.png)
Large language models (LLMs) are powerful: They can reason through problems, analyze sentiment, code software, and more. Yet they are also limited in some fundamental ways. In particular, they struggle with certain areas of data analytics, which means they can’t be straightforwardly applied to some of the tasks for which enterprises might like to use them.
A wave of companies in the finance sector is trying to tackle the problem with foundation models trained in-house on transaction data. Mastercard is one of them. Its “large tabular model” (LTM) is based on the same “transformer” architecture that underpins LLMs, but is designed specifically for tasks relying on tabular data, with use cases spanning payments, personalized retail, and security. We spoke to Steve Flinter, Distinguished Engineer, Artificial Intelligence at Mastercard’s AI Center of Excellence, about its model’s potential.
What made you decide a foundation model was necessary for Mastercard?
The lightbulb moment was seeing what was happening with LLMs and how they were continuing to evolve. What was particularly compelling was the idea that a single model, or a relatively small number of foundation models, could power a wide range of use cases across the business rather than building a separate model for each one, which is traditionally what you have to do.
Let’s take an area like fraud. What is it your model can do that previous systems couldn't?
Imagine someone is buying a wedding ring. That purchase often comes with a broader pattern of legitimate behavior, like booking a restaurant or buying flowers. Someone using a compromised card may show a very different pattern, such as repeated high-value purchases that don’t reflect typical customer behavior. What we hope, and this is the hypothesis, is that our model will be able to detect the differences between those types of patterns.
But how is that different from what you could do with traditional machine learning?
The LTM can learn patterns across transactions more holistically, rather than being constrained to a specific, pre-defined correlation. This allows us to explore and surface signals in the data that traditional models might miss.
LLMs predict the next word. Could you use an LTM to predict a customer’s next transaction?
Yes. That capability is useful for counterfactual analysis: given a string of transactions, what might the next ones look like? If we predict one type of transaction, and we see something that's radically different, that deviation can be a powerful signal for detecting potential risk or fraud.
But aren’t you worried about hallucinations—models making things up?
In this space the idea of “hallucinations” doesn’t really apply, as what we can see in the transaction stream only reflects a small part of someone’s behavior. So the idea of 100 percent accuracy with any model is inherently unrealistic in this space.
Do you think we're heading towards a world of many domain-specific models, or will the general-purpose models ultimately absorb this type of capability?
My view is it's probably going to be both. The LLMs are not going anywhere, but I do think there is a space for domain-specific, and industry-specific, and even company-specific LTMs, because the type of data that companies can access is simply not available to the big LLM labs.
What I would expect to see over the medium- to long-term is that companies like ours that have the resources in terms of the human capital, the talent, the know-how, and that have the data and access to the necessary compute, will start to build their own models that are relevant to their industry or their specific use cases.
Is that the breakthrough that enterprise gen-AI is looking for, in terms of diffusing across more of the enterprise and delivering value?
It's certainly a valid hypothesis. These models are designed to extract more value from data and operate more like an insights engine. So, instead of building and maintaining thousands of narrow models, you can create a smaller number of foundational models that power a wide range of use cases across an enterprise. That’s a meaningful shift in how organizations scale AI and deliver impact.
Do you think making a foundation model will eventually become an option for smaller businesses with less resources?
Over time, yes. While building these models today requires deep expertise and infrastructure, rapid advances in tools, platforms, and partnerships are making them increasingly accessible. What’s exciting is that we’re moving toward a world where organizations of all sizes can tap into increasingly powerful AI capabilities, whether it’s by building, customizing or leveraging models to turn data into better decisions, stronger security and more personalized experiences.
Discover more from Payments Unbound:
