A financial services company sought to extend its support beyond regular business hours by implementing an AI-powered chatbot. The challenge lay in addressing customer inquiries during off hours, in various formats and without human intervention, while maintaining high accuracy and access to information.
Our approach involved a phased strategy, beginning with a proof of concept focused on a narrow use case and expanding to cover a wider array of data. Rigorous testing highlighted OpenAI's GPT-3 Davinci-003 model as the optimal choice. Starting with a Minimal Viable Product (MVP), Phase 1 concentrated on known questions and data extraction from a predefined source. Phase 2 involved generating synthetic data to accommodate diverse chatbot process flows, encompassing multiple versions of FAQs, and culminating in model fine-tuning.