Underwrite.ai

There is nothing here but the model, which makes this among the cleanest results on this axis in the index. The company builds a custom machine learning model from a client's own historical loan and application data, maintains it, and returns a decision through an interface. It does not sell a workflow, an origination system, a document pipeline, a data asset or a case management layer, so there is no residue to survive removal: take the model out and the company has no product at all.

The technique is named rather than implied, with gradient boosting identified explicitly and the argument for it made in technical terms, contrasting logistic regression over fifteen to twenty variables with a boosted approach across hundreds that captures threshold effects and interactions automatically. The explainability method is named as well. A vendor willing to state which algorithm class it uses, and to explain why it beats the incumbent method, is describing a model company rather than a software company that added models.

Full automation, presented as the achievement, with no published floor beneath it. The decision returns in two to three milliseconds through an interface, which by construction leaves no room for a person in the path, and the flagship case study describes a lender adopting the machine learning model as its sole underwriting methodology, with the dramatic default reduction attributed to that adoption.

Sole is the operative word: the human underwriter was removed entirely and the result is offered as the reason to buy. The material gestures at a boundary without defining one, noting that staff who previously reviewed routine applications can concentrate on exceptions that genuinely require human judgement, which implies exceptions are routed somewhere but never says what makes an application an exception, who sets that threshold, or whether the model itself identifies cases it should not decide. Across two passes nothing published describes an override path, a confidence threshold, a manual review requirement, or any decision the vendor declines to automate.

Standard model metrics published against a named baseline, which is rare enough in this index to carry the grade on its own. The Korean market study reports area under the receiver operating characteristic curve for both approaches, 0.906 for a logistic regression the company describes as already well tuned with carefully selected variables, against 0.958 for the boosted model.

That is the discriminatory power measure a credit risk function actually uses, disclosed for the challenger and the incumbent, with the incumbent described as a fair fight rather than a straw man. The installment lender case gives a before and after on a defined outcome measure. The technique and the explainability method are both named, and ongoing monitoring with revalidation against actual outcomes is described as part of the service.

Against that: no sample sizes, observation periods or out of time validation are stated, the separate claim of prediction accuracy above 98 percent is loosely specified and sits oddly beside the precise area under curve figures, and no model documentation package is published.

Specific results attached to no identifiable institution. The published outcomes are described with more methodological care than most vendors manage, covering an installment lender whose first payment default rate of 32.8 percent and overall defaults above 60 percent fell to 8.5 percent early default under the model, a Korean market comparison reporting area under curve figures for both the incumbent and the challenger, and a United Kingdom auto lender moving from days to milliseconds.

Every one of them is anonymous. Across two passes no customer is named anywhere, no case study carries an attributed quote or a named executive, no customer count or decision volume is published, no funding is disclosed, and no analyst recognition or independent review was located. The company operates as a series limited liability company, a structure that discloses little. The result is a record where a reader can assess the method in some detail and cannot verify that anyone is using it, which is the inverse of most large vendors in this index.

Two disclosures worth more than most of the safety pages in this index, both of which cost the company something. The first is explainability described as a method rather than a promise: Shapley additive explanations named outright, with a worked example decomposing a prediction into per variable contributions, and a statement that every decision carries documentation of the factors that drove it.

Naming the technique lets a buyer's own quantitative staff evaluate whether it suits their adverse action process. The second is rarer. The company publishes the conditions under which its product is the wrong purchase, stating that lenders without historical performance data may need to start with industry standard approaches, and that where existing scoring already separates good from bad borrowers well the gains from machine learning may be small. A vendor telling a prospect not to buy is a signal no marketing claim substitutes for. Absent across two passes: model card, red team result, incident disclosure and any acceptable use boundary.

One relevant design choice and little documentation around it. The design choice is anonymisation at intake: the company states that clients provide anonymised loan and application data from which the custom model is built, which means the training corpus need not carry identifying borrower detail and reduces what the vendor holds about individuals. For a model development business that is the right architecture, and stating it is better than leaving it implicit.

Around that, a privacy policy and terms of service are published and reachable without contact. Across two passes nothing further was located: no description of what data accompanies a live decision request as against the training set, no retention schedule, no subprocessor list, no data processing terms, and no reference to the Gramm Leach Bliley Act despite a customer base of community banks and credit unions that the statute governs directly.

Nothing was located. Across two passes no certification, attestation report, trust centre, security page, penetration test summary, subprocessor list or security contact address appears on any surface, and no framework is named anywhere in the published material. That is the complete absence this grade describes rather than a thin or stale disclosure. The requirement is not theoretical for this business model.

A lender must transfer its historical loan and application records to the vendor for model development, and must then route live application data through the vendor's interface for every decision, so the vendor sits inside both the training pipeline and the production credit path. The anonymisation stated at intake reduces the sensitivity of the training corpus and is worth crediting, but it says nothing about the live decision path or about the vendor's own controls. Community banks and credit unions, the segments this company targets, are examined on third party risk management and would need this before onboarding.

An unregulated supplier with genuine command of the rules that bind its customers. The published material engages the substance rather than gesturing at compliance: it states the requirement that a declined borrower be told why, identifies black box models as creating a compliance problem for that reason, describes the prohibition on deciding by protected characteristic, explains disparate impact as the analysis that detects indirect discrimination, and works through how a facially neutral variable becomes a proxy.

