Pave

The removal test leaves the lender holding exactly what it already had. Raw bank transaction data, credit reports and loan performance history are inputs the customer supplies, and the entire product is the transformation applied to them: unifying, categorising and enriching that material into more than 4,000 attributes and into scores trained on repayment outcomes.

Nothing survives the models except the customer's own unprocessed data, and the company is unusually direct that it adds no data of its own, which sharpens rather than weakens the position, since analysis is demonstrably all it sells.

This vendor publishes a limit on its own authority and publishes it as a correction, which is rarer than stating a boundary unprompted. Its own material names the misconception directly, that Pave replaces proprietary models, and answers that the products drive lift in a lender's models and are not a replacement, adding that lenders retain full control of those models.

The claim is defended by architecture rather than policy, since the product surface is an interface returning attributes and scores and has no decision output at all, so there is nothing for a lender to over delegate to. That is the Recordsure pattern, a vendor declining credit for more autonomy than it has, and it is the correct division for this function because the accountable credit decision stays inside the regulated lender where the obligations sit.

The modelling design is itself the risk argument and it is a sound one. Scores are trained on actual loan performance history rather than on a proxy for creditworthiness, and they are built separately per credit product and per small business industry on the explicit reasoning that a generic score misrepresents both a seasonal restaurant and a trucking operation. Building narrower models against observed outcomes is the correct answer to a real failure mode.

The enhancement rather than replacement position also means the lender's own validation covers the combined system and lift is directly measurable by the customer. What is absent is the conventional evidence: no discrimination statistic, no backtest, no stability monitoring across the 4,000 attributes, and no documentation package a model risk function could review.

Volume is stated in a strong operational unit at 100 million credit risk evaluations a month, and the framing around it is informative in its own right, describing lenders moving from one time origination decisions to continuous risk evaluation across payments, limits and servicing.

Outcome claims are specific and, unusually, cover both sides of the underwriting trade off: approval increases around 80 percent alongside defaults reduced roughly 45 percent, which is the pairing that matters because either number alone can be achieved by moving the cut off. Partners are named including a decisioning platform also indexed here, a major cloud data platform and a lending marketplace. What is missing is a single named lender, so the customer base is evidenced by throughput rather than by identity.

Two mechanisms limit exposure and one question stays open. Customer data is described as anonymised, and the secure sharing architecture means analytics reach the customer inside their own governed environment rather than being extracted.

The open question follows directly from the product design: scores are stated to be trained on loan performance history for each credit product, and a model that distinguishes a fuel card from a personal loan needs repayment outcomes across many lenders to learn from, so a pooled training corpus is implied by construction. Nothing states whose performance data trains those models, whether contribution is a condition of use, or whether a lender can decline to have its outcomes inform scores sold to its competitors.

Three properties stack and two of them are architectural rather than promissory. The company states plainly that it neither provides nor aggregates data but sits on top of the customer's own sources, which limits what it holds in the first place. Data passed for analysis is described as anonymised.

And delivery runs through a major cloud platform's secure sharing mechanism, chosen for a reason the company explains: data never leaves that platform, stays encrypted at rest and remains subject to the customer's own already defined access control policies, replacing an earlier practice of moving large unencrypted files between machines. A vendor that publishes why it abandoned a weaker method is describing a real control. Held at B because no retention schedule, subprocessor list or deletion commitment was located.

No attestation, certification, trust centre or enumerated framework was located. Real security reasoning is published, in the account of why bulk file transfer was replaced with platform native sharing to keep data encrypted and inside customer controlled boundaries, and that is a control described rather than asserted. It is not an independent assessment, and a lender processing 100 million evaluations a month through a third party will need one before its own examiners are satisfied.

No supervisor, statute or instrument is named, and one specific question goes unaddressed that a lender's counsel will raise immediately. Assembling and furnishing information bearing on a consumer's creditworthiness for use in credit decisions is activity that consumer reporting law defines and regulates, and the company's repeated statement that it does not provide or aggregate data, sitting instead on top of the customer's own sources, reads like the boundary that keeps it outside that definition.

If so it is a meaningful and defensible position and it is never articulated as one. Nothing addresses reporting agency status, dispute handling obligations, or the adverse action requirements attaching to attributes that shape a decline.

The inclusion case and the proxy risk are the same fact, which is the position recorded for Omnisient. Cashflow underwriting is the best evidenced method available for extending credit to people a bureau file describes poorly, and the company's stated purpose is finding healthy underserved borrowers and approving more of them, which is a real and measurable social good when it works.

The same transaction data also encodes where a person shops, banks and worships, what they spend on healthcare, and who they send money to, so cashflow attributes carry proxy exposure to protected characteristics that a bureau score does not. One attribute family is named Willingness, and inferring willingness to repay from spending behaviour is a judgement about character rather than capacity. No fairness testing, disparate impact analysis or per population performance was located.

No guarantee, indemnity or falsifiable accuracy commitment binds the vendor. Two features work in the applicant's favour without being designed for them. Because the vendor supplies inputs rather than decisions, the accountable party remains the regulated lender, where adverse action duties already sit.

And because the output is thousands of named attributes grouped into affordability, stability, willingness and assets rather than an opaque single number, the material needed to construct a reason for a decline is at least present in the data. Neither is recourse. An applicant is not told that cashflow analysis shaped the outcome, cannot see which attributes counted against them, and has no route to correct a miscategorised transaction.

The data chain is disclosed by negation and the negation is unusually clean: the company states it does not provide or aggregate data and sits on top of the customer's own sources, naming those as loan performance outcomes, bank transaction history and credit reports. That closes the question most credit vendors leave open, because there is no bureau, broker or alternative data supplier in the path to disclose.

The infrastructure provider through which analytics are delivered is named, as is the decisioning partner. What remains undisclosed is the model layer itself, with no statement of what the scoring and categorisation models are built on, and no subprocessor list.

Ingestion is deliberately frictionless, with the customer passing existing data to the interface without transformation, which removes the usual first obstacle in a data heavy deployment. Delivery is the more interesting half: analytics are shared through a major cloud data platform's native sharing mechanism, so results land as tables inside the customer's own warehouse and feed model training, dashboards and campaigns without a file transfer.

A decisioning platform also indexed here is named as a partner for small business underwriting, the fifth recorded instance of an indexed vendor depending on another. No loan origination system, core banking platform or servicing system is named.

The delivery architecture answers part of this axis without being framed as a residency position. Because analytics are shared through a cloud data platform's secure sharing capability, the output rests inside the customer's own environment under its existing access controls rather than in a vendor held store, which is a materially different posture from an interface returning data the vendor also retains. Held at B because nothing is published about where processing itself occurs, no region selection is offered, and no residency commitment appears for lenders whose own regulators specify one.

No pricing, packaging or basis of charge is published and every route in is a demo request. The unit question has a clear shape given the volume the company reports, since a product measured in evaluations per month is presumably charged per evaluation or per band of them, and knowing whether lead scoring calls cost the same as underwriting calls would materially change how a lender designs its funnel. Nothing indicates it.

Coverage is deep within a defined lane rather than broad across institutions. Consumer and small business lenders are both served, across the full credit lifecycle from lead scoring through underwriting, limit setting, servicing and collections.

The differentiating depth is in the scores themselves, which are built per credit product rather than generically, so a charge card, a fuel card and a personal loan each get their own model, and small business scores are built per industry on the reasoning that a trucking business, a restaurant and an online retailer have entirely different cashflow shapes. That is real segmentation. The limits are that buyers are lenders specifically rather than institutions generally, and no footprint outside one market is evidenced.