Feedzai

Scoring 120 billion events a year in real time is not achievable without models, and the company has now gone further by building a foundational model purpose built for financial risk data spanning fraud, money laundering and wider crime decisions across the whole lifecycle.

Its chief science officer articulates why this domain resists borrowed approaches, noting that the next transaction is far less predictable than the next word in a sentence and that financial risk is adversarial because fraudsters adapt in real time. A rules layer exists and is deliberately retained, but it is combined with models rather than substituting for them, and rules alone are what the platform positions against.

Control sits with the institution to an unusual degree. Customers deploy their own models, author custom rules, import and export machine learning models, incorporate third party scores and tune the system to their own risk appetite and regulatory requirements, so the decisioning logic is neither opaque nor owned solely by the vendor. Integrated case management gives analysts a working surface and explainability is treated as a product property. What is not described is the automation boundary: no stated threshold at which a transaction is blocked without review, no sampling of automated decisions and no escalation architecture.

Three properties give a validator real material. Explainability is treated as a differentiator rather than an afterthought, transparency and robustness are named pillars of the published framework, and most usefully an institution can import and export machine learning models, meaning a bank can bring a model its own risk function has already validated rather than accepting the vendor's.

The performance claim for the foundational model is stated comparatively and conditionally, which is more honest than a bare superlative. Still missing are published accuracy or recall figures, model documentation and a validation package.

The scale figures are the largest in this index and are stated consistently across sources: roughly 9 trillion dollars of payment volume secured annually, about 120 billion events risk assessed, and more than a billion consumers protected, across banks, payment networks and acquirers. A major British banking group is named as a customer with its economic crime prevention platform lead quoted on the record describing years of collaboration on artificial intelligence.

Eighteen years of operation, availability through a major cloud marketplace and recognition on a national innovation ranking round it out. What is absent is per customer outcome measurement, with no published reduction in losses or false positives attributed to a named institution.

Feedzai publishes a named framework for trustworthy artificial intelligence built on transparency, robustness, freedom from bias and security, and supports it with research led commentary on responsible use in regulated finance, which is a more considered posture than the category norm.

It is also unusually direct about the cross institution question, stating that its foundational model matches bespoke supervised models on a single customer's data and surpasses them when trained on data from several institutions and geographies. That is the clearest statement in this index that pooling customer data improves the product, offered as a benefit rather than buried. What remains undefined is the boundary itself, beyond the assertion that network insights are anonymised and aggregated.

The data footprint is the widest in this index by volume, covering transaction records, behavioural biometrics and device intelligence for more than a billion consumers across multiple jurisdictions, and feeding a shared network layer on top. Network derived insights are described as anonymised and aggregated, which is a stated position rather than a documented method. No published privacy framework, retention schedule, subprocessor list or cross border transfer statement was located in this pass.

No trust centre, enumerated certification list, attestation scope or audit period was located in this pass. Customers include tier one global banks, payment networks and acquirers, all of which impose extensive assurance requirements before transaction data moves, and listing on a cloud marketplace carries its own publisher review, so the control environment is certainly substantial and simply not surfaced. The grade records what a buyer can verify without entering procurement.

Feedzai supplies technology and holds no licence, the expected posture, and it occupies the dual position seen among the intelligence vendors in this index, selling to the largest banks and to the regulators who oversee them. Product scope maps onto anti money laundering obligations, onboarding requirements and the scam liability regimes now reshaping instant payments in several markets, and the platform is designed to be configured against each institution's own regulatory requirements. No individual supervisory instrument is named as a design target and no formal admission process is evidenced.

This is the strongest fairness position in the fraud lane of this index, where the norm is silence and two comparable vendors grade D. Feedzai names freedom from bias as an explicit pillar of its published trustworthy artificial intelligence framework, publishes material on fairness in automated decisioning, and engages with the ethical scrutiny the sector is under rather than avoiding it. The limit is that a framework is not a result.

No demographic accuracy, no decline rate analysis by population or geography, no independent audit and no fairness testing outcomes were located, so a buyer gets a stated commitment without evidence that it holds.

No accuracy guarantee, remediation commitment or published error rate was located, and the exposure is proportional to the scale: a platform scoring transactions for more than a billion consumers in real time will decline legitimate payments and freeze legitimate accounts at volume. The affected consumer has no relationship with the vendor, is not told which system produced the outcome, and has no described route to see or contest it. Scam liability regimes are also shifting reimbursement obligations onto institutions, and nothing states how the vendor's judgement interacts with that allocation.

Feedzai builds its own foundational model rather than adapting a general purpose one, which is an explicit and unusually consequential in house claim, and it describes the training basis in real terms covering onboarding, digital activity, payments, transfers and anti money laundering workflows across its own network. Supporting imported customer models means part of the chain can belong to the institution outright.

What is not published is the fourth party layer: no subprocessor list, no infrastructure detail beyond the marketplace relationship, and no account of which institutions contribute to the shared network a buyer's scores will draw on.

Integration reaches every channel and rail rather than one entry point, covering card, instant transfer, digital and onboarding flows in one omnichannel platform, with the newer network score delivered through a single interface and listed on a major cloud marketplace so a smaller bank can adopt it without an enterprise programme.

The platform enhances existing decision engines rather than replacing them, accepts third party scores and imported models, and ships integrated case management so alerts land where analysts work. Named individual connectors are not published, which is the one gap.

Delivery is cloud native, available directly and through a major cloud marketplace, operating across geographies for global banks and payment networks. Residency deserves more attention here than for most, because the company states that its foundational model performs better when trained across several institutions and geographies, which implies data movement between jurisdictions as a design property. No hosting regions, residency options, tenancy separation or transfer mechanisms were located.

No rates, tiers, billing unit or minimum were located. One route deserves note as a partial exception: the network scoring service is offered through a cloud marketplace, which is a procurement path that sometimes carries listed pricing, and it is positioned for banks large and small, implying a lighter commercial entry than the enterprise platform. Neither is quantified publicly.

Coverage spans the payment ecosystem end to end rather than one seat in it, addressing banks explicitly at both large and small scale, payment networks, acquirers, processors and merchants, and extending to the regulators who supervise them. Risk types are enumerated in operational terms including card fraud, account takeover, scams, money mule activity, bust out and chargeback fraud for acquirers, alongside anti money laundering and onboarding. Reach is global across geographies and payment rails, which few vendors in this lane can claim.