As generative artificial intelligence advances, payments companies are looking to the emerging technology to save them money and make the payments process faster and more efficient.
Some of the giants of the payments industry, such as Fiserv, Visa and Global Payments say they are already using, or plan to incorporate artificial intelligence into their business.
Sid Tiwari, a partner at the consulting firm McKinsey who specializes in banking and payments, says there is another major application for artificial intelligence: fraud prevention.
Generative AI can use large language models — a vastly advanced version of the program on your laptop and smartphone that guesses the next word when you type an email or compose a text message — to learn from large data sets and be trained to recognize patterns, Tiwari said. If a bank account is showing unusual activity, such as a number of unusually large withdrawals, an AI program can flag those transactions as potentially fraudulent.
Tiwari spoke with Payments Dive about the potential applications of artificial intelligence and some of the specific uses of the ascending technology.
Editor’s note: This interview has been edited for clarity and brevity.
PAYMENTS DIVE: How are payments companies using artificial intelligence?
SID TIWARI: Machine learning has been used for a long time and large companies have used it to detect fraud. What is happening now is it's much more accessible to get advanced [AI] models and easier to apply AI across the spectrum. You can use it at [a payment’s] origination time to make sure there are no errors, or apply it to create instructions for a settlement, or use it for generation reports and reconciliation. Back in the day, it was much more focused on specific things, it was much more rule based. Now, you can apply it across the board.
How is artificial intelligence going to change the payments industry?
I have been in the industry long enough to tell you developing a model used to take months. You had to build a model and then apply it. Today, you could take some data, put it into a model. For example, you can tell the model 'these are the patterns I'm looking for, such as making the transactions too often, flag it as fraud,' and that's enough for the model to classify it as fraud. The timeline to develop something has gone very narrow. It's become easier to make progress in terms of implementing some of these AI features. Back in the day, it was much more complicated, you had to do more math. Now you can apply this across the board, and reduce the cost of making payments, by reducing payment fraud.
How do these models work?
You need data that shows a pattern. So you could generate a data set, and what it could do is simulate that [payments] environment so the model can be tested properly. You have the model act like someone trying to do a fraud. Imagine how much easier it becomes to go from zero to 100, and now you can launch [the fraud prevention guardrails] much faster.
What safeguards need to be in place?
Three key guardrails need to be in there. The first is a human in the loop. You cannot have an AI algorithm out there alone. There needs to be a checker on the process, and the checker needs to be a human. Number two, there are algorithmic ways to generate interaction. I can tell the model any transaction that is above $100 and not part of a softl-ine product must be classified as fraud. The model will check the context, check the outcome and make a decision that is related to that context. The last one is simply doing periodic audits to make sure [the system] is performing at the level at which you want.
What are some other ways artificial intelligence might improve the payments landscape?
I'm seeing a huge amount of investment in building specific models which will do specific actions. Models are being trained so they can automatically process invoices. You could see small use cases which are completely automated end to end by large language models. As you know, when we started [generative AI] was all Q&A, but we’re seeing companies are evolving and moving toward decision-making. These models are not Q&A bots anymore. They feed into the decision-making process, providing information which could help you and me make a decision.
How can AI reduce those costs?
During one day you might not have many payments happening, but then the next day is Christmas, which is different because there are more payments happening. In the future, you can simulate [and learn from] those kind of load scenarios and that would help large payment companies become more efficient
