ThetaRay

Learning normal behavior from unlabeled data lets the models surface typologies nobody has described yet, which a rule written in advance cannot do. Strip them out and nothing is left, since static rules are what the product exists to replace.

Which steps an agent performs alone, and what stands between its output and a filing decision, are not set out. Alerts go to financial investigation teams rather than being actioned automatically, and Ray, the agentic investigation suite, evaluates counterparty behavior, checks profile alignment, correlates past alerts and recommends next steps while analysts validate and finalize every decision.

A model that derives its own definition of normal is harder to interrogate than a rule an analyst wrote, and supervisors increasingly want demonstrated effectiveness. Patents describe the method, but accuracy figures, a validation summary and model documentation are unpublished.

Onafriq, a pan African payment network connecting over 500 million wallets and accounts in more than 40 African markets, deployed ThetaRay in under three months and judged 83 percent of sender alerts and 94 percent of receiver alerts on complex cross border payments worth investigating. Shift4 increased productive alerts by 70 percent, and Payoneer and Santander are named for broader risk coverage and detection.

Model training goes unmentioned in the privacy policy, adversarial testing is not described, and whether detection learned at one institution informs models serving another is left open, a live question for a product that learns what normal looks like from each client's payment flows.

The privacy policy covers site visitors and service users, names Google Analytics and HubSpot among tracking tools and keeps data as long as needed without a stated period. A subprocessor list in the trust center names one party, Microsoft, for Azure cloud infrastructure, Entra ID and Intune. The data at stake is heavy: payment messages carry every party, intermediary agent and free text narrative, and correspondent work processes data about other banks' customers.

The audit period, auditor and certification scope are not shown, and the documents sit behind an access request. ThetaRay's trust center, which monitors its controls live, lists a SOC 2 Type 2 report and ISO/IEC 27001:2022 alongside the cloud and privacy extensions ISO/IEC 27017, 27018 and 27701.

Current regulatory framing is a strength: FinCEN's modernization agenda and the European Union's new money laundering regulation, with the Anti Money Laundering Authority set up to enforce it, both move the standard toward demonstrated effectiveness. As a technology supplier ThetaRay needs no license, and correspondent banking obligations are addressed directly.

Anomaly detection treats deviation from dominant flows as risk, so remittance routes, diaspora transfers and emerging market business payments draw extra scrutiny. Corridor or demographic error analysis is not published, and Screena's screening side adds the usual name matching asymmetry. The trust center lists ISO/IEC 42001, the AI management system standard, among the frameworks it tracks, with no certificate or certifier shown.

ThetaRay publishes no accuracy commitment, and consequences land on people outside the customer relationship: a sender whose remittance is held or a respondent bank whose correspondent link is cut is never told a model made the call. Inside the institution, every Ray decision comes with clear reasoning and audit ready documentation, so a result can be traced and contested.

Watchlist providers, the model behind the Ray agents and the storage and analytics vendors are not named. In house algorithms handle detection, screening came from acquiring Screena, and the trust center's subprocessor list names Microsoft for Azure infrastructure, a host rather than a model supplier.

An overlay suits institutions that cannot remove a long serving monitoring platform without revalidating everything downstream. Microsoft Azure partners with ThetaRay to deliver SaaS monitoring and Travelex to deliver SaaS screening, though no developer documentation is linked.

Hosting region, residency option and tenancy separation are unstated, with transfers covered only by European adequacy safeguards. The SaaS service runs on Microsoft Azure, which the subprocessor list names for cloud infrastructure, while data originating under one supervisor is analyzed under another and correspondent flows cross several regimes at once.

Engagement starts with a demo request and no billing unit is published, though for a product applied to cross border payment volume that unit is a buyer's first question.

Supervisors and the supervised both buy it, from banks, fintechs and payment providers to regulators, across more than 40 countries. Correspondent banking is the specialism, monitoring a counterparty's underlying customers without seeing them, while credit unions, insurers and capital markets are not addressed.