RAIN / AI Fraud Detection
Payment fraud is not a single class of event. The Federal Reserve’s voluntary FraudClassifier distinguishes who initiated a payment and how the event occurred; the FBI describes business-email compromise that exploits apparently familiar requests. A proposed RAIN study would evaluate specific review signals against those distinctions, rather than promising one universal fraud score.
Exploratory use case
THE FRAUD RESEARCH / PAYMENT DECEPTION QUESTION
A CASE WORTH RECONSTRUCTING
In a fictional historical example, an accounts-payable employee receives a plausible supplier message asking for a change to the bank details used for a forthcoming payment.
A proposed review signal brings the changed recipient details and available request history together. The study compares this case with both confirmed deception and legitimate supplier changes; it does not assume that every change is fraudulent.
The responsible team checks the request through its established verification process and determines what action is appropriate. The study neither blocks the payment nor contacts the supplier.
Illustrative research design informed by public fraud typologies. RAIN has no implemented fraud-detection product, demonstrated detection rate, or automatic payment intervention.
Fraud research / payment deception / PROPOSED WORKFLOW
Start with a bounded task such as supplier payment-change review, and include legitimate changes as well as resolved fraud cases.
Distinguish confirmed outcomes from allegations and unresolved cases; record the evidence and review date behind each label.
Reconstruct only the records that existed at the chosen decision time so later investigation findings cannot leak into the proposed signal.
Compare signals at an agreed review capacity and inspect false alerts, missed cases, and differences between fraud mechanisms.
Assess whether the evidence helps the responsible team ask the right verification question, while leaving live payment controls unchanged.
THE INPUTS THAT MATTER
Choose a source to see the context it adds to this proposed review.
Approved records of the payment instruction and changes to recipient details, considered alongside legitimate business changes in the same sample.
SCOPED INPUT · HUMAN REVIEWPermitted metadata and correspondence showing where a payment-change request originated and what evidence of verification was available at the time.
SCOPED INPUT · HUMAN REVIEWThe institution’s available record of who initiated the payment and relevant access events, without treating successful authentication as proof that the request was honest.
SCOPED INPUT · HUMAN REVIEWConfirmed case classifications, review dates, and unresolved outcomes; an uninvestigated event should not silently become a trustworthy negative example.
SCOPED INPUT · HUMAN REVIEWWHAT A PILOT SHOULD PROVE
Suggested evaluation criteria. Set the baseline and acceptance thresholds with the sector team before a trial.
THE OPERATING BOUNDARY
The proposed scope excludes blocking transfers, freezing accounts, assigning liability, or contacting customers. A signal is a prompt for human review, not a finding of fraud.
Later chargebacks, complaints, and investigation notes can reveal an outcome that was unavailable earlier. Keep those labels separate from the inputs used in a historical replay.
Account compromise and deception of an authorized person need different context. The FraudClassifier is a voluntary taxonomy, not a legal liability framework or an endorsement of RAIN.
RESEARCH & CONTEXT
Primary sources reviewed September 2026. These explain the sector context; the scenario and pilot measures are RAIN proposals, not reported customer results or source endorsements.
Federal Reserve Banks / FedPayments Improvement
FraudClassifier Model: Frequently Asked QuestionsThe voluntary model classifies fraud by the initiating party and mechanism, distinguishing authorized-party events from unauthorized-party events without determining liability.
Federal Bureau of Investigation
Business Email CompromiseThe FBI describes deceptive requests that appear to come from known business contacts, including changed payment instructions, and the importance of independent verification.
03 / A FEW DETAILS
No. The current implementation is a plant simulator. This exploratory concept does not include live account actions, payment blocking or automated enforcement.
No. An unusual pattern is a prompt for review. The proposed study would evaluate it against appropriate records and human investigation outcomes.
AI Fraud Detection / EXPLORE THE FIT
Start with the records, the review team, and the evidence a useful outcome would need. RAIN’s current implementation is a battery-storage simulator; this sector workflow requires its own validation.