Call center forecasting & optimization

Booking.com’s Customer Service (CS) center employs thousands of agents to handle over a million phone and email interactions, in 43 different languages, every day. Scheduling a workforce of such scale and complexity has been a challenge for some time. But with rapid growth in new markets, the daily 13-week forecast accuracy of CS Workload – a key metric – was deteriorating. In some markets by more than 25%. The goal was to halve forecast errors with a system that agents and managers could easily understand and adopt.

We developed a prototype which could easily scale on proof of concept and tested in two language markets – one European and one Asian. Our approach was based on four principles:

1. Understand the business dynamics that inform the patterns.
2. Smartly combine data and select only the most important drivers to reduce complexity.
3. Thoroughly test different techniques and manage outliers by identifying and understanding them.
4. Build a system that integrates a learning loop to correct forecast deviations.

The total workload forecast error decreased by 60% for the pilot European language and 50% for the pilot Asian language. This was primarily driven by a 75% improvement in Contact Volume Forecast accuracy. A 60% improvement in Handling Time Forecast accuracy was also a major factor. When applied to the remaining 41 Customer Service languages, only two did not show improvement. These two outliers will be the initial focus of our test & learn approach.