Case-Based Reasoning for General Electric Appliance Customer Support

William CheethamAAAI 2006

Introduction

In the late 1990s, General Electric (GE) faced a significant operational challenge within its appliance service division. Field service visits, costing upwards of $50 per trip, were a major expense. Analysis revealed that nearly 20% of these visits resulted in "customer education" rather than mechanical repairs—meaning the issues could have been resolved over the phone. However, call center agents struggled to provide these solutions due to the increasing complexity of modern appliances and a high annual staff turnover rate of 60%.

To bridge this gap, GE developed and deployed the "Support The Customer" (STC) system. Published in 2006, this landmark project utilized conversational Case-Based Reasoning (CBR) to capture the tacit knowledge of veteran engineers and make it accessible to entry-level call-takers, effectively augmenting human intuition with a structured digital knowledge base.

The STC System: Methodology and Interface

The core of the STC system is Conversational Case-Based Reasoning. Unlike rigid rule-based systems, CBR solves new problems by retrieving and adapting solutions from a curated library of past cases. This library was developed by a dedicated team of 40 technical experts who abstracted real-world problems into diagnostic paths.

The STC system interface, showing the diagnostic questions and potential results presented to the call center agent.

A defining feature of the system is its Mixed-Initiative Interaction. The AI does not replace the agent; instead, they work as a team. The human agent manages the natural language conversation with the customer, while the STC system handles technical information retrieval and diagnostic logic. The system also employs "auto-answering" capabilities, using data such as model numbers or repair history to skip redundant questions and streamline the call.

To ensure consistency and safety, the diagnostic logic was first mapped using Visio-based decision trees. This allowed GE to verify that every path met legal and quality standards before being digitized into the case base.

Key Insights: Augmentation Over Automation

The success of the STC system stemmed from recognizing the limitations of AI at the time. In 1999, natural language processing (NLP) was not robust enough to handle complex customer descriptions. By keeping the human "in the loop" to translate customer speech into system inputs, GE bypassed the NLP bottleneck.

Furthermore, the system provided value even when a phone resolution was impossible. By identifying exactly which part was failing, GE could utilize the "Intercept" process—mailing the part directly to the customer or ensuring a technician arrived with the correct component on the first visit. This reduced "fishing" behavior, where customers would call repeatedly hoping for a different answer, and ensured a unified diagnostic standard across the organization.

Operational Impact and Results

The implementation of the STC system delivered profound financial and operational benefits. Between 2000 and 2005, GE saved an estimated $44.5 million by avoiding unnecessary field service visits.

The percentage of calls resolved over the phone increased significantly following the implementation of the STC system.

As shown in the performance data, the probability of resolving a customer's issue over the phone rose from a baseline of 3.9% in 1999 to nearly 25% by 2005. Beyond the metrics, customer satisfaction surveys indicated that users preferred the guided, efficient diagnostic process provided by STC over traditional troubleshooting methods.

Conclusion and Future Outlook

While the STC system was a major success, it highlighted the significant "knowledge acquisition burden" required to maintain a high-quality case base. The project required a full-time case author and continuous feedback loops with call-takers to keep the logic current.

As the technology matured, the authors noted a shift toward web-based self-service tools. The ultimate goal of the project was to eventually transition this diagnostic logic directly to the consumer, removing the need for a human intermediary where possible and further scaling the benefits of automated expertise.

Summary updated

BibTeX
@inproceedings{cheetham2006case,
  author = {William Cheetham},
  title = {{Case-Based Reasoning for General Electric Appliance Customer Support}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure