AI System for Referring Financial Fraud Signals for Review

The problem

When transaction, device, time and account behaviour data are held in different sources, staff must manually align the fields before preparing an evidence-backed fraud signal.

The fraud analyst must compare every suggestion in the evidence-backed fraud signal with the source record; incomplete or conflicting evidence must not replace the final decision.

If the effect of travel or a shared device on behaviour is not referred for separate review, a valid transaction could be blocked; the record must not progress without a human decision.

What we set out to improve

• Time taken for the draft evidence-backed fraud signal and its supporting evidence to reach the fraud analyst for review
• How records referred for human review because travel or a shared device has changed behaviour are resolved
• Number of suggestions corrected after identifying the risk of blocking a valid transaction

How the system works

1. 01 Retrieve transaction, device, time and account behaviour data from authorised sources
2. 02 Verify field and document versions
3. 03 Prepare the evidence-backed fraud signal with supporting evidence
4. 04 Check for exceptions: travel or a shared device has changed behaviour
5. 05 Have the fraud analyst accept, correct or reject the suggestion
6. 06 Decision gate: transaction blocking, notification and risk classification
7. 07 Record the decision and source separately

Methods we used

• A source-identified data schema for transaction, device, time and account behaviour data
• A structure that separates deterministic rules from AI suggestions
• A screen showing the evidence-backed fraud signal alongside its supporting evidence
• A human review queue for behaviour affected by travel or a shared device
• A record identifier that prevents duplicate processing

Where people stay involved

Data and security

Transaction, device, time and account behaviour data should be processed using only the fields required. Role-based access should be applied, and unnecessary identity, health, financial and internal organisational data should not be sent to the model. Source retrieval, the AI suggestion and the fraud analyst's decision should be recorded separately.

Who this suits

Frequently asked questions