Silent Eight

The removal test leaves nothing, because the models are built individually for each client rather than configured from a common product. The company's operational model is constructing custom systems trained on the processes human investigators were already carrying out inside that specific institution, so the software replicates that bank's own reasoning and decision making from its historical case data and keeps learning from analyst feedback. Agentic components, machine learning and natural language processing carry the screening, investigation and closure work end to end.

The boundary is stated plainly and the architecture supports it. The platform is described as human in the loop, delivering real time auditable decisions, and the agents are positioned as acting as an extension of the team by automating tasks without replacing people, which is an explicit refusal of the position Arva AI takes in the same category. The chief executive frames the goal as responsible, auditable and human centred artificial intelligence.

What makes this more than a claim is that the models are trained to replicate the institution's own investigators, so the standard being applied belongs to the accountable bank rather than to the vendor, and quality assurance is a named module in the platform rather than an afterthought, meaning the system's own output is checked as part of the product.

The most complete set in this category and it works on four levels. Quality assurance is a distinct module within the platform, so the system's own decisions are checked rather than assumed correct. Performance benchmarking lets an institution compare its compliance operation against peers worldwide, which is external reference rather than self assessment and is rare anywhere in this index.

Explainable techniques mean each decision carries its reasoning, and outputs are described as auditable in real time. And the track record is unusually long, with agents running in live environments at global banks since 2018 and quantified results published across investigation time, manual workload and money saved. What remains unpublished is a direct accuracy or false negative figure.

Among the strongest evidence surfaces in this index. Three global banks are named as customers, two of them among the largest in the world, and the group chief executive of one is quoted directly saying the platform significantly enhanced the speed and accuracy of spotting financial crime risk across the transactions that bank undertakes, which is the most senior testimonial recorded here.

The relationship with the other extended in February 2026 to add transaction screening, and expansion of an existing deployment is stronger evidence than a new logo. Agents have run in live bank environments since 2018. Outcomes are quantified at investigation times down up to 60 percent, manual workloads down 70 percent, and more than 19 million dollars saved across three institutions by January 2025. Around 55 million dollars has been raised across seven rounds, and the company won two industry regulatory technology awards in 2025.

The architecture answers most of this axis by construction: models are built and tailored per client and trained on that institution's own case history, so a bank's system learns from its own analysts rather than from a corpus pooled across competitors, which is separation by design rather than by policy. Two statements leave the boundary partly open and should be read together with it.

The company describes its artificial intelligence becoming more advanced as it has onboarded more banks, and the platform offers performance benchmarking that lets an institution compare itself to peers worldwide, both of which imply something does travel between deployments. Nothing states what is common and what is contained.

No data protection agreement, retention schedule, subprocessor list or deletion commitment was located. The access required is unusually deep, since building a custom model demands the bank's historical risk management data and its investigators' past decisions, which the company acknowledges was the hard part of its early strategy. That material identifies both customers who were investigated and the staff who judged them. Nothing published describes how it is handled, how long it persists, or what happens to a trained model if the relationship ends.

No attestation, certification, trust centre or enumerated framework was located. Two of the world's largest banks have granted this vendor access to their historical risk management data, which is among the most demanding third party assessments in financial services and has evidently been passed, and none of the resulting assurance is published. For a company positioning itself as regulator ready and preparing for a public listing, publishing that control set would be consistent with the rest of its posture.

Regulatory readiness is claimed repeatedly, with the platform described as regulator ready, built to handle complex regulatory demands and designed to adapt to evolving global compliance frameworks, and no supervisor, statute or instrument is named anywhere.

The gap is narrower in practice than it looks, since deployment inside globally systemic banks means the technology has been examined by those institutions' own regulators as part of routine supervision, but that is scrutiny conducted privately rather than a disclosure a buyer can read.

Explainability is the central commitment rather than a feature, appearing consistently as transparent regulator ready artificial intelligence, explainable solutions, detailed insight through explainable techniques and real time auditable decisions, and third party analysis identifies the black box problem as precisely the risk the company positions against.

That matters here because a screening decision that closes an alert on a named person should be examinable, and explainability is the mechanism that makes any bias assessment possible at all. What is absent is the assessment itself. Name screening carries documented false positive differentials across naming conventions and transliteration, and no demographic testing, error analysis by population or independent evaluation was located.

No commercial guarantee or indemnity was located, and accountability is reconstructable to an unusual degree. Decisions are auditable in real time and carry explainable reasoning, so an institution challenged by a supervisor on why an alert was closed can produce both the conclusion and the logic behind it, and the quality assurance module means erroneous closures surface through the product rather than through a later examination.

That is a materially better position than a system returning a score. The subject of a screening decision has nothing, which is largely inherent to a regime that prohibits telling them, and the vendor adds nothing beyond it.

The disclosure that matters most for this product is made clearly: the training data is the client institution's own historical case records and its investigators' past decisions, so a buyer knows exactly what the model learned from and that it did not come from an external corpus of uncertain provenance. Models are built in house rather than adapted from a third party foundation.

What is not disclosed is anything else, with no model provider named for the language processing components, no watchlist or sanctions data source identified despite screening being a core function, and no subprocessor list or hosting arrangement located.

Integration is evidenced by deployment rather than by a connector list, with the technology described as designed for seamless integration and scalability and demonstrably running inside the alert and case infrastructure of several of the world's largest banks, including a 2026 extension into transaction screening at one of them. Replicating an institution's existing investigation process necessarily means connecting to the systems that generate and hold those alerts. What is not published is any named platform on either side, so a prospective buyer cannot establish what connecting involves without engaging.

No hosting provider, region selection, residency commitment or private deployment option was located. That is a notable omission given a customer base of globally systemic banks operating across many jurisdictions with strict and differing requirements on where compliance and customer data may be processed, and given that the custom model approach implies each client's data and trained model must be held separately somewhere.

No pricing, packaging or basis of charge is published. Third party analysis records the model as subscription based licensing of the compliance technology, which identifies the commercial shape without quantifying it, and the custom model approach means implementation effort almost certainly varies substantially by client. Nothing indicates whether charge scales with alert volume, users, models built or institution size.

The buyer profile is deliberately concentrated at the top of the market, serving tier one global banks and insurance companies, with named deployments spanning a British headquartered international bank, a second with an Asian and African footprint, and a Gulf institution, which is genuine geographic reach through very large customers rather than many small ones.

Functional coverage within financial crime is complete, running from name and transaction screening through investigation and decision making to quality assurance and performance benchmarking. The limit is the flip side of the strategy: a business built on custom models for the largest institutions does not naturally serve the mid market.