Provenir

Genuine embedded machine learning across the product: real time graph learning for fraud and relationship profiling, per customer models built by an in house data science team, generative assistance for analytics and an agentic execution layer. Held at B rather than A on the orchestration test that separated Ondorse from CleverChain.

The company's own framing is that it consolidates data, models and agents into one governed environment, which makes the platform the container and the models the contents: strip the models and a low code decisioning engine with data orchestration, a rules layer and a prebuilt data marketplace still stands, which is substantially what this vendor sold before the AI layer arrived.

Better than the category norm and short of an A. Three controls are described: agents act within guardrails the customer defines, strategy changes are validated against real production data before going live rather than after, and integrated case management provides a referral and investigation route for decisions pulled out of the automated path. The pre production validation step is the strongest of these because it is a vendor built mechanism rather than a customer setting.

Held at B because the guardrails themselves are entirely customer defined with no published default or floor, and the human in the loop framing appears in launch coverage rather than as a described control on the vendor's own pages.

Strong on the customer's models and silent on its own, which is exactly the hyperexponential precedent and the second clean instance of it in the index. What is published is real and useful to a supervised buyer: model management with monitoring dashboards, explainability extended to cover Python and other model types, simulation and scenario modelling, and validation of strategy changes against real production data before deployment.

All of it governs models the customer built or deployed. Nothing describes the vendor's own layer: no documentation of the graph learning engine, the generative assistant or the agentic protocol, no versioning, no drift disclosure and no evaluation results. A platform that sells model governance is not thereby transparent about its own models.

Clears the bar several times over. Named customers with quantified outcomes: tbi bank reports a fourteenfold capacity improvement with decisions in milliseconds, and MTN reports a hundred and thirty five percent increase in high risk transactions stopped alongside a hundred and thirty percent increase in pre approvals. SoFi is documented as live on a student loan refinancing build within ten weeks.

Beneath the case studies sit platform scale figures of four billion decisions a year across a hundred and twenty customers, and four independent analyst placements from separate firms: Forrester strong performer in AI decisioning platforms, Chartis category leader in retail credit solutions, IDC MarketScape major player in decision intelligence platforms and a Datos fraud orchestration citation. All four evaluators are current and still operating, which the benchmark half life rule requires checking.

The nearly answer, and a distinction worth carrying index wide. The vendor states repeatedly that models are built from the individual customer's own historical data and outcomes and are not generic market models trained on industry averages. Phrased as a data commitment that would be a strong stewardship disclosure, close to the reference set answers.

It is instead phrased as a competitive claim about model quality, and it is never accompanied by a statement that customer data is excluded from shared or pooled training, retained only for that customer, or withheld from third party model providers, which matters more here than for most vendors because the platform integrates external large language models. An ambiguous claim earns nothing, and that has to hold when it costs a grade.

Above the norm on independent verification rather than on published detail. A third party privacy compliance verification seal is displayed with a checkable registry identifier, which is a privacy specific external attestation rather than the usual self assertion, and compliance and data protection are presented as a distinct platform layer rather than a policy page.

A published privacy policy, cookie policy, modern slavery statement and transparency in coverage disclosure sit alongside it. Held at B because retention periods, subprocessors and cross border transfer mechanisms are not published.

Three credentials in the site footer and two of them are independently verifiable in one click, which is unusually good presentation: an information security management badge linking through to the certificate itself, and a third party privacy verification seal linking to the verifying firm's register with a unique identifier.

Held at B rather than A because the third badge is an accountancy body mark carrying no report name, examination period or type, and there is no trust portal, no penetration testing disclosure and no subprocessor list. The contrast is worth recording, since the same footer holds the best and the weakest credential presentation on one row.

A software vendor with no licence, registration or supervised standing of its own. Regulated environments are addressed as a design target, with the platform described as transparent, compliant and audit ready, but the regulatory obligation sits entirely with the lending institution that buys it. No sandbox participation, no supervisory programme enrolment and no direct examination exposure.

More than silence and less than a disclosure. Bias mitigation in decision making is named as a platform capability and explainability runs through the product, and the company maintains a dedicated ethics and compliance page in its site footer, which is itself uncommon.

But no fairness testing method, disparate impact measurement, protected class handling or fair lending validation approach is described in the material reviewed, and a named capability without a described method is an assertion.

Queued check, cheap and specific: the ethics and compliance page has not been opened, and this is the highest value unopened page on this vendor, because a credit decisioning vendor is the standing test case for whether AI governance is becoming a purchasing requirement outside commercial insurance.

Nothing published on recourse for the declined applicant. For a platform whose agents execute credit decisions in real time, the relevant disclosures would be adverse action reasoning, the route by which a rejected consumer contests a decision, and how responsibility is allocated between the platform and the lender when an automated decline is wrong. None appears. The liability structure follows the collections pocket pattern: the lender is the regulated party and absorbs the statutory consequence of a model it did not build and cannot fully validate.

A counterexample to the standing rule that buying forces disclosure while building permits silence. This vendor plainly buys: its marketplace advertises prebuilt integrations to large language models and its launch coverage describes integrating leading public and private models with private instances available through a cloud provider's managed model service. Yet its own dedicated AI page names not one model, provider or version, referring only to large language models generically.

The disclosure exists in trade coverage of a product launch and not in the vendor's own material, which is the wrong way round. Queued check: the February 2026 launch coverage enumerates the integrated models and would move this grade if the vendor carries the same list itself.

The data marketplace is the real integration asset, offering prebuilt on demand connections to global identity, fraud and credit data sources so a buyer can add a bureau or an alternative data provider without building the integration. Low code drag and drop configuration is positioned as removing the dependency on the buyer's own development team. Held at B because the depth is toward data suppliers rather than toward systems of record: no core banking, loan origination or servicing platform connectors are named publicly.

Cloud native delivery is stated and private model instances are offered for sensitive workloads, reportedly hosted through Amazon Bedrock, which is a real deployment option. Held at C because data residency itself is undocumented: no regions, no countries, no residency commitments and no statement of where decision data is processed or stored. Local legal entities in the United Kingdom, Singapore, the United Arab Emirates, Brazil and Mexico imply regional capability but are not a published residency position.

Nothing published. No pricing page, no tiers, no unit or per decision rates, no minimum volumes and no indicative bands. Every route through the site ends at contact us, book a meeting or talk to an expert, and software directory listings carry no figures either. A buyer can learn nothing about cost without entering a sales process.

As broad as coverage gets in this index. More than a hundred and twenty financial services customers across over sixty countries, processing upwards of four billion decisions a year, with named buyers spanning a global bank (BBVA), captive auto finance (GM Financial), consumer and card lenders (NewDay, Novuna, Resurs), a credit union (Meridian), challenger banks (tbi, Bigbank), telecom operators (Telia, MTN), a Brazilian fintech (Jeitto) and a credit bureau (Dun and Bradstreet).

Separately addressed segments include banks, credit unions, fintechs, payments, auto financing, buy now pay later, credit cards, digital banking, embedded finance, retail lending and small and midsize enterprise lending, supported by legal entities on five continents.