Mastercard unveils new AI model built for the realities of payments

Mastercard unveils new AI model to push artificial intelligence beyond the chatbot boom and into the operational core of global payments.

Mastercard unveils new AI model 

Its latest move is the development of a new generative AI foundation model trained not on text, images or video, but on vast volumes of structured transaction data.

The company believes this approach could sharpen fraud detection, strengthen cyber defences and improve everything from loyalty programmes to small business tools.

The significance of the initiative lies in the type of model Mastercard has built.

While the current wave of artificial intelligence has been dominated by large language models trained on unstructured content, Mastercard has opted for what it calls a large tabular model, or LTM.

This is designed specifically for structured data environments, making it particularly relevant for payments, where information sits in rows, tables and highly organised datasets.

A different use of generative AI for financial services

Mastercard’s new model has been trained on billions of anonymised transactions, and the company plans to expand this to hundreds of billions of payments records.

Over time, it also intends to incorporate merchant location data, fraud signals, authorisation records, chargeback data and loyalty programme information.

In effect, Mastercard is building a payments-specific intelligence layer that could identify patterns across its network with increasing accuracy.

This matters because the payments industry generates enormous volumes of repetitive but highly nuanced data.

Traditional machine learning models often require extensive human intervention, with data scientists engineering features to help systems detect anomalies or suspicious behaviour.

Mastercard’s LTM is designed to learn more independently from raw structured data, potentially identifying relationships that human analysts may overlook.

Cybersecurity emerges as the first major use case

The clearest commercial application so far is cybersecurity. Mastercard says early testing shows the model outperforming standard industry machine learning techniques, particularly in reducing false positives.

That is a notable claim. False declines remain one of the most frustrating challenges in digital commerce, often affecting legitimate but unusual purchases.

A high-value, infrequent transaction such as the purchase of a wedding ring can easily be flagged by existing models as suspicious.

Mastercard argues that its new foundation model is better able to detect weak but meaningful signals in the data, allowing it to distinguish more effectively between genuine consumer behaviour and actual fraud.

If borne out at scale, that could improve both security and customer experience.

From thousands of models to a more flexible intelligence engine

The longer-term ambition is even broader. Mastercard currently maintains thousand’s of AI models across different markets, customer segments and use cases.

A sufficiently flexible foundation model could reduce that complexity, lowering maintenance burdens and creating a more adaptable system for deploying intelligence across the network.

That prospect helps explain why Mastercard is now building APIs and internal toolkits to make the model accessible across the organisation.

The company is not presenting this as a chatbot strategy. Rather, it is positioning the LTM as an insights engine for commerce.

In a payments industry increasingly shaped by data, speed and security, that may prove the more consequential application of generative AI.