A U.S. telecom company cut costs and boosted upsells by adopting advanced keyword search, transformer models, and open-source large language models (LLMs).
SITUATION
Our client's proprietary GPT-4 based application incurred excessive licensing costs while predicting customer churn from call transcripts. They needed to optimize these costs while maintaining accuracy and performance. Additionally, the client aimed to increase sales through upsells by better understanding customer buying behavior from the same call transcripts.
84% Reduction in Large Language Model Usage Costs
The Everforth Apex solution involved transitioning from the GPT-4 based application to a sophisticated integration of keyword search capabilities, transformer models (DeBERTa), and open-source LLMs (Llama). This was complemented by the development of machine learning pipelines and workflows utilizing Databricks. Integral to this solution was our collaborative effort with the client’s product team. Together, we crafted transformer architectures and harnessed the strength of open-source LLMs to construct robust pipelines.
This sophisticated and innovative solution was designed to interface seamlessly with Azure Data Lake Storage (ADLS) and Snowflake, empowering our client to efficiently retrieve and preprocess the transcription data. To enhance the performance, the Llama models have been finetuned with QLORA and PEFT techniques on a comprehensive list of 89 churn indicators. Additionally, this solution applies a similar strategic approach to identify upsell opportunities, tailoring its analysis to various questions. This dual capability solution empowered our client with actionable insights, enabling them to not only retain their customer base but also to identify and capitalize on growth opportunities.
RESULT
Replacing the GPT-4 solution resulted in a significant 84% reduction in LLM usage costs, lowering monthly expenses from $250,000 to $40,000. Additionally, by listening to customers' product concerns, understanding their buying behaviors, and automatically recommending relevant additional products, the system significantly increased upsells by 20%, further enhancing the company's profitability and operational efficiency. The Everforth Apex team's efforts surpassed anticipated benchmarks, driving significant cost efficiencies and increased sales.