It also publishes a use case addressed specifically to regulatory compliance audits. That is more regulatory content than most vendors in this lane put in front of a reader. It remains literacy rather than standing. The company holds no licence and claims none, and across two passes no supervisory examination outcome, no citation of the governing statutes by name, and no position on the artificial intelligence rules emerging in the jurisdictions where it has published work was located.

The strongest fair lending disclosure located in this sweep, and the only vendor graded in it to publish a capability rather than an assurance. Three specifics carry the grade. The company provides disparate impact analysis tools that generate reports comparing outcomes across demographic groups, positioned so a lender can identify and address bias before deployment rather than discover it in an examination.

It states that its models are designed to avoid variables that proxy for protected classes. And it explains the proxy problem correctly and concretely, using postal code correlating with race through historical housing segregation as the worked example, which demonstrates the mechanism rather than asserting awareness of it. Paired with named explainability serving adverse action requirements, this is fair lending built into the product.

What keeps it out of the top band is that the capability is handed to the lender and no results are published: no testing outcomes of the vendor's own models, no model card, and no independent fairness audit was located across two passes.

Published contractual terms and a trial that reduces commitment, with the substantive allocation unaddressed. Terms of service and a privacy policy are published and reachable without contact, which is more than several much larger vendors in this index offer, and a 30 day free trial lets a lender test performance on its own portfolio before committing, which is a meaningful practical protection against the model simply not working.

Across two passes no warranty, indemnity, liability cap or service level was located. The uncovered exposure is shaped by the service model. This vendor does not supply a tool the lender operates; it builds and maintains the model, so the quality of the artefact is its work product rather than the client's configuration of it, and nothing published states what a lender is owed if a custom model underperforms in production, degrades without detection, or produces outcomes that fail the disparate impact testing the vendor itself supplies the tools for.

Training provenance stated plainly, which is the disclosure that matters most for a bespoke modelling service. The company states that each model is built from the client's own anonymised historical loan and application records and calibrated to that specific portfolio and lending environment, so a buyer knows exactly what the model learned from: their own book, not a pooled corpus assembled from other lenders and not a pretrained artefact of unknown origin.

That answers the question a credit risk function asks first and that many vendors leave deliberately vague. The technique is named as gradient boosting and the explainability method as Shapley additive explanations, both standard and inspectable rather than proprietary black boxes. Input data categories are described by type, covering bureau records, bank transaction data, utility and rent payment history and employment verification. What is not named is any specific supplier, library, framework or infrastructure provider, and nothing describes what happens to a client's training data or model when the relationship ends.

A single well specified integration point and nothing published behind it. The product is consumed as an interface call returning a decision in two to three milliseconds, which is the right shape for embedding in an origination flow and fast enough to sit inside a live application without the borrower waiting.

The company describes the surrounding requirements candidly, noting that the model must connect to the loan origination platform, pull data from credit bureaus and verification services, and return decisions in a format downstream systems can process, and it identifies integration planning as a real implementation consideration rather than glossing it. A self serve trial provides a route in without a sales process. What is absent is anything an engineer can inspect.

Across two passes no public interface documentation, developer portal, sandbox, code repository or software development kit was located, and no origination platform, bureau or verification provider is named as supported, so a lender cannot establish before contact whether its own stack is covered.

A hosted interface service with nothing published about where it runs. Across two passes no statement was located on hosting arrangement, cloud provider, available regions, tenancy, or any data residency commitment, and no customer hosted or in perimeter option is described.

The gap matters more than it would for a purely domestic vendor because of the published footprint: work is described across the Korean credit market, a United Kingdom auto lender, rural Mexico and Philippine provinces, each of which carries its own rules on where lending and personal data may be processed.

The training arrangement raises the question in a second form, since building a custom model requires the client's historical loan and application records to move to the vendor, and while those records are stated to be anonymised, nothing describes where they are held during and after model development, how long they are retained, or whether they are destroyed when a client leaves.

A named unit and a named rate, which almost nothing in this index offers. The company maintains a public pricing page and the rate located is two dollars per transaction for a credit decision returned through the interface, disclosed by the company itself in a directory profile as a core differentiator. A per decision unit is exactly the right shape for this product and lets a lender model cost against application volume before any conversation.

A 30 day free trial is offered with a self serve start, and terms of service and a privacy policy are published and reachable. Two qualifications belong on the record. The specific rate was located on a vendor supplied directory profile rather than read from the pricing page itself, so a buyer should confirm it is current.

And the commercial model involves custom model development and ongoing maintenance performed by the vendor for each client, which is professional services work whose treatment under a per transaction rate is not described anywhere.

Segments addressed on paper, none evidenced by a customer. The company publishes distinct use case material for small and midsize business lending, start ups and new lenders, underserved markets, community banks and credit unions, auto and mobility lenders, peer to peer platforms, established lenders, and regulatory compliance audits, which is a wide and coherently chosen set for a firm of this size and shows a deliberate focus on lenders too small to build modelling teams.

Geographic range is genuinely unusual for a small vendor, with published work spanning a Korean market study, a United Kingdom auto lender, and credit risk modelling for populations with little or no bureau data including unbanked borrowers in rural Mexico and less developed Philippine provinces.

What is missing throughout is a named institution in any of those segments or geographies, any count of customers, and any statement of scale, so the coverage claim rests entirely on the existence of the pages describing it.