Advancements in generative artificial intelligence, and in particular large language models (LLM) such as ChatGPT, have simultaneously caused mass excitement and concern about their impact on each and every industry. As an industry that has already dipped
its toe into the field of AI, payments is a little-discussed yet key testing ground for the real world implementation of LLMs.
Where can we find artificial intelligence in payments today?
A decade ago, data intensive processes in payments were mostly managed by rule-based systems. As machine learning became more feasible and increased efficiency across the card payments value chain - from fraud detection to customer onboarding - rule-based
systems were gradually augmented with machine learning throughout the payments industry.
The current pace of innovation in machine learning will be significantly increased by the introduction of LLMs, as the general benefits of ChatGPT and its competitors are realised. This will include assistance for software developers as they implement new
payment applications, support for product teams as they refine business models for new payment products, and much more.
Large language models driving innovation in unexpected domains
Payment operations might experience a seismic shift when LLMs enter the card payments value chain, especially in sections that have remained unaffected by artificial intelligence thus far.
A great example of LLM’s impact on payment operations is how issuers interact with their cardholders. Currently, issuers use a combination of chatbots directing cardholders to helpful resources and humans handling all decision making. Recent advancements
in LLMs will enable issuers to utilise advanced chatbots and other automated customer service channels to perform actions upon the request of cardholders, acting as agents capable of executing decisions.
Ambitious banks, traditionally a role filled by neobanks, will further leverage the capabilities of LLMs by utilising enormous training data sets and constant improvement from past interactions with cardholders to build virtual agents that act as the designated
decision maker. This could include making authorization and onboarding decisions, increasing credit limits and handling disputes.
Another exciting area would be LLM’s text understanding capabilities; LLMs are able to analyse and make decisions based upon a larger and more diverse set of data and within a shorter period of time than a human can. If implemented cautiously, generative
AI could bring about a significant decrease in the rate of escalated complaints and time spent by cardholders.
Barriers for artificial intelligence to overcome
Predictability and explainability remain significant hurdles in the path of AI and its adoption across industries. Predictability is achieved once an LLM produces the same output when prompted with a certain set of inputs, whereas explainability is achieved
once an LLM’s output can be explained in a way that makes sense to a human.
Financial institutions are required to demonstrate to auditors how every decision was taken, making the current absence of predictability and explainability a fatal flaw for LLMs in the payments industry. This predicament was encountered by the insurance
industry in the past decade; AI-driven agents demonstrated better accuracy than human operators, but insurance firms were not able to move the virtual agents into production due to regulatory shortfalls.
Artificial intelligence, already in use throughout the card payments value chain, has demonstrated its value in the payments industry. The leap taken by ChatGPT and other LLMs demonstrates great potential to vastly improve many payments processes,
but previous challenges persist: banks and their regulators aren’t yet ready for the implementation of advanced AI.