System for Bringing Order Fraud Signals to Human Review

The problem

In day-to-day operations, order, payment, device and delivery signals arrive in separate records and files, preventing the team from reviewing the same case in one place.

As a result, the link to the source can be lost while a reasoned risk queue is prepared, leading to repeated checks and uncertainty about the basis for the decision.

Incomplete automation may overlook legitimate behaviour that appears unusual, while poorly controlled automation creates a risk that a legitimate order will be blocked.

What we set out to improve

• Time taken for the draft reasoned risk queue to be ready for review
• Accurate routing of records where legitimate behaviour appears unusual to the correct specialist queue
• Recording of suggestions amended or rejected by the risk operations specialist

How the system works

1. 01 Retrieve order, payment, device and delivery-signal data from an approved source
2. 02 Verify the source system, record identifier and freshness information
3. 03 Transform the fields into the target process schema
4. 04 Prepare a draft reasoned risk queue
5. 05 Compare the draft with business rules and source records
6. 06 Exception check: legitimate behaviour appears unusual
7. 07 Risk operations specialist approval, amendment or rejection
8. 08 Write the input, suggestion, amendment and final decision to the audit log

Methods we used

• Source-traceable data extraction for order, payment, device and delivery signals
• Structured output that restricts the reasoned-risk-queue fields through business rules
• An exception gate that separates cases where legitimate behaviour appears unusual
• A queue that shows the source and suggestion together for the risk operations specialist
• Duplicate-processing and rollback records to manage the risk of blocking a legitimate order

Where people stay involved

Data and security

Only the fields required for the task are processed from the order, payment, device and delivery signals. Source-system permissions are preserved, and personal and commercially sensitive data is minimised before being passed to the model. Every read, suggestion, human amendment and write to the target system is logged.

Who this suits

Frequently asked questions