Legal, Risk & Compliance · Module
Fraud & Risk Scoring
Explainable risk scores that hold risky Actions for review before money moves
Fraud & Risk Scoring scores sign-ups, orders, payments, refunds, expenses, vendor bank changes and agent Actions with explainable rules and models. Risky Actions are held for review before they execute.
What it does
Fraud & Risk Scoring is designed to score risky events across the platform with explainable models and rules, hold risky Actions for review before they execute, and manage fraud cases and blocklists. Covered events include sign-ups, orders, payments, refunds, expenses, vendor bank changes and Digital Employee Actions.
Signals and rules
The signal catalog draws on platform events: velocity, device, IP, geography, contact-point reuse, bank detail changes, KYC results and refund ratios. A configurable rules engine applies weights and thresholds and can be tested on historical events.
Scoring at Action time
Scores attach to Action Bus evaluation for configured Actions, with outcomes of allow, review or block. Blocking applies only where it is not adverse to a person, or where policy allows it. Every score shows its top contributing signals and rules. Held Actions wait in a review queue with context and an SLA, and a person approves or rejects.
Cases and lists
Fraud cases link events, parties, amounts, investigation notes, outcomes and recovery. Blocklists and allowlists cover contact points, devices, bank accounts, GSTINs and IP ranges, each with reason, expiry and approver.
GA
Tenant-trained or shared-feature models (no cross-tenant raw data) are evaluated, monitored for drift and registered in AI Assurance, and run in shadow before enforcement. Expense and procurement patterns such as duplicate claims, split purchases and vendor-employee bank matches, COD and RTO order risk for Commerce, and agent Action anomaly scoring for AI Governance kill-switch decisions are planned, with a feedback loop from case outcomes. Consortium signals and graph link analysis follow in W5, subject to verified data-sharing legality.
Planned for W4 (Year 3). Fraud blocks on individuals are always human decisions.
Features
What Fraud & Risk Scoring includes
Signal catalog
Velocity, device, IP, geo, contact-point reuse, bank detail changes, KYC results and refund ratios.
Rules engine
Weighted rules and thresholds tested on historical events.
Scoring at Action time
Scores in Action Bus evaluation with allow, review or block outcomes under policy.
Explainability
Top contributing signals and rules shown for every score.
Review queue
Held Actions with context, approve or reject and SLA.
Fraud cases
Linked events, parties, amounts, notes, outcome and recovery.
Blocklists and allowlists
Contact points, devices, bank accounts, GSTINs and IP ranges with reason, expiry and approver.
ML risk models
Tenant-scoped models with evaluation, drift monitoring and AI Assurance registration.
Expense and procurement patterns
Duplicate claims, split purchases and vendor-employee bank matches.
Order and agent anomaly scoring
COD and RTO order risk for Commerce; unusual agent behaviour feeding AI Governance.
Feedback loop
Case outcomes label events for rule and model tuning.
Workflow
Held refund
- 1
A refund Action is evaluated by the Action Bus
- 2
The fraud score and its reasons are attached
- 3
Above threshold, the Action is held in the review queue
- 4
An analyst reviews context and approves or rejects
- 5
Confirmed fraud opens a case and may add a blocklist entry with expiry and approver
- 6
The outcome labels the event for tuning
AI
AI skills and Digital Employees
Tenant risk models
Non-LLM tabular models with feature contributions, run in shadow before enforcement.
Case summaries
Summaries linking events, parties and amounts, processed in-region.
Plain-language explanations
Template-constrained explanations from rule and feature contributions.
Agent Action anomaly scoring
Statistical anomaly alerts for AI Governance.
Receipt duplicate detection
Perceptual hashing to flag duplicate expense receipts; no automatic rejection.
Digital Employees that use this module
They act through the same governed Actions as people; risky steps wait for approval.
Frequently asked questions
Will customers be blocked automatically?
Blocking is limited to cases that are not adverse to a person or where policy allows. Fraud blocks on individuals are human decisions.
Is data shared across tenants?
No cross-tenant raw data is used. Models are tenant-trained or use shared features only.
Can it score AI agent actions?
Yes. Agent Action anomaly scoring is planned to feed AI Governance kill-switch decisions.
See Fraud & Risk Scoring in your workflow
Tell us about your team and the outcomes you want. We will walk you through the apps, the Digital Employees and the plan for your industry.