- Diverse Industry Adoption: Major financial institutions like BNY, NatWest, and Barclays are actively integrating AI for various functions, including data sharing, customer service, and operational efficiency.
- Advanced Fraud Detection: The industry is shifting toward “payment foundation models” trained on massive, global datasets that can identify fraud patterns and consumer behaviors more effectively.
- Rise of Agentic Commerce: Future AI agents will leverage consumer data to autonomously predict needs and complete transactions, such as pre-ordering coffee or managing a weekly food shop.
Imagine you are walking towards a coffee shop to fuel up for an important meeting. A button appears on your phone. You hit yes. You walk into the shop, and your coffee order, exactly what you wanted, is handed to you. You walk back out. This isn’t Starbucks’ mobile app or Amazon Fresh. This, perhaps, is the future of agentic commerce.
At the 2025 Sibos conference in Frankfurt, Germany, Trade Finance Global (TFG) sat down with Georgios Kolovos, Head of Payments and Fintech leader at Nvidia, to discuss the integration of AI into the banking and payments sector, the diversity of uses for the technology, and how AI may revolutionise e-commerce.
How is AI being used in the banking and payments industry?
The International Data Cooperation (IDC) predicts that by 2030, AI will contribute almost $20 trillion to the global economy. Major banks have been quick to follow the trend. In the last six months, in spite of numerous new partnerships, financial institutions appear not to have coalesced around the most valuable areas for investment.
Last year, BNY launched an internal proprietary platform named Eliza, which can plug in with any language learning model on the market to increase data-sharing ease. In March, NatWest announced its partnership with OpenAI with a strategic focus on deploying AI to meet customer needs, including using digital and virtual assistants. Commerzbank partnered with Google Cloud and Microsoft the same month, integrating AI to increase cloud computing and data analytics. In June, Barclays announced a rollout of Microsoft 365 Copilot to 100,000 employees to give them access to AI agents and information. In September, Temenos announced the launch of an AI-powered platform to streamline payments and account services.
These initiatives, from data-sharing, customer needs, cloud computing, employee information, and AI platforms, indicate institutions in different stages of AI development. Yet, the announcements are indicative of the diversity of uses for AI models in the industry, from fraud, compliance, authorisation, and settlement.
Kolovos broke the stages of AI development for the sector into three main parts:
- Firstly, a physical transition from CPU infrastructure, used in traditional computing to perform tasks sequentially, towards GPUs, which can perform tasks in parallel, increasing the system’s efficiency.
- Secondly, companies could begin using graph neural networks (GNNs) that can model the connectivity between different data points, including payment transactions, to spot patterns and anomalies useful in fraud detection.
- Lastly, companies could train large language models (LLMs), such as OpenAI’s GPT-4, which can predict trends and patterns, from liquidity needs to consumer behaviour.
Evolving from fraud detection
Fraud detection has become a hot use for AI in the payments sector by facilitating real-time fraud detection in payments by picking up on anomalies and flagging them.
In May, Stripe announced it had built a Payments Foundation Model, an AI model trained on tens of billions of transactions that had decreased the detection rate for attacks on large businesses by 64 per cent “practically overnight”.
Foundation models differ from traditional AI models, which are trained on carefully curated datasets to perform one specific task, for instance, tracking payments made for one type of commodity between a small set of countries. Instead, foundation models are trained on huge, diverse data sets that encompass transactions for every type of purchase, internationally, across decades. These models can learn principles about human behaviour, market dynamics, and business interactions and can quickly be adapted for new tasks.
Kolovos said, “These new payment foundation models are becoming an interesting step change for payments companies or financial services institutions, because data then is not staying in silos for fraud or for compliance”.
This means the new payment foundation models could have ramifications far beyond fraud detection by enabling insights into the broader lifecycle of consumer and business customers.
A personal shopper? Agentic commerce
Until now, customer-facing AI in the banking sector has primarily taken the form of online chatbots or answering machine voice AI.
Yet, financial institutions have an information advantage in customer dynamics because they have a lot of consumer data. Every purchase and payment made through a financial institution provides data points for the bank, as does every sale and transaction.
Agentic commerce depends on training LLMs to predict future consumer behaviour and then using AI to complete the transaction.
If an institution holds information about consumer preferences and retailer options, an AI agent can complete the transaction between them. This takes us back to Kolovos’ coffee scenario. If an institution knows precisely what type of coffee you like, when, and what is available at the coffee shop, it can predict and pre-load the purchase. It’s not just coffee. Agentic commerce could manage everything from travel agents to your weekly food shop.
In a study released in June, Gartner predicted that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024. Additionally, 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024.
“I hope it’s not in four or five years. I would expect you to see a lot more coming in the next year, to 18 months, I would say,” said Kolovos.
The environmental challenge
One of the challenges associated with generative AI is its impact on the environment, particularly increased water needs and energy demands. In West Des Moines, Iowa, a data centre that serves OpenAI’s GPT-4 faced a lawsuit by residents in July 2022, which found the centre used 6 per cent of the district’s water the month before opening.
The more institutions that shift towards developing their own LLMs, the higher the environmental pressures from the technology.
Kolovos said, “AI is not a one-to-one replacement. When you use accelerated computing, what might take a server 24 hours may be reduced to five minutes. You need to look at the overall perspective.”
