Refine Intelligence

Agentic AI investigates an alert, decides whether to contact the customer, reads stated intent and resolves or escalates the case. A structured digital outreach service would remain without the models, and much of the measured gain comes from replacing branch phone calls with consistent digital requests.

Initial investigation, outreach and resolution run automatically, with 88.6 percent of alerts resolved without a person in production and the rest escalated to analysts. Automatic resolution thresholds and any sampling of cleared alerts are not published.

Resolution in under a minute rather than 10 to 30 minutes is an operational figure, and each case lands in an audit trail. Accuracy and the false clearance rate, the error where a genuine scam gets resolved automatically, are unpublished.

Valley Bank's median AML inquiry resolution fell from 16 days to two minutes, with 85 percent of digital requests completed and front line outreach work down 60 percent. The case study gives no period, sample or method.

Intent aware models are trained on a proprietary dataset of customers' explanations of anomalies, which implies responses gathered across banks inform every bank's models. Retention of responses, and whether a bank can keep its customers' answers out, are not stated.

Bank customers who receive the messages are addressed by a privacy policy updated March 17, 2026, which limits phone number use to transactional alerts and sets out rights under the CCPA and GDPR. Source of funds explanations, the most sensitive data held, are left out, and no subprocessor list appears.

Industry standard procedures and commercially acceptable means are all the privacy policy offers on security, and no certification or trust center is named. That is thin for a platform messaging customers under a bank's name.

Outreach is designed to avoid tipping off and to support customer and enhanced due diligence, yet no regulator, statute or rule is named. Bank Secrecy Act confidentiality and Regulation E would be the expected references for a service contacting customers about suspicious activity and Zelle payments.

Every customer gets the same structured questions through one digital channel, a design choice that removes one source of differential treatment. Response and clearance rates by customer group are unpublished, though customers less able to respond digitally are exactly who automatic resolution could miss.

Flagged customers are asked by push notification, email or SMS to explain a payment's purpose or source of funds, and the answer can resolve the alert instead of leaving them waiting. What happens when an explanation is not accepted is not described, and no guarantee backs the outcome.

Google Analytics and the Azure host are named, while the model provider and any messaging provider are not. Origin of the proprietary explanations dataset, the differentiator, is undisclosed.

Working downstream of existing fraud detection with a zero integration method of pulling data, the service names no monitoring, case management or core platform. Developer documentation is not linked.

Microsoft Azure in the United States hosts the servers, per the privacy policy. Private or on premises delivery is not described for the SaaS service.

A demo request is the only path, with no price, plan or basis of charge. Charges could follow alerts handled, customers contacted or institution size, and value scales directly with alert volume.

Fraud and anti money laundering teams at banks are the buyers, with pages for check fraud, scams and Zelle fraud and an outreach product for due diligence requests. Credit unions and payment companies carrying the same obligations are not addressed.