Launched in November 2022, GPT-3.5 drew in a million users from its early days. For banks, the question is no longer whether conversational agents will enter the customer relationship, but under what conditions they will entrust them with speaking.
The banking sector, however, is accustomed to quickly absorbing technological ruptures, from online banking to mobile apps. Nevertheless, generative artificial intelligence is changing in nature. It does not merely perform an operation: it formulates, argues, suggests. The bank that responds yields to the bank that converses, with all that responsibility entails.
This is the finding stated in a research note published recently titled “Opportunities and Challenges of Artificial Intelligence for Banks: A View Through ChatGPT.” Based on a review of international literature, the note maps what this tool can bring to institutions, and what it may cost them.
Eight years in gestation, a single turning point
The apparent suddenness of ChatGPT masks a long maturation. OpenAI was founded in 2015. GPT-1 appeared in 2018, GPT-2 in 2019, the year when Microsoft invested one billion dollars in the lab, then GPT-3 in 2020, with an announced collaboration between the two players.
The commercial breakthrough arrives with GPT-3.5 in November 2022, then GPT-4 arrives on March 14, 2023. For banks, the message is clear. The technology has crossed the commercialization threshold, and its improvement cycles are now counted in months.
Bank marketing learns to listen
The first area of application: communication. The note describes a tool capable of exploiting the history of behavior to personalize exchanges. A client who conducts multiple foreign exchange operations could thus receive a tailored exchange offer or investment advice abroad. On social media and websites, the model would answer frequent questions, for example about loan interest rates, by understanding the context and intent.
The interest goes beyond automation. The analysis of customer comments makes it possible to extract sentiments, preferences, complaints, and suggestions, and to quickly assess the perception of a new product. The performance of past ads can even feed the production of optimized headlines and descriptions. A caveat weighs on each of these promises: they assume that the bank has beforehand implemented the technical integrations giving the model access to the relevant data.
The back office, a discreet laboratory of productivity
