Mastercard bets on new AI model to fight financial fraud
Mastercard just made a pretty interesting move in the fight against financial fraud. The payments giant developed a new artificial intelligence model called the Large Tabular Model, or LTM, designed specifically to analyze structured data and identify suspicious transactions with far more accuracy than traditional solutions can deliver.
If you ever wondered how a company that processes billions of transactions a day manages to separate the good from the bad in real time, the answer starts getting a lot clearer here.
The financial sector has been dealing with fraud for decades, and the sophistication of attacks just keeps growing. This is not the old story of someone skimming a card at a gas station anymore. Today the schemes are automated, distributed, and hard to track with conventional tools. That is where Mastercard steps in with a different approach, betting on an AI model that does not deal with text or images but instead works with the structured data that already exists inside payment systems. The ambition is real, and the details of how this works in practice, what the risks look like, and what regulation has to say about all of it are worth every paragraph ahead. 👇
What the Large Tabular Model is and why it is different
When most people think of artificial intelligence, they picture models that read text, recognize faces, or generate images. The LTM from Mastercard is a whole different thing. It was built to work with tabular data, those spreadsheets packed with columns and rows that describe every detail of a transaction: amount, time, location, merchant history, card usage patterns, purchase frequency, and so on. This is exactly the kind of information that flows through payment systems all the time, but it used to be processed by fixed rules and simpler statistical models that had limited performance when facing new fraud behaviors.
The core difference lies in how the LTM learns. Traditional fraud detection models rely heavily on rules defined by human experts, something like: if a transaction happens outside the country and the amount exceeds a certain threshold, block it. That works to a point, but fraudsters learn fast and adjust their attacks to get around those barriers. The Mastercard model goes beyond that because it learns patterns from a massive volume of historical data, recognizing subtle anomalies that a simple rule could never catch. It understands the context of a transaction in a much richer way, taking dozens of variables into account at the same time.
Another thing that stands out is the scale at which this model operates. Mastercard processes, according to the company itself, more than 143 billion transactions per year. That means any fraud detection solution needs to work in milliseconds, without slowing down the payment flow, without generating too many false positives, and without letting illegitimate transactions slip through. The LTM was developed with that reality in mind, optimized to run at high speed within the company’s processing infrastructure. There is some seriously sophisticated engineering behind something that, for the end user, just looks like the card working normally. 💳
Gradual rollout strategy and the risks involved
One of Mastercard‘s smartest decisions was not replacing all of its fraud detection systems with the LTM at once. Instead, the company adopted a parallel implementation approach, running the new model alongside its existing systems. That caution makes a lot of sense when you think about the risks of a broad-scope approach like a Large Tabular Model. If a widely deployed model has a flaw, the consequences can be systemic, affecting millions of transactions at once. By keeping the previous layers of protection active while the LTM is validated in production, the company significantly reduces the risk of a large-scale incident.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
This coexistence strategy between the new and the old also allows Mastercard to compare the LTM’s results with those of traditional systems in real time, identifying areas for improvement and tuning the model before expanding its role in block or approve decisions. This is a very common practice in critical systems engineering, where reliability needs to be proven before any definitive transition happens.
The company also has plans to expand the scale of data used in training the model and increase its overall sophistication over time. On top of that, Mastercard is planning to offer access through APIs and SDKs so that internal teams can build new applications on top of the LTM. That suggests the model will not only be used for fraud detection but could become a broader platform for transactional data intelligence, with applications that are still being explored internally. 🔧
How data fuels the model’s intelligence
At the heart of the LTM is the quality and quantity of the data it consumes. Mastercard has access to a volume of transactional information that very few companies in the world can match, and that is a huge competitive advantage in building any machine learning model. Every approved or declined transaction, every chargeback filed by a customer, every behavioral pattern of card usage over time, all of it feeds into the system’s training. With that level of richness in the data, the model learns to tell whether an unusual purchase is simply a vacation trip or whether it is actually a fraud attempt in progress.
One of the classic challenges in fraud detection is what is known as class imbalance. In the universe of transactions, fraudulent ones represent a tiny fraction of the total, which makes it very difficult to train models since the system can simply learn to classify everything as legitimate and still end up with a seemingly very high accuracy rate. The LTM was developed to handle this problem more effectively, using specific training techniques that give more weight to fraudulent transactions even though they are rare. The result is a model that manages to be more sensitive to fraud patterns without firing off unnecessary alerts for completely normal transactions.
Beyond that, Mastercard mentioned that the LTM has the ability to generalize across different markets and regions. That means a fraud pattern detected in European transactions can help the model identify similar behaviors in Brazil or Asia, even when local characteristics are different. This transfer of learning between distinct contexts is one of the most significant advances of the system, because modern fraud often spreads globally before local defenses can react. Having a model that learns across regions gives Mastercard a considerable speed advantage. 🌍
Regulation and the limits of AI use in payments
All of this ability to process data and make automated decisions does not exist in a vacuum. Regulation around the use of artificial intelligence in financial services is evolving fast in many parts of the world, and Mastercard needs to operate within those boundaries. In the European Union, for example, the AI Act already establishes clear requirements for AI systems considered high risk, and credit scoring or transaction blocking systems fall right into that category. That means the company needs to ensure transparency about how the model makes decisions, along with maintaining mechanisms for customers to dispute unjustified blocks.
