Carrington Labs

The company states its position directly: most lenders already have the data and the decision engine, and the gap is the model in between. It builds that model from the lender's own borrowers and repayment outcomes, using machine learning and explainable AI over transaction data. Its scoring product runs solely on bank transaction data rather than bureau files. The model is the product. Remove it and nothing is left, because the company sells neither workflow nor data.

The boundary is drawn clearly and repeatedly: the company does not replace origination, decisioning or servicing systems, its models supply decision-ready signals that sit alongside the lender's bureau data, policy rules, scorecards and existing models, and use is described as lender-controlled with decisions kept in the institution's hands. A second-look pattern is offered as a contained way to test cash flow data on borderline applications without rebuilding the origination stack. No review requirement or threshold is specified for the signals themselves.

Disclosure is detailed. The score is decomposed into five named behavioral categories: velocity, liquidity, stability, leverage and resilience. A credit officer can see what drives it rather than receiving an opaque number. The accuracy claim is stated against a defined baseline, at up to 30 percent above traditional credit models, and the evidence base is quantified as millions of loans and billions of data points.

Explainability is described as expanded in the second version rather than assumed. No validation methodology or discrimination statistic is published to support the accuracy figure, and there is no independent testing or monitoring approach.

The evidence base for the scoring product is stated specifically: the performance of millions of loans and billions of data points. The parent is a listed company with published financial growth recognition. One lender customer is named, a month to month car lease company, alongside three named technology partners through which the models are delivered. No deployment count, portfolio volume or customer outcome figure is published.

The position is mixed and no boundary statement resolves it. Bespoke models are built from each lender's own borrowers and repayment outcomes, which limits cross-client exposure, while the scoring product is explicitly trained on the performance of millions of loans pooled across the business and its parent's consumer lending operation. Nothing states what a lender contributes to that pool by adopting the product, or whether its borrower outcomes improve a score sold to competitors.

The consent basis is stated explicitly rather than assumed, with the platform described as turning customer permissioned transaction data into decision ready signals. That matters because bank transaction history is among the most revealing datasets about a person's life. No data processing agreement, retention schedule or subprocessor register is published, and nothing describes how long transaction data persists after a decision is made.

Lenders supply complete borrower repayment histories to train bespoke models and route permissioned bank transaction data through the platform. No control documentation accompanies either flow. There is no published attestation or certification, and no trust center or enumerated control set.

Models are described as built for each institution's specific regulatory requirements and designed for explainable, governed and lender-controlled use within the customer's compliance framework, and the protocol server exposes outputs built to those requirements rather than generic scores. No regulator, statute or lending rule is named in any market, so a buyer cannot see which regime the compliance framing is calibrated against.

One mechanism does real work here: model features can map directly to adverse action reasons, which is the operational form the fair lending obligation takes, since a lender that cannot say why it declined someone cannot comply regardless of model quality.

The inclusion argument is framed well, that thin-file customers lack traditional credit history but are not low-information, and the score is positioned as an inclusive measure reaching thin-file and no-file applicants underserved by traditional methods. No fairness testing, disparity analysis or approval outcome by borrower group is published.

No guarantee, indemnity or correction process is published. The applicant's position improves indirectly: because features map to adverse action reasons, a decline can be explained rather than merely issued. Nothing states whether a borrower can see which transaction patterns counted against them or correct misclassified transaction data. Nor does anything say whether they can contest a score derived from their own bank account activity.

The input is stated precisely rather than vaguely, with the score powered solely on bank transaction data, which tells a buyer exactly what the model does and does not consume and rules out the undisclosed alternative sources that make peer products hard to assess. Models are the company's own, developed within its parent group, so no external model dependency exists. No open banking aggregator or data access provider is named, and those intermediaries determine coverage and data quality.

Three delivery partners are named individually, covering a decisioning platform through which lenders gain end-to-end control of credit strategy, a major customer platform's sales cloud, and an origination system, so the models reach several different points in a lender's stack. A protocol server additionally exposes institution-specific model outputs directly inside agentic workflows, which is a distinct delivery path from conventional interface access. No core banking, servicing or bureau system is named.

No hosting provider or region is published, and there is no residency commitment or private deployment option. The company is based in Australia and serves lenders internationally, with three countries excluded from availability. That makes the question of where transaction data is processed a live one for any buyer outside its home market, and it is unaddressed.

The nearest thing to a commercial commitment is implementation speed: a tailored model piloted in days and a lender onboarded in weeks. That bounds the implementation cost without indicating the license cost. There is no published pricing, packaging or basis of charge.

Buyers span banks and non-bank lenders across consumer and small business lending, and coverage runs the full borrower lifecycle rather than the application alone, taking in underwriting, loan and limit sizing, risk-based pricing, post-origination limit management and early warning detection. Availability is global with three countries excluded. One customer is named and no market-by-market presence is evidenced.