DataVisor

Unsupervised learning finds coordinated attacks nobody has labeled, graph analysis exposes rings across accounts and devices, and agents tune rules and draft suspicious activity reports. Rules and labeled models, the approach the company set out to displace, are what removal would leave.

Every agent output is reviewable before it applies, people review suspicious activity report submission, regulatory reporting and strategic rule changes, and access is permission based. A complete audit trail records each agent action, output and human review step, and rollback of agent actions came with the 2026 conversational agent.

Plain language reasoning accompanies every detection, rule suggestion and report draft, and an 87 percent false positive reduction with 100 percent true positive retention names both error directions. That figure has no method, data set or customer behind it, and no validation package is published.

NASA Federal Credit Union cut manual review alerts by 40 percent and review time by 85 percent while raising detection precision 20 percent, and TaskRabbit reports a 60 times efficiency gain. Executives at Galileo and One Financial are also quoted, and no story says when its baseline was taken.

Customer data is neither shared with nor used to train public models, a categorical line. Whether patterns learned in one institution's data inform detection at another, outside isolated deployments, is not stated.

The March 29, 2018 privacy policy covers website visitors only, with no retention periods or named subprocessors. Onboarding identity documents, transaction histories and investigation case files on the platform fall outside it.

SOC 2 Type II and PCI compliance are stated, while the CCPA and GDPR badges beside them are regulations rather than credentials. Audit period and report access are not given.

Bank Secrecy Act obligations frame the AML platform, with automated suspicious activity and currency transaction report filing to FinCEN and AI drafted narratives reviewed before submission. Batch or system to system submission is not named.

Clusters of legitimate customers who share a bank, device type or remittance corridor can be flagged by unsupervised detection, and a group flag is harder to interrogate than a rule. Analysis across customer populations and fairness testing are unpublished, and AI drafted suspicion narratives add exposure the subject never learns about.

Institutions can reconstruct, contest and reverse what the system did, but no guarantee or indemnity stands behind it. Customers blocked in real time or named in a suspicion report are given no route, and for a filing cannot be told at all.

Ownership of the proprietary, patented unsupervised engine is clearly disclosed, while the agents appear to rely on outside language models with no model provider named. Hosting and external data sources are also undisclosed.

Affirm's executive describes feeding the risk scores into online decisioning, and scoring runs under 100 milliseconds. Not one core banking, payment processing or identity system is named, and no API documentation is linked.

Isolated deployment options keep a customer's data separate, but nothing names the platform's host, region or residency terms. The privacy policy allows storage wherever the firm or its providers have facilities.

Onboarding checks, AML monitoring and case management are usually priced apart, and whether the unified platform is licensed as one or by module is not stated. Engagement runs through a demo request.

Banks and credit unions, fintechs and payment companies, and digital marketplaces each have dedicated material and named customers. Functions span onboarding, transaction monitoring and regulatory reporting in one lifecycle view.