GreenLyne

The search is the product. What is sold is the discovery of a loan size and price combination that makes a qualifying mortgage viable for a borrower a conventional credit box declines, and that combination does not exist without the model that finds it. The Loan-to-Cashflow metric replaces the credit score as the default predictor rather than supplementing it, and the platform describes agents evaluating each borrower individually rather than by segment.

Strip the models and there is no residual product, no data service and no workflow tool underneath. Consistent with the whole alternative data credit cohort in this index, where Zest AI, Scienaptic, Carrington Labs, Pave, Prism Data, credolab, GiniMachine and the corrected Uplinq all hold A on this axis.

The design choice here protects the applicant rather than the lender, which is unusual enough to be worth naming. The Automated Second Look takes an application the lender's existing system has already declined and searches for terms under which it works, so an automated process is inserted specifically at the point where the answer was already no. That is structurally the pattern Persona earned an A for, routing a would be failure into an additional path rather than letting the denial stand, applied to credit rather than identity.

The lender's own underwriting system remains the decision maker and GreenLyne supplies candidates and pricing to it. Held at B rather than A because nothing published describes the oversight mechanics inside GreenLyne itself, including what a lender reviews before acting on a surfaced borrower, or what happens when the search proposes terms the lender's credit policy will not permit.

Worth recording that this vendor does something the rest of its cohort does not: it makes a loss side claim at all. The standing finding against Zest, Scienaptic, GiniMachine, Carrington Labs and Uplinq is that an approval uplift claim is only half a result, because approving more borrowers is only good if the new loans perform, and none of them publish the loss side. GreenLyne claims a 23 percent improvement in loss prediction, which is the right kind of number.

It is then unmethodologised in every respect that would make it checkable: no baseline model, no out of sample test, no vintage curve, no delinquency comparison between borrowers approved through the Second Look and conventionally approved borrowers, and no definition of what improvement is measured against. Securitisation readiness raises the stakes, since a loss forecast that is wrong is priced into a security someone else holds.

No named lender, bank or credit union appears anywhere, and every figure is the company's own. The 23 percent loss prediction improvement carries no baseline, comparison model, sample or time period. The dataset claim of more than 18 million loans across national lenders describes what the models were trained on rather than who bought the result.

The company does appear in independent editorial coverage of AI and mortgage bias in Shelterforce, placed alongside a National Fair Housing Alliance and FairPlay AI study, which is genuine third party attention and is not a customer reference. The stated goal of powering one trillion dollars of mortgages to underserved first time buyers by 2030 is an ambition with no progress figure attached. One named institution reporting a measured result would move this axis a long way.

The stated training corpus is more than 18 million loans described as spanning all national lenders, and nothing explains where that came from, on what basis, or whether contributing lenders consented to their loan performance informing a product sold to their competitors. That is the same consortium question raised by FiVerity and Vyntra, where shared intelligence across institutions is the asset, but here it appears as a scale claim with no governance attached. Nothing states whether a client lender's own portfolio performance feeds models served to other lenders, which is the first question a bank's data governance team asks.

No privacy posture published, and the data being handled is at the sensitive end of consumer finance. The platform is built on real time cash flow data, which means transaction level bank account history revealing where a household shops, who it pays, whether it gambles, whether it receives benefits and where it worships, all inferable from a transaction feed.

That places the product inside Gramm Leach Bliley obligations and, because cash flow data is used to assess creditworthiness, likely inside consumer reporting duties as well. Nothing states what is collected, retained, how consent is obtained, or how a borrower withdraws it.

No SOC 2, ISO 27001, penetration testing programme or trust centre found. The data at issue is consumer bank transaction history and credit file information for identified individuals, which is among the most attractive material a lending technology vendor can hold, and no attestation about its protection is published.

No licence, regulated entity or supervisory relationship named, and the perimeter engagement is more visible than most without being met. The product is framed around qualifying mortgages, which is a defined regulatory category under the ability to repay and qualified mortgage rule carrying a specific legal safe harbour, and around Community Reinvestment Act goals, which is a supervised obligation of the bank buyer.

