3 Generative AI Use Cases in Banking and Payments

Fabiane Ziolla Menezes

Content Strategist

Contact the author

Practical examples of how companies in the banking and payments sectors are using generative artificial intelligence have begun to proliferate. As we showed in this special series, players are experimenting with AI models and their applications as they explore how to build the right technology and governance structure for them. They are doing this primarily through partnerships with big tech companies dominating the field, from infrastructure providers – developers of cloud services and CPU chips – to startups.

As the Bank for International Settlements (BIS) explained in its latest Annual Economic Report¹, “the ability to bring mathematical order at scale to everyday unstructured data, whether they be text, images, videos or music,” has turned AI models “into highly refined prediction machines,” capable of detecting patterns in data and “fill in gaps.” These are key characteristics to improve operational efficiency and “lower costs in back-end processing, regulatory compliance, fraud detection and customer service.” These are precisely some of the primary areas where financial institutions and fintechs are experimenting with AI.

Most Common Use Cases of AI
in Banking and Payments

Improving fraud detection

Large language models (LLMs), a type of foundation model on which other systems can be built, can look at a huge volume of data of various kinds and make not-so-obvious associations between them. This capability is allowing financial companies to move from basic machine learning to GenAI-driven features to detect fraud.

JPMorgan Chase is one of the banking giants making this shift. As described² by the bank’s Global Head of Payments, Trust and Safety, Ryan Schmiedl, the institution is doing that by examining patterns in emails for signs of compromise. The executive explained that this is key to identifying threats as fraudsters trying to create deceitful emails often use different patterns from that of the corporation’s employees. The bank uses only data within its ecosystem to train the LLMs so as not to generate false positives.

Mastercard is also using GenAI to boost its fraud detection capabilities. According to the company, fraudsters steal millions of payment card numbers through various criminal tactics and place part of these numbers on illegal websites to try to sell this data to other criminals. Mastercard claims³ it can now predict all the details of these compromised cards, alerting and enabling banks to block them much more quickly than before. It is doing that by scanning transaction data across billions of cards and millions of merchants “at faster rates than previously imaginable.”

The company’s new feature is powered by a proprietary AI model, built in-house and trained on data from the roughly 125 billion transactions that go through Mastercard’s network annually. Mastercard’s feature uses the cardholders’ visit history to merchants to determine whether the business involved in a transaction is a place the customer would likely go. The technology generates a score by creating “paths” through Mastercard’s network to find an answer to this question. The lower the score, the greater the chance of fraudulent purchases. All this process happens in 50 milliseconds⁴.

Boosting productivity and code development

GenAI technologies can also work as assistants to employees, helping them do their jobs more efficiently by, for example, searching and synthesizing documents. That is the case with Deutsche Bank, which has been testing Google Cloud’s LLMs at scale⁵ to provide new insights to its financial analysts. The bank has now reportedly developed a proprietary solution capable of answering questions about publicly traded companies and creating summary documents to advise its clients.⁶ And this is only one of the use cases under development in the Frankfurt-based bank, which has about 16,000 employees working on tech, data and innovation areas – about half of whom are engineers working on code⁷.

Coding is another area where AI models can work alongside humans. As we explained in the second article of this AI series, Big Tech companies have enabled platforms aimed at developers so they can access, test, and fine-tune foundation models to their programming needs. That is the case with Microsoft’s GitHub Copilot service and Google’s recently launched Gemini Code Assist⁸. Australia and New Zealand Banking Group (ANZ Bank) has recently conducted a six-week trial of GitHub Copilot with 1,000 of its 5,000-tech staff (software engineers, cloud engineers, and data engineers). Between June and July, they were separated into two groups, one using GitHub Copilot’s Visual Studio Code extension and the other not, and asked to complete six algorithmic coding challenges per week using Python. The test evaluated not only the impact on productivity but the sentiment of the engineers while working with an AI assistant.

The group with access to GitHub Copilot completed their tasks 42 percent faster than the other.  Also, the code produced with the AI assistant had fewer code “smells” (meaning any trace of deeper problems in the application or codebase) and bugs on average.  The overall sentiment of the staff toward the technology was also positive. “They felt it helped them review and understand existing code, create documentation, and test their code; they felt it allowed them to spend less time debugging their code and reduced their overall development time; and they felt the suggestions it provided were somewhat helpful, and aligned well with their project’s coding standards,” says the trial report⁹. With that, ANZ Bank decided to deploy the AI generative programming assistant for all its staff.

Recently, Nubank, the largest neobank in Latin America and one of the largest in the world, has stated that it is shifting its business model to be “AI-first.” In recent interviews, Nubank CTO Vitor Olivier said that each company area will have a specific “AI copilot” and that some are already testing it. In June, the neobank bought Hyperplane¹⁰, a San Francisco-based startup created by Brazilians that is building foundation models for banks. With it, Nubank expects to develop its own models and applications to boost areas such as risk management, collections, and marketing.

Another key area for this acquisition, according to the company’s co-founder, David Vélez, is credit¹¹. He believes that it is where the smallest AI-driven improvement can make a stark difference. He says that many of the already known machine learning applications failed to advance in this area because they perform analysis based on regressions, that is, with an eye on the past.

Vélez said that Nubank carried out tests with Hyperplane’s neural networks and managed to prove that the startup’s modeling actually outperforms the neobank’s internal algorithms. An important metric to evaluate this was the so-called ‘area under the curve,’ which measures how much the model can predict default. According to Vélez, this metric improved using Hyperplane’s algorithms.

Personalizing customer experience

That is perhaps the area where financial institutions and companies have ventured first with GenAi, given LLM’s capabilities towards mimicking human conversation. In March 2024, Latin America’s largest digital wallet, PicPay adopted OpenAI’s conversational application ChatGPT for its customer support channels¹².

The implementation began with a small number of customers and gradually expanded to the entire customer base — the digital wallet has over 35 million active users. During the testing period, the net promoter score for customer service increased by 45%, and the complaint resolution rate was 20% higher. PicPay says more than half of the users served by AI have resolved their demands without having to speak to a person.

According to the company, the tool can answer questions about products and services, with information fed by the user experience team. It has also learned about customer behavior to respond to demands in a contextualized way, bringing personalization to service.

Unlike previous chatbots, programmed to answer generic questions, GenAI can understand more specific questions, such as “How much did I spend on electricity on average in the last few months?” or “How much did I pay David last week?” because she can identify the electricity costs and who David is on the customer’s contact list. The expectation of financial institutions and fintechs training these AIs is that, over time, they will be able to answer specific questions and provide increasingly complex guidance.

One of the US’ leading automotive lenders, Ally Financial, has also deployed AI to improve its customers’ experience, but from solutions built in-house. In June 2023, the financial holding company launched a proprietary, cloud-based AI platform called Ally.ai, where it keeps its foundation models and applications. Ally then ran its first successful use case, using Ally.ai to transcribe and summarize calls answered by over 700 of its customer service staff. The company says¹³ 90% of the summaries using Ally.ai require no further human modification. As such, they are instrumental in helping associates provide quicker responses to customers. More importantly, the staff reported that Ally.ai summarizations have helped them focus their energy on meaningful customer interactions.

Sources

Fabiane Menezes

Fabiane is a journalist with more than 15 years of experience reporting on business, finance, innovation, and cities in Brazil. At PCMI, she keeps an eye on revolutionary fintech movements and the businesses behind these big changes in Latin America.

See Full Bio