From grammar to fluency: Our foundation models now power Plaid's intelligence products

Plaid reports that its transaction and sequential foundation models, introduced in May 2026, are now powering intelligence products for underwriting, ACH payment risk, and cash advance decisions by learning patterns in individual transactions and longer-term financial behavior. The company says the models improved outcomes in testing, including identifying more ACH return-risk dollars, reducing default risk at fixed approval rates, and lowering cash advance losses, while requiring no integration changes for existing customers. To make sequential-model decisions interpretable, Plaid uses integrated gradients to connect transaction-level contributions to familiar reason-code categories, enabling LendScore Arc to provide ranked explanations for lending decisions and adverse action notices. Plaid also optimized real-time Signal scoring through model distillation, GPU-specific compilation, and lower-precision computation, reducing batch scoring latency from about 136 milliseconds to 38 milliseconds. Current reported results include a 20% increase in subprime borrower approvals for LendScore Arc, 26% more risky ACH dollars detected by Signal at a 1% decline rate, and 10% fewer lost dollars for Cash Advance Index, while future models are planned to incorporate balances, connection history, and financial product usage.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.