--- title: Foundation Models in Fraud Explained: The Tech Behind ChatGPT, Applied to Fraud source_page: https://www.sardine.ai/media/the-saturday-fraud-strategist/episodes/foundation-models-in-fraud canonical: https://www.sardine.ai/media/the-saturday-fraud-strategist/episodes/foundation-models-in-fraud format: text/markdown date: 2026-11-07T01:01:00.000Z description: Are foundation models in fraud real or hype? How they differ from XGBoost, where they lift precision, and where data quality and latency still set the limits… --- **Quick links:** [Human page](https://www.sardine.ai/media/the-saturday-fraud-strategist/episodes/foundation-models-in-fraud) · [Home](https://www.sardine.ai) · [Customers](https://www.sardine.ai/customers) · [Blog](https://www.sardine.ai/blog) · [Demo](https://www.sardine.ai/demo) --- # Foundation Models in Fraud Explained: The Tech Behind ChatGPT, Applied to Fraud **Published:** 2026-11-07T01:01:00.000Z Are foundation models in fraud real or hype? How they differ from XGBoost, where they lift precision, and where data quality and latency still set the limits… Foundation models in fraud are on every conference agenda this year, and I keep hearing the same question. Is this the real thing or the next buzzword? When I landed at Money20/20 Vegas, I wanted to ask someone who builds with them every day. Pedro Sanzovo works on identity and fraud products, and has been in fraud his whole career. He started at Palantir, built a fraud company for merchants in Brazil, and sold it to Signifyd, where he led product work on machine learning models. Six months ago he joined Plaid. We kept this one short and practical. What is a foundation model, and how is it different from the models most of us run today? Where does it beat XGBoost, and where does it hit the same walls? And what do you do with an embedding when your team writes rules? Pedro had some good answers, and a few things I think you’ll want to hear. What you’ll hear in this episode: What a foundation model is, and how transformer models differ from the chained rules and XGBoost models fraud teams have used for years Why legacy models depend on handcrafted features and clean supervised labels, and why that is hard for smaller companies How financial data foundation models treat transactions as a sequence of events and predict what comes next What Pedro means by the precision uplift of up to 34% he is seeing, and why data quality still decides the outcome Why latency constraints, such as 300 milliseconds for card transactions, shape which model belongs where How foundation model explainability compares with XGBoost, and why credit and fraud face different expectations Why embeddings are hard to turn into rules, and how AI agents might help You should listen to this episode if you: Lead or work with a fraud data science team and are deciding whether to explore foundation models Run a smaller fraud team and struggle with labeling and feature engineering Want a plain explanation of how foundation models differ from XGBoost Need to weigh precision gains against latency and explainability Are curious how embeddings and AI agents could work together to produce fraud rules ### Episode notes & key takeaways From rules to XGBoost to transformers Pedro explains the history simply. For 10 to 15 years, fraud prevention ran on chained rules, and then XGBoost allowed far more data to feed small decisions with very low latency. That became the state of the art for about a decade. I pointed out that those models look a lot like rules from an algorithmic perspective, and Pedro agreed that they share the downsides. They are labor intensive to build, they need data scientists to code features, and they depend on clear supervised labels. If you label something incorrectly, it is trash in and trash out, and that is especially hard for smaller companies. The takeaway is that legacy models work, but they put the burden on people. What changes with a foundation model Transformer models back the LLMs we all use, and they work by treating information as a sequence and predicting the next most likely output. Pedro’s question is what happens if we do the same with financial data. Plaid and others now codify financial information as a sequence of events and create embeddings that replace handcrafted features. Teams rely less on feature engineering and need less precise labels, though some labels are still required. Pedro sees a precision uplift of up to 34%, which matches what I see in the industry. The same problem still exists, though. Poor data means poor results, and even something as simple as precise time stamps is easier said than done. These models can absorb more data, but we don’t have a limitless supply of clean data, and teams may need to start collecting data they never stored before. Latency, precision, and explainability decide where each model fits I asked how all of this fits with machine learning and the newer decision models. Pedro looks at target variables and constraints. The target is usually precision and recall, and the biggest constraint is latency. A card transaction decision has a budget of about 300 milliseconds or less, while onboarding or promo abuse leaves much more room. His view is that a good data science and data engineering team should explore the options, know what it is optimizing for, and let the data decide. If a model beats your precision under the same constraints, it wins. Explainability is his other concern. XGBoost never solved it fully, yet it is much easier to work with than a transformer. Credit has a regulatory mandate for explainability and fraud does not. The right answer depends on the use case. From embeddings to rules with AI agents I added that embeddings are a shortcut to feature engineering, but you can’t use them in rules. Pedro immediately thought of agents. His team is experimenting with feeding embeddings to agents and asking whether they can turn some of them into rules and find gaps. A human still needs to decide whether the work is good. Things are not getting simpler, and Pedro says they are getting more fun. I agree that we have more options and more toys to play with. Each one comes with tradeoffs that teams have to understand before they commit. ### Final takeaway Foundation models in fraud are the newest tool in the kit. They are not a replacement for good data. Pedro’s advice is to be clear about what you are optimizing for, respect your constraints, and let the results decide. If you are choosing between legacy models and foundation models, start with the quality of the data underneath. We’re only beginning to understand how these models change the work, and I think the best teams will treat them as a promising option to test and not a finished answer. Resources & links Not ready to stop the conversation about my, and hopefully your, favorite subject? Subscribe to The Saturday Fraud Strategist newsletter. Connect with Pedro Sanzovo | LinkedIn GM and Head of Fraud at Plaid Connect with Chen Zamir | LinkedIn Host of The Saturday Fraud Strategist Helping fintechs build smarter fraud defenses Co-author of “The Fraud Fighter’s AI Playbook”