We received an inquiry from the press today about sharing our thoughts on how can ChatGPT be used in the banking industry, the opportunities and threads it will bring, and how should banks and credit unions approach this technology?

I will share my personal opinion from all of the knowledge and experience I’ve gathered over the past 12 years helping financial institutions automated their banking processes. 


How can ChatGPT and GPT4 be used in the banking industry? 


The technology behind large language models (LLMs) such as ChatGPT in my view represent the next wave of evolution in process automation across all industries. As this technology matures it will help the banking industry and others make the leap from automated processing to autonomous processing in the form of autonomous A.I. agents that specialize in many areas of banking knowledge and processes. It’s important to understand that these technologies are evolving at lightning speed. ChatGPT itself was release less than 5 months ago on November 30, 2022 and just 2 weeks ago new frameworks like AutoGPT and BabyAGI, which introduce a new way of using and combining these highly advanced A.I. models towards solving complex goals, proved that it is not only technically possible but economically feasible to assemble self-directing autonomous agents aimed at highly complex and valuable goals like performing thorough research to generate concise recommendations for a potential purchase or planning an event completely independent of human input aside from providing their primary directive.

For banking, this wave of autonomous intelligent agents will start showing up across every area where there’s a set of tasks to complete that are defined using specialized banking knowledge as the key ingredient to perform them. Most of the work performed at banks is knowledge work. Some examples in lending will be tasks like autonomously ensuring that all required information and documentation for each specific loan application is collected as fast as possible without involving human agents in the process. In this case, an autonomous agent who will be collaborating with the in-house team in real-time will review every incoming loan application, decide which pieces of information and documentation are missing, and then engage applicants over multiple touch-points across different communication channels to assist in collecting this information without the involvement of a human agent.

A very similar scenario will show up for validating fraud alerts among deposit account applications. In this case, the autonomous agent will review any alerts received from credit bureaus, assemble an internal task list using specialized knowledge and engage with different systems as well as the customer to gather enough evidence to verify the identity of the applicant.

It’s important to know that the first wave of autonomous agents will not be perfect. They will not be able to handle every possible scenario at first. However, with the rapidly evolving advancements in short and long-term memory facilitated by combining vector databases with large language models (LLMs) like GPT4, these early autonomous agents will be able to recall previous similar interactions. This will allow them to learn overtime and improve its accuracy on their own. So while they might not be able to handle 100% of all tasks, they will be able to tackle 50% or more of current knowledge-based tasks at first. We will see the rise of autonomous banking agents for every area of a bank’s operation, which will be transformational for most financial institutions who operate with limited resources.


What are the opportunities and/or dangers of using it in financial services? 


One of the biggest opportunities and threats that autonomous banking agents powered by large language models (LLMs) like ChatGPT will bring to banking is that they will democratize specialized banking knowledge. Today, learning how to properly process a loan, open an account, or service existing loans is highly specialized knowledge that takes years to master and is only accessible to a select group of people who go through rigorous training to acquire it. Autonomous A.I. agents will learn this specialized knowledge relatively fast, which will provide a lot of leverage to most financial institutions. This will also have a net positive effect on existing teams at financial institutions as this technology will give them back time to focus on higher order of priorities and do a lot more with less. It’s reasonable to assume that their jobs will also evolve away from manual repetitive work and closer to collaborating and managing these autonomous agents who will be directly collaborating with human agents to move business goals to completion faster.  

However, because of the nature of banking knowledge, which follows very strict forms and patterns, there is a real risk that eventually many of these autonomous agents will converge. What this means is that because most of these autonomous agents will be prompted/assembled to follow very similar knowledge and perform very similar tasks, they will all start to behave similarly. This will commoditize personalized banking knowledge as autonomous agents will provide the ultimate personalized service experience, which will make it harder for community financial institutions, who are known for their personal touch, to differentiate themselves.

An additional risk is the fact that the very nature of these autonomous agents allows them to make their own decisions on how they approach a task within the limitations of their initial prompting. This means that the degree in which an autonomous agent makes accurate or inaccurate choices is directly related to the quality and robustness of its prompting. I believe that prompt engineering will become a highly coveted skill within banking and every other vertical.  


How should bank approach this technology?


I align with Sam Altman’s (CEO of OpenAI) analogy about ChatGPT, GPT4 and other large language models being the equivalent of a “calculator for writing”. I just see how any of us would ever want to go back to crunching complex calculations by hand now that we have the calculator or better yet, now that we have excel. 

I would advise embracing it fully, as early as you can to get an understanding of this completely different technological paradigm. It’s very easy to “play” with it and it will become easier and easier very fast however, just like the analogy of the calculator or excel, the only way to really understand how to think about it is to simply grab one and start inputting things into it. It then becomes really simple and it will allow your team to have a foundation to decide how, when, and where you’d like to start applying this new technology to help your institution to a lot more with less.

Aside from this, we’ll share a Shastic pro tip with you about a best practice to get started with this type of autonomous automation technologies: for whatever are of the operation you are considering applying this to, begin by creating a highly detailed map of your current workflow across different people, roles, and systems. The more detailed the better. We’ve found that all institutions who go through this process with us learn things they did not know about their own processes or align their understanding of how certain things should work with how they actually work every day. Starting from an accurate view of your existing workflow is the best way to avoid the risks that autonomous agents powered by large language models will bring.