Compliance monitoring · Production

Monzo builds a real-time ML fraud detection model to identify and block APP fraud payments

The problem

Authorised push payment fraud is the fastest growing type of fraud in the UK, with over 35,000 people losing £145 million in the first half of 2018 alone. Bank transfers are instant and nearly impossible to cancel, making recovery very difficult once a victim has been tricked into sending money.

Workflow diagram · grounded in source
1
Payment event triggers review
trigger
“Whenever a Monzo customer receives or spends money, our systems have to decide whether to approve or decline the payment”
2
Model scores for financial crime
ai_action
“We feed these transactions into a model that tells us how likely it is that the account activity indicates financial crime. This model considers dozens of transactional signals and several properties about the customer. We don't have ver…”
3
Name-match comparison
validation
“on every inbound bank transfer we compare the name the sender entered when they sent the payment with the actual name on the account and store the result of this comparison. If the person sending the bank transfer used a different name t…”
4
Approve or decline payment
output
“our systems have to decide whether to approve or decline the payment. There are hundreds of different reasons a payment might decline, but for authorised push payment fraud we're looking for activity that might indicate that the bank tra…”
Reported outcome

Monzo's fraud detection controls help more victims recover their money than they negatively impact legitimate customers, though the outcome is described qualitatively.

Reported metrics
APP fraud victims in UK (H1 2018, industry-wide)35,000
APP fraud losses in UK (H1 2018, industry-wide)£145 million
Victims helped vs. legitimate customers affectedmore than one victim helped for every legitimate customer affected
Source
https://monzo.com/blog/2019/02/07/authorised-push-payment-fraud
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Monzo's fraud detection controls help more victims recover their money than they negatively impact legitimate customers, though the outcome is described qualitatively.

What results were reported?

APP fraud victims in UK (H1 2018, industry-wide): 35,000; APP fraud losses in UK (H1 2018, industry-wide): £145 million; Victims helped vs. legitimate customers affected: more than one victim helped for every legitimate customer affected (source-reported, not independently verified).

How is this compliance monitoring AI workflow structured?

Payment event triggers review → Model scores for financial crime → Name-match comparison → Approve or decline payment.

Related compliance monitoring cases