Predictive Outbound Call Platform: Algorithm, Design, and Production Deployment at Ctrip

The article introduces the author team from Ctrip's Basic Business R&D Call Center, which leverages softswitch, intelligent routing, and automatic speech semantics to provide human‑machine interactive voice services for Ctrip users.

Ctrip operates over 15,000 seats with a daily outbound call volume of around 300,000. To lower agent waiting time and boost outbound efficiency, they developed a predictive outbound platform.

Principle Overview

The predictive outbound platform automatically executes outbound tasks based on predefined strategies and number lists. Calls are placed sequentially; successful calls are transferred to agents or IVR, while failures are logged with detailed results (e.g., powered off, invalid number, no answer).

The core idea is to predict that in N seconds X agents will become idle, then pre‑dial Y calls so that a new customer call is ready exactly when an agent finishes the previous one.

Predictive Algorithm

Accuracy depends on agent answering ability and called party connection rate. The formula used is:

OutboundVolume = (IdleAgents - OngoingOutbound*SuccessRate - QueueLength*NoAbandonRate + PredictedFreeAgentsAfterNSeconds + OngoingPostProcessing*ConfigFactor) / (SuccessRate*NoAbandonRate) * ManualAdjustmentFactor

Key considerations include:

The algorithm treats queue behavior statistically and may degrade if agent behaviors vary widely.

Fluctuations in answer rate can cause sudden queue spikes.

Task initiation based on idle agents can be slow; large variations in available agents affect prediction.

Average call duration variance impacts queue time predictions.

Customer waiting time influences call loss, which the algorithm cannot control.

Therefore, the model must be continuously iterated with historical data for each business scenario, and manual overrides should be allowed for real‑time control.

Platform Design and Construction

After completing the algorithm model, the team built a new SoftIVR outbound platform, tightly integrated with Ctrip’s SoftPBX (a SIP‑based softswitch). SoftPBX now serves multiple business units (vacation, hotel, flight) handling about 50,000 calls daily, offering transparent data, easy deployment, smooth scaling, and reduced O&M costs.

Figure: System logical architecture.

Feature 1: Multi‑point Communication Architecture between Outbound System and SoftPBX

SoftIVR uses a floating‑IP active‑standby mode where a single SoftPBX handles outbound calls. To overcome capacity limits, an OutCallService layer connects one‑to‑one with SoftPBX, and multiple OutCallService instances compete for SoftPBX IPs, providing horizontal scalability.

Figure: Communication structure.

Feature 2: Concurrent Generator for Outbound Triggers

The platform adopts a multi‑instance, per‑business concurrent processing model, solving two problems:

Single‑instance sequential traversal leads to long scheduling cycles as business count grows; per‑business concurrent scheduling improves real‑time dispatch.

Single task‑table scanning becomes slower with larger data volumes; partitioned tables dramatically reduce selection time and lessen impact on other operations.

Figure: Task scheduling structure.

The SoftIVR platform replaces vendor solutions, runs on standard VMs, and offers low‑cost deployment, comprehensive monitoring, and open APIs. It currently supports hotel and train ticket outbound services, handling over 100,000 calls per day.

Within two months, the predictive outbound platform was launched and validated in production.

Production Instance

SoftPBX’s mature load handling enabled a smooth transition to production. The first customer was the hotel business, with 200+ seats deployed across Shanghai and Xinyang within two weeks, including server setup, firmware updates, and CCDesk release.

Emergency plans ensure fallback to manual outbound if the platform fails, and thorough training minimizes operational impact.

Gradual gray‑release mitigates risk during the switch‑over.

Results to date:

Hotel predictive outbound covers 200+ seats.

Outbound success rate improved by ~7% compared to manual dialing.

Average outbound calls per agent per day increased by over 20%.

As SoftPBX seats migrate, the predictive outbound platform will continue to enhance efficiency for more agents, with further feature extensions planned.

Further Reading

How Ctrip Optimized React Native

Pitfalls in Ctrip Train Ticket RN Development

Common User Password Encryption and Cracking Methods

Ctrip Mobile Architecture Evolution and Optimization

Ctrip Risk Defense System Transformation