A system that determines whether a loan qualifies is operating on a regulatory definition. The comparison that makes the gap concrete is already in this index: Upstart obtained the first CFPB no action letter for fair lending in 2017 and demonstrated to that regulator that its platform did not introduce unlawful bias. GreenLyne makes claims in the same territory with no equivalent engagement on the public record.

Fairness is in the objective function rather than in a policy statement beside it, which is what separates this from the mission language that left Uplinq at C. The company describes generating fairness optimised qualifying mortgages, states that failing to make loans that should have been made is a failure of the same kind as making loans that should not have been, and names fair lending experts and community advocates alongside mathematicians and mortgage specialists in its own account of who built it.

The Second Look is itself a recognised fair lending remediation mechanism rather than a marketing construct, and the stated intent is to move past proxies toward individual assessment, which is the correct direction for disparate impact exposure. It sits with Stratyfy at B and below Upstart and Zest AI at A for exactly the reason recorded against Stratyfy: no outcome data by protected characteristic is published.

There is no approval rate comparison across groups, no disparity metric, no less discriminatory alternative search result and no independent fair lending audit, so a company selling fairness as the output does not show the demographic evidence that would demonstrate it. Two further gaps belong on the record. Cash flow underwriting can encode its own disparities, since income volatility and thin deposit histories track occupation and geography. And nothing describes how adverse action reason codes are produced when a cash flow derived assessment contributes to a decline, which is a requirement rather than a courtesy.

No warranty, service level or remedy published. The recourse position is genuinely two sided here and better than most in this category on one side of it. A borrower surfaced by the Second Look is better off than under the counterfactual, since the alternative was a decline that stood, so the failure mode most likely to harm an applicant is the miss rather than the false positive, and the miss is invisible to everyone including the lender.

The exposures that do need answers are unaddressed: a borrower approved at a size and price the search optimised toward has no way to know an external system set those terms, and the investor buying a security backed by loans this engine sourced is relying on a loss forecast with no published error characteristics and no stated liability if it is wrong.

The site states that individual borrower risk assessment is powered by frontier AI models, and no model, provider, version or hosting path is named. That phrase deserves attention rather than acceptance, because a frontier general purpose model is a different kind of component in a credit decision than a purpose built statistical model: it is externally controlled, versioned by its provider rather than the lender, retrained on a schedule nobody downstream sets, and not straightforwardly reproducible, all of which cut against the model documentation and change control a bank must maintain for models affecting credit outcomes. The 18 million loan dataset is the other half of the chain, and its provenance is stated only as spanning all national lenders, with no source, licensing basis or currency given.

The ambition is stated as an intelligence layer enhancing loan marketing, manufacturing and trading infrastructure across the industry, and the platform is described as end to end from origination through securitisation and investor reporting. What is missing is any named integration. No loan origination system, no core banking platform, no pricing engine, no document or verification provider is identified.

For a product whose whole mechanism is intercepting applications that another system has declined, the connection to that system is the load bearing integration, and how the Second Look receives a declined application and returns terms to the lender's workflow is not described anywhere.

No deployment model, hosting arrangement, tenancy structure, region or residency commitment published. Nothing indicates whether the platform runs as a hosted service the lender sends applications to, or can be deployed inside the institution's own environment, which is a live question for depository buyers handling borrower cash flow data under vendor management scrutiny.

No pricing, model, unit or structure published. The commercial shape is unusually unclear even by the standards of this axis, because the platform spans origination through securitisation and investor reporting, and nothing indicates whether it is licensed as software to the lender, priced per loan found or funded, or takes economics from the securitisation side. Those are materially different arrangements with different incentive implications, which matters for a product whose output is a recommendation to lend.

Banks, credit unions and non bank lenders are the primary buyers, with the site also addressing contractors, merchants and capital markets participants, which points at point of sale home improvement lending and the investor side of the securitisation chain. Product scope is mortgage and home equity credit in the United States, with the credit invisible and credit thin population, stated at more than six million mortgage ready households, as the defining segment.

Held at B rather than A because the additional buyer types are presented as site sections rather than described capability, and the substantiated coverage is depository and non bank mortgage lenders in one national market.