Conversational AI in Financial Services

At Cognis, we have completed several proof-of-concept initiatives, where the results exceeded expectations. During a recent test involving an onboarding scenario, which used a fine-tuned model, 91 percent of the model’s answers were rated as “Acceptable”, with 2 percent rated as “Unsure”. The success rate climbed to 96 percent leveraging GPT-4. We attribute these results to both, the sophistication of the models and the quality of the source documents, which are subject to Good Disclosure Principles.

It is important to note that using GPT-4 can be expensive. The test runs (end-to-end application scenarios) that used GPT-4 cost between 12 and 20 cents (USD) each. In contrast, using the fine-tuned model cost less than 3 cents per application. We recommend modelling these financial scenarios as there is a tradeoff where a better model can result in a higher conversion rate but at a higher cost.

Given these results, we believe conversational AI represents a significant opportunity to improve the onboarding customer experience and reduce abandonment rates.