Transforming Call Center Operations with AI Models

This technical report details AT&T's innovative approach to enhancing customer service through the implementation of small language models (SLMs) powered by H2O.ai. The company, which handles approximately 5 million customer calls annually, aims to extract valuable insights from these interactions to improve service quality. Initially, AT&T utilized the GPT-4 model but faced challenges related to cost and processing time. To address these issues, AT&T distilled large language models into smaller, fine-tuned models, achieving significant reductions in processing time from 15 hours to 4.5 hours and increasing transcript processing capacity from 45,000 to 250,000. The report outlines the architecture of the new model deployment, which includes a fine-tuned classifier and the H2O Danube model, and highlights the business value derived from this transition, including a 91% accuracy rate and substantial cost savings. The findings underscore the effectiveness of SLMs in optimizing customer service operations within the telecommunications sector.