Most fraud doesn't look like fraud in isolation. A login looks normal. A bank connection looks normal. A transfer looks normal. It's only when you line those events up, in order, across time, that the pattern shows itself. Most fraud models never get to see that lineup. They see one frame at a time.
For over a decade, the industry's answer has been feature-engineered ML like XGBoost: hand-crafted features taught a model which combinations tend to precede fraud. That works for fast, in-the-moment decisions, but it evaluates a snapshot, not the sequence that led there. Every new typology needs a new model, and labeled fraud data lags real outcomes.
Foundation models are the next frontier in AI, and they've already reshaped search and productivity. Now Plaid is building one for fraud: the Fraud Foundation model, the next chapter for Plaid Protect.
From Snapshots to Storylines
The Fraud Foundation Model is a transformer-based model pretrained on behavioral and transactional sequence data before it's pointed at a specific fraud task. Transformer-based is about architecture: the attention mechanism that weighs every event against every other, however far apart. Foundation model is about training paradigm: pretrained broadly, then adapted to many tasks, rather than rebuilt from scratch each time.
As Plaid's engineering team has put it: "Finance looks like a ledger, but it behaves like a language." Two accounts can look identical on paper while the order and timing of events reveal one is a normal recovery and the other a fraud pattern taking shape.
Why Plaid Can Build This
Plaid sees the same device, credential, and identity appear across thousands of institutions and apps. A single app sees one event; Plaid sees the pattern it belongs to. Traditional fraud detection today is siloed to one company's four walls, so activity coordinated across institutions is invisible to any one of them. A network built over more than a decade lets us train on the full shape of an attack, not just the fragment any single company sees.
Introducing the Fraud Foundation Model for Protect
Every generation of the Trust Index has gotten smarter by expanding its intelligence: identity, then network, then transaction signals. This is the next evolution, but a different kind of upgrade: instead of adding more signals, it changes how the model learns from them.
The Fraud Foundation Model has been trained on hundreds of millions of proprietary data points, then fine-tuned on customer-specific fraud labels. Because it pretrains on sequence data rather than confirmed labels, it picks up on warning signs ahead of the labeling cycle. The new foundation feeds into Trust Index's existing scoring framework rather than replacing it, and in internal tests it is showing 40% improvement in fraud detection performance over previous Trust Index models. For fraud teams, that means catching more fraud earlier, with less manual review and less friction for legitimate users.
If you want to see it in action, contact our team to learn more.