It is worth noting that the article published by Mastercard itself emphasizes the responsibilities around the data the LTM handles. The company explicitly mentions privacy, transparency, model explainability, and auditability as pillars of the system’s operation. Regulatory scrutiny of any system that influences credit decisions or fraud detection outcomes is expected and will likely intensify as these models take on more responsibility within the payments chain.
In the United States, regulators including the Consumer Financial Protection Bureau and the Office of the Comptroller of the Currency have been advancing discussions about the use of AI in the financial system, especially in the context of fair lending, anti-money laundering, and consumer protection. The use of automated models for blocking transactions raises important questions about liability: when an AI model blocks a legitimate transaction and the customer suffers some loss because of it, who is responsible? Regulation is still being shaped to answer these questions clearly, and companies like Mastercard play an active role in these discussions since they understand the practical implications of these automated decisions in everyday payments better than anyone.
Another relevant point in the regulation discussion is data privacy. A model that learns from billions of transactions inevitably handles sensitive consumer information on a global scale. Mastercard needs to demonstrate that this data is treated in accordance with legislation like the European GDPR and the California Consumer Privacy Act, ensuring that the LTM’s training does not violate the rights of data subjects. The company says it works with anonymized and aggregated data in training processes, but the auditability of those practices is something regulators tend to demand with increasing rigor as these systems grow in scale and influence. 📋
Could tabular models be the future of banking AI?
One of the most interesting takeaways from this Mastercard initiative is the possibility that Large Tabular Models represent the beginning of a new generation of AI systems built for banking and payments infrastructure. While the tech market has been intensely focused on language models and image generation, the financial world fundamentally runs on structured data in tables. Amounts, dates, category codes, terminal identifiers, transaction sequences: all of it lives in tabular format, and building models native to that kind of data makes a lot more sense than trying to adapt architectures originally developed for text.
That said, it is important to keep expectations grounded. So far, the performance evidence for the LTM is limited to reports published by Mastercard itself, meaning the information comes from the technology provider. That does not invalidate the results, but it means the performance claims should not be considered conclusive until there is independent validation or large-scale production data available for external analysis. It is an important distinction that the original article makes a point of highlighting.
There are also significant practical challenges that will determine whether tabular models really gain traction in the industry. Robustness under adversarial conditions is one of them. Sophisticated fraudsters may try to manipulate their transaction patterns specifically to fool the model, and the LTM’s ability to withstand that kind of attack still needs to be proven at scale. The costs of post-training and long-term maintenance are also a relevant concern, since a model of this size needs to be continuously updated to keep up with evolving fraud tactics. And of course, regulatory acceptance will play a crucial role in defining the pace and extent of adoption for these systems. Ultimately, these are the factors that will determine whether tabular models become a central piece of financial infrastructure or remain a niche solution. 🧩
The real impact for everyday card users
For the average consumer, the most immediate effect of a system like the LTM should be a reduction in two problems everyone has dealt with at some point: having a legitimate purchase blocked for no apparent reason, or finding an unknown charge on a statement. Both sides of that problem are costly. False positives create frustration, especially at inconvenient moments like during a trip or an important purchase. False negatives, meaning when fraud goes undetected, cause direct financial loss. A more accurate model helps reduce both types of errors at the same time, which improves the card experience and lowers the operational costs of disputes and refunds.
Mastercard says that in LTM testing there was a significant improvement in the fraud detection rate compared to previous systems, without a proportional increase in wrongful blocks. If those numbers hold up in real production at scale, the impact could be quite substantial. Issuing banks that use the Mastercard platform will be able to configure their risk thresholds with more precision, adapting the model‘s behavior to the characteristics of their own customer base. That creates a layer of customization that more generic systems simply cannot offer.
It is worth remembering that financial fraud is not just a problem for big companies. Small merchants are also affected by fraudulent chargebacks, and consumers of all profiles have had negative experiences with unauthorized charges. A smarter payments ecosystem benefits the entire chain, from sellers to buyers. Mastercard‘s bet on the LTM is, at its core, an attempt to raise security standards systemically, using the scale and the data the company already has as raw material to build something that smaller players would have a very hard time replicating on their own. 🚀
What to expect going forward
Mastercard has made it clear that the LTM is just the beginning of a broader strategy. With plans to expand the data base used in training, increase the model‘s sophistication, and make tools available for internal teams to develop new applications, the company is positioning the Large Tabular Model as a central piece of its artificial intelligence infrastructure for the years ahead.
The payments market as a whole should be watching this move closely. If Mastercard‘s approach proves effective at scale, it is likely that competitors and other financial sector players will accelerate their own investments in tabular models. The race for high-quality structured data and AI architectures optimized for this type of information could become one of the main fronts of technological competition in the payments industry over the next few years.
For now, the LTM represents a well-grounded bet by Mastercard in an area that still has plenty of room to grow. The challenges of robustness, operational cost, and regulatory compliance are real and should not be underestimated. But the direction is clear: artificial intelligence applied to tabular data has the potential to transform how the financial sector handles fraud, risk, and real-time decisions. And whoever gets there first with a solution that actually works in production is going to have an advantage that will be very hard to catch up to. 🏁
