Intelligent Outbound Call Automation for Hotel Order Confirmation at Ctrip

Author Zhou Zhenwei, a data science engineer at Ctrip’s Data Intelligence Department, shares a case study from the second Ctrip Cloud Machine Learning Salon, focusing on improving call‑center automation for hotel order confirmations.

When a user books a hotel through the Ctrip app, the order must be confirmed by the hotel. Traditionally, the call center waits for the hotel’s reply and, if the reply is delayed, manually calls the hotel to urge confirmation. This process can cause long user wait times and consumes significant human resources.

The call center’s responsibilities are divided into four areas: pre‑confirmation processing, post‑confirmation processing, general customer service (order modifications, cancellations, invoicing, etc.), and complaint handling. Except for complaint handling, the other three functions involve repetitive, rule‑based steps that are suitable for automation.

To address inefficiencies, Ctrip introduced a predictive intelligent outbound (outbound call) system. Two machine‑learning models are used: a response‑time prediction model that estimates whether a hotel’s reply will exceed a tolerance window, and an outbound‑effectiveness model that predicts whether a manual call would be useful.

The new workflow operates as follows: if the response‑time model predicts a delay beyond the tolerance, the system schedules an early outbound call; before each call, the effectiveness model decides if the call is likely to succeed. Effective calls are prioritized for human agents, while predicted‑ineffective calls are deferred and handled by an IVR (interactive voice response) system, with human follow‑up only if IVR fails.

Both models are supervised learning models built from historical order data, incorporating features such as booking time, days to check‑in, room type popularity, hotel‑partner relationship, and past outbound call statistics. After extensive offline training, the XGBoost algorithm was selected for the effectiveness model.

After deployment, the intelligent outbound system increased total order volume by 25%, reduced the proportion of orders requiring manual outbound calls by one‑third, and maintained similar user confirmation times, demonstrating that predictive resource allocation can significantly improve call‑center efficiency without harming user experience.

The article concludes that big data and machine learning can effectively automate and optimize call‑center operations, and that similar predictive approaches can be applied to other business processes to achieve higher efficiency and better service quality.