random routing ignores all of this. We set out to build something better: match customer propensity with agent capability in real-time, before the call connects.
We benchmarked XGBoost, LightGBM, and CatBoost head-to-head. CatBoost won for practical reasons, not theoretical ones.
Categorical features without leakage. Our feature space was heavy on categoricals — customer segment, product category, call reason, agent skill tier. CatBoost's ordered target encoding avoids the leakage problems we hit with standard target encoding in the other frameworks. Less preprocessing, fewer subtle bugs. Automatic interaction discovery. CatBoost found that the interaction between time-of-day and customer segment was highly predictive — something we hadn't hypothesized and might never have engineered manually. Graceful handling of sparse segments. Some customer-product combinations had very few training examples. CatBoost's symmetric tree structure and ordered boosting handled these thin slices without overfitting, where the alternatives needed explicit regularization tuning per segment.Purchase history (recency, frequency, monetary), service interaction patterns, current product portfolio, tenure, and demographic segment.
Time of day, day of week, holiday proximity, recent marketing touchpoints, current promotion eligibility, and queue wait time. That last one surprised us — longer waits meaningfully correlate with lower purchase propensity.
Historical conversion rate by product type, average handle time, customer satisfaction scores, and current shift duration. Agent fatigue is real and measurable.
Customer-agent segment compatibility scores, product affinity signals, and whether the opportunity is cross-sell vs. upsell.
We ran a 50/50 split for 8 weeks:
| Metric | Control | Treatment | Lift |
|---|---|---|---|
| Sales conversion | 18.3% | 20.5% | +12% |
| Average order value | $42.10 | $43.80 | +4% |
| Customer satisfaction | 4.2/5 | 4.3/5 | +2.4% |
| Handle time | 8.2 min | 7.8 min | -4.9% |
At our call volume, the 12% conversion lift translated to roughly $15M in annualized incremental value. The handle time reduction was a bonus we didn't expect — better customer-agent matching means fewer awkward conversations.