Plaid Adds Transaction-Sequence AI to Cash-Flow Underwriting

Plaid launched LendScore Arc on October 6, adding a transaction-sequence model to its cash-flow underwriting suite. The concrete workflow change is that a lender can evaluate how deposits, bills, transfers and spending unfold over time, rather than relying only on a flattened set of account averages and ratios.

The Wall Street Journal reported that the launch accompanies a second version of Plaid’s LendScore product. Plaid’s consumer-lending materials describe Arc as a transformer-based credit-risk model and also list specialized scores for auto, home and short-term lending.

What the sequence adds

A traditional cash-flow feature set might summarize monthly inflows, expense volatility or ending balances. Plaid says its sequence approach instead learns the order, timing and relationships among transactions. The company’s technical overview uses the difference between a temporary shock and a recurring liquidity squeeze as an example of context that sequence can expose.

That additional context may matter for members with limited conventional credit history or income patterns that are poorly represented by a bureau file. Plaid says lenders can use up to 24 months of consumer-permissioned cash-flow data. Its API documentation places the reports in prequalification, underwriting, servicing and other permissible-purpose workflows under the Fair Credit Reporting Act.

Those descriptions are vendor claims, not proof that the model will improve approvals, pricing or losses at every institution. A credit union would still need to establish performance on its own products, members, policy rules and economic conditions.

The lender still owns the decision

The useful implementation path is a bounded comparison, not an immediate replacement of existing underwriting. Credit unions can start with one product and a defined population, run the new signal alongside current decisioning, and predefine the measures that would justify adoption: approval lift, loss and delinquency behavior, pricing consistency, manual-review volume and outcomes across protected groups.

A current Experian-hosted interview with Mission Federal Credit Union offers a relevant operating lesson. The credit union describes alternative data as a supplement rather than a replacement and recommends beginning with one pain point, one product and one measurable objective. Because the interview appears on a vendor site, its implementation observations are useful context rather than independent outcome evidence.

Model approval should also bind inputs to a permissible purpose and documented member authorization, test missing-data and account-linking patterns, and preserve a route for disputes. If Arc or any derived feature affects a denial or less favorable terms, the credit union must be able to produce specific, accurate reasons consistent with Regulation B; a vendor score or opaque feature label is not a substitute.

A controlled pilot, not a shortcut

Before launch, lending, compliance and data teams should agree on a model inventory record, version controls, performance and fairness thresholds, adverse-action mapping, monitoring cadence and a shutdown path. The credit union should also know what happens when a member revokes access, linked data are incomplete or the vendor changes model versions.

The control set aligns with the loan-pricing framework: separate eligibility, price and exception authority; monitor outcomes by segment; and keep a human escalation path. It also complements the VantageScore 4.0 transition analysis, where implementation discipline matters as much as the additional data.

LendScore Arc expands what cash-flow underwriting can observe. It does not transfer accountability for what the lender decides.