The rate at which generative artificial intelligence is evolving sets it apart from previous innovations. Citi’s Prag Sharma, told finews.com how data-rich banks can benefit from the technology.

«Overall, artificial intelligence, especially its latest incarnation, 'Generative AI,' will make us more efficient,» Citi’s global head of artificial intelligence Prag Sharma told finews.com in an interview.

Large Language Models (LLMs), which generate human-like text by predicting the next word in the sequence, have been trained on enough data and now understand the statistical properties of language. What is missing and will remain absent is «common sense.»

Literacy Versus Experience

«Language describes many things, but I believe that it doesn’t capture all the unique insights and experiences we as humans have had,» he said. Bearing in mind that «no single technology on its own has ever made an outsized impact,» Sharma sees the immediate impact of generative artificial intelligence (Gen AI) as a «point solution» to replace specific activities, especially those which rely on manually sifting through a lot of data.

«If you’re doing only a single repetitive task every day, such as generating personalized images, then you may need to reconsider» your profession. Otherwise, there are few cases where technology can replace humans entirely, he said.

Teething Problems

Moreover, «there’s a lack of understanding about what the technology can do,» and people are making mistakes, he said referring to a case in which an experienced US lawyer allegedly relied on chatGPT for a motion document he filed in a personal injury claim without reviewing it.

Accountability stays with the end user who is consuming the outputs of these large language models. Essentially, ChatGPT shouldn’t be treated differently from how one uses the internet. After all, «you can’t do an internet search and simply pick the first result and run with it, just because you are in a hurry,» Sharma said.

Sharma appreciates supervisors being technology and sector agnostic towards AI, saying the EU’s approach that distinguishes between low and high-risk use cases is «going down the right path.»

«Ultimately, we should be able to explain how we are using something,» he said.

Out of Scope

To date, 80 percent of that data, including emails, texts, legal documents, marketing material and news articles, is unstructured. Although a lot of documentation has been scanned and digitized, most of it has not been fed into LLMs.

«The question is now 80 percent of data is in play, what can we do with it?» Sharma asked.

Proprietary Data

This is precisely why banks, which hold proprietary client data that doesn’t originate from the internet, are well-positioned to benefit from the next Gen AI wave.

«I think we need to combine that (data) in ways to offer unique insights that do not exist today and haven’t been commoditized yet,» he said, adding that this is where banks are going to find future value.

Closed Networks

Citigroup is currently exploring the technology among an initial cohort of 100 end-users, for writing code and for desk-based research and information retrieval activities.

Yet, further down the line, it will become easier for corporates to use models in combination with sensitive data as models are getting smaller which means you can deploy them in-house in a closed environment, he said.

Although «banking jobs will be AI-assisted,» similar to how computers assist banking jobs today, Citigroup's recent research found that finance is still people-driven with clients preferring to resolve queries around mortgages, car loans and life insurance purchases with another human.  

Paradoxically for banks, Gen AI means «the relationship and trust with the client will become even more important,» Sharma added.