Plaid launches credit, fraud and payment risk AI models

Plaid just launched new AI models for credit, fraud and payments, trained across its network of more than 12,000 financial institutions and millions of consumers. Early testing showed a 20% lift for deep-subprime borrowers, up to 40% improvement in fraud detection, and 26.5% more ACH returns prevented.

Plaid notes that millions of US adults fall outside of the traditional credit lens, making them difficult to evaluate with traditional scores alone. To help with this, the firm is tapping into cash flow data, which shows how people earn, spend, and pay their bills, giving lenders a fuller picture of their ability to repay.

Instant Link enables borrowers who connect their financial accounts to Plaid Consumer Reporting Agency to consent to share their cash flow insights for future credit applications, creating a faster experience. Lenders have access to those insights in under two seconds.

Then, the new Plaid LendScore 2 core model helps lenders make decisions. The company claims that the model is 42% stronger at predicting a borrower’s ability to repay a loan compared to traditional credit data alone. New specialised models extend those gains for specific industries such as auto and home lending.

In addition, LendScore Arc is a transformer-based credit risk model that learns from the order and timing of a borrower’s transactions. In early testing, Arc delivered “significant predictive lift” over the core model, including 20% lift on deep subprime and 24% lift on superprime borrowers.

Away from credit, Plaid has rolled out an AI foundation model purpose-built for fraud. Trained on patterns and hundreds of millions of data points from the Plaid Network, it analyses the full sequence of events rather than a snapshot. In internal evaluations, it delivered up to 40% relative improvement over previous baselines.

Plaid’s sequential foundation model also powers Signal, an ACH payment risk model which reads an account’s transaction history in sequence to better predict payment risk. In testing, the sequential foundation model helped Signal prevent 26% more ACH returns without increasing false flags.

The company also announced Plaid Layer, which is Plaid’s instant onboarding product. Instead of asking every new user to type in their name, address, and bank details, Layer checks whether someone’s already on the Plaid network using just their phone number. If they are, we’ll let them share identity and bank information they’ve already saved with Plaid in a single tap. The Layer network now has over 67 million remembered users – 49% growth from this time last year.

For eligible users, that means skipping the cumbersome parts: retyping an address, digging up a bank login, re-uploading a document. For the businesses using Layer, it means higher conversion – anywhere from 5% to 25% uplift.