Loxon Solutions

The Clearwater precedent, third application this session. A twenty five year old credit management suite where the lending origination, collateral management and workflow products are parameterised deterministic software, and the learned component sits in two of the seven modules: the early warning system, where the vendor's own company timeline dates the first artificial intelligence projects to a specific era of its history, and the collection platform, where behaviour analysis drives contact strategy.

Strip the models and retail lending, corporate lending, collateral management and the collection workflow all continue to run. The vendor also states plainly that an early warning prediction model may be expert based rather than statistical, which is the honest description of an optional analytical layer rather than a claim that models carry the product.

The human sits in the described path rather than beside it. The early warning system is documented as supporting the human workflow processes of monitoring, explicitly including manual decisions, customer relationship handling and root cause analysis, with built in automatic and manual workflows, scheduled recurring human events such as customer visits and calls, manual control of watch listed cases, controlled manual capture of qualitative information, and escalation and pre workout routing.

The system produces signals and classifications for a banker to act on rather than executing action. Held at B because no thresholds, confidence bands or limits on unaided action are published, and the division between what the model decides and what the officer decides is implied by workflow rather than stated.

The strongest axis on this vendor and a deliberate contrast with others graded the same week. Its early warning product page sets out the actual apparatus of credit model validation in the supervisor's own vocabulary: thorough backtesting including root cause analysis and signal significance testing, expert based validation covering plausibility checking and reject inference, statistical techniques such as rating migration matrices for low default portfolios, and an explicit discussion of the discriminatory power of a behavioural approach.

It also gives the institution full parameterisation to change the models and the qualitative inputs itself, which is the property that actually helps a supervised buyer validate what it deploys. Held at B rather than A because the vendor publishes no results, no performance figures for its own models, no versioning or drift disclosure and no external assessment, and because the passages read as methodological guidance rather than as a claim about tested implementations.

Two named engagements and no attributed numbers. Raiffeisen Bank Hungary is identified in the company's own history as the first customer, the institution the original retail lending product was built inside. The Egyptian Collateral Registry is named as a delivered national infrastructure project with a testimonial stating the system was developed and implemented to Egyptian law and regulatory requirements, and that Loxon remains its backbone.

Alongside them sit more than eighty banking groups and over two hundred staff. Held at B because no outcome figure attaches to any named institution, the testimonial carries no named individual, and the recovery and portfolio quality improvements the products are sold on are never quantified anywhere.

Silent. Nothing states whether customer credit and behavioural data is used to train or improve models, whether anything is pooled across the eighty banking groups on the platform, or how long data persists. Where models are described the framing is that the institution supplies its own data and parameterises its own models, which points toward per customer isolation, but that is an inference from product design rather than a published commitment and the axis does not credit inference.

Nothing published on data handling. No retention position, no deletion terms, no subprocessor disclosure, no statement on how borrower and delinquent account data is segregated between tenants in the multi tenant environments the company operates, and no privacy specific attestation. The collection platform reaches consumers directly through messaging, email and push notification, which raises consumer contact data questions the published material does not address.

No certification, attestation, audit report or trust page found in the material reviewed. For a vendor holding retail and corporate credit files and delinquency data for eighty banking groups, and running multi tenant cloud environments, the absence of any published information security certification is a notable gap rather than a neutral one. Queued check: cloud marketplace listings often surface a vendor's security and compliance attestations in a structured field even when its own site carries none.

A software vendor with no licence or supervised standing. It is closer to the regulatory machinery than most, having built a Basel calculation engine early in its history, publishing commentary aimed at Gulf central bank rulebook expectations and expected credit loss reporting, and delivering a national collateral registry for a state, but every one of those is a customer relationship or a content exercise rather than a regulatory position held by the firm itself.

Nothing at product level. The company runs a blog series on the European Union artificial intelligence regulation and on digital operational resilience, which is category explainer content aimed at its buyers rather than disclosure about its own systems, and the standing test applies: the blog explains what the rules mean for the category while the product pages describe the specification.

No fairness testing, no treatment of protected characteristics in default prediction and no governance framework appears in the product material, which matters because behavioural default models built on payment timing and spending patterns are precisely where proxy discrimination arises.

No recourse position published. The exposure is specific to this product set: an early warning classification moves a performing borrower onto a watch list and can trigger repricing, facility withdrawal or restructuring, and a collection strategy assigns contact intensity, yet nothing describes whether the borrower is told, how a wrong classification is challenged, or how responsibility divides between the vendor supplying the model framework and the bank operating it. The bank is the regulated party and carries it.

No model provider, base model or third party analytical engine is named as underpinning its own products. The one disclosed third party dependency runs the other way: the company implements Oracle's financial services analytical applications as a services line, which is somebody else's analytics delivered to a customer rather than a component inside the Loxon products. Where generative capability is discussed it appears in blog material rather than in a product specification naming a supplier.

The lending and collection products are themselves systems of record, and the early warning system is offered either integrated with the retail lending, corporate lending or collection modules or as a standalone deployment against another vendor's stack, which is a real architectural statement rather than a claim.

The company joined the Amazon Web Services independent software vendor programme in April 2026 and publishes products on that marketplace, and it delivers Oracle financial services analytical applications as an implementation partner, so it integrates with a named third party analytics estate. Held at B because no core banking connectors are enumerated and integration is described at the level of collecting from source systems.

The deployment half is answered and the residency half is not. Both software as a service and on premises versions of the collection platform exist as separately described products, the newer collection system is cloud native, the early warning module deploys standalone or integrated, and the company runs multi tenant cloud environments on a named public cloud with published operational engineering roles supporting them. Nothing states data residency, region availability, or how data is kept inside a jurisdiction, which matters for a vendor selling across the European Union, the Gulf and Africa simultaneously.

No pricing published on the vendor's own site: no rates, tiers, bands or module prices across a seven product portfolio sold in both software as a service and on premises form. Queued check that is cheap and specific and may well overturn this grade: the company lists at least two products on a public cloud marketplace, and marketplace listings routinely carry contract pricing that a vendor never publishes on its own site.

More than eighty financial brands and banking groups reported as customers across Central and Eastern Europe, the Middle East, Africa and Asia Pacific, with offices beyond Budapest in Romania and the Gulf. Buyer types span banks and leasing companies, retail and corporate segments, and both conventional and Islamic banking operations, which is a genuine capability distinction rather than a marketing line since Islamic products require different profit and default treatment. Held at B rather than A because there is no presence in the largest banking markets, no tier one global institution named, and the customer count is reported by the company without a directory.