Since IBM Synthetic Datasets first debuted earlier in 2025, one of the use cases we’ve been hearing increasingly from financial institutions is being able to detect fraud in instant payments. Our team is happy to announce that artificial instant payments, specifically like ones from peer-to-peer (P2P) institutions, are added in this October’s update of IBM Synthetic Data Sets for Core Banking and Money Laundering.  Read the announcement here

What is Instant Payments Fraud?

Instant payment fraud is when scammers trick victims or exploit weaknesses in real-time payment systems to steal money, making it difficult to recover due to the speed and irreversibility of transactions.

Instant payments have grown as a payment type with the financial ecosystem embracing speed, convenience, and increased flexibility and transparency into payment status. Peer-to-Peer (P2P) instant payment platforms like Venmo, Zelle, PayPal, and others are growing rapidly in popularity because they are designed to enable easy, rapid, and trustworthy payment to others. Typical examples include reimbursing a friend for dinner or concert tickets, or reimbursing a roommate for rent. Businesses also benefit from serving their customers better and having improved cash flows.

However, as adoption of instant payments grows, this adoption also increases fraud risk. P2P also enables easy and rapid payment to untrustworthy others such as scam artists. As such, crime involving P2P is quickly increasing. Bloomberg’s article¹ around instant payments quotes a 2024 AARP report² that found that nearly one in five adults surveyed had been targeted for fraud on such platforms.  Additionally, in 2024, the US Federal Trade Commission (FTC) received 2.6 million reports of fraud with $2.95 billion reported losses³ coming from imposter scams. P2P is attractive to criminals because the transactions happen immediately and are not reversible.

It has been difficult for banks and others to determine if a particular transaction (P2P or otherwise) is a scam.  Tens of millions of transactions are made per day in the US alone, and those transactions involve a huge variety of payers and receivers.  To make matters worse, IPX1031 reported in their American Fraud and Identity Theft Statistics 2024 that the majority of Americans (57%)⁴ aren’t reporting scams to authorities for various reasons.

With the huge transaction volume and poor knowledge of which of those transactions are scams, real data provides banks and others with a mediocre foundation on which to build AI models to detect them.  Poor models are not only a problem in missing criminal activity, poor models can be a huge annoyance to businesses and consumers when transactions are legitimate but erroneously flagged as a scam and blocked.

Detect instant payments fraud with real time inferencing with transactional AI on IBM Z and LinuxONE

In order for financial institutions to combat instant payment fraud, there are many approaches to take, but a competitive one is AI-powered fraud detection solutions using AI on IBM Z and LinuxONE. Given that 70% of the world’s transactions by value⁵ run through a resilient, secure and high availability transaction processing IBM Z and LinuxONE server, co-locating AI models on the same processing chip as the payment transactions results in low latency and high-performance AI inferencing.

By bringing fraud detection models created on other platforms for inferencing on IBM Z and LinuxONE, predictions are made in real-time. Instead of extracting a portion of the transactions off-platform for scoring with additional cost and latency, running the AI models on IBM Z and LinuxONE means 100% of transactions are scored and done so in real time as transactions are in flight. One bank in the US moved their AI model from off platform, and deployed it on IBM Z, and went from only scoring 20% of their transactions to scoring 100% of their transactions on IBM Z – this resulted in the bank being able to detect more fraud accurately and saved up to $20M in costs, while lowering their risk and increasing customer satisfaction.

Labelling of complex criminal activity in IBM Synthetic Data Sets is designed to aid in building instant payments fraud detection models

Financial institutions can use IBM Synthetic Data Sets (SynDS) to tackle instant payment fraud on IBM Z and LinuxONE. SynDS provide huge numbers of realistic synthetic transactions with labelling of 100% of criminal activity – including scams.  With the release of Version 1.1.0 on October 31, 2025, this SynDS labelling will include artificially generated data designed to be similar to data used by P2P institutions – both the legitimate reimbursements such as dinner and rent, and the fraudulent activity.  SynDS not only marks fraudulent activity as fraud, but also labels each fraudulent transaction with the identity of the criminal perpetrating the fraud, designed to enable better identification of particular patterns and criminals.

SynDS further labels all banking transactions -- P2P and otherwise – with their underlying purpose, be it dinner reimbursement, payment of salary, loan payment, accounts payable, and dozens more.  These labels are designed to help financial institutions better understand normal patterns and hence better detect the anomalous behavior of a scam.  These transaction-purpose labels have the potential to also help grow business, e.g. identify certain products and services that may be of interest to customers.

SynDS goes even further than identifying scam transactions.  Like other illicit funds, money stolen through scams and P2P is often laundered through a web of accounts and transactions.  SynDS also labels this follow-on laundering allowing financial institutions to build AI models that capture a big picture of illicit funds flows – from initiation through complex laundering.

Like scams, detection of money laundering is a huge challenge, with only an estimated 1-2% being detected each year according to Ashish Pradhan from Financial Crime Insights⁶.  Thus, as with scams, real data provides a poor foundation on which to build accurate AI detection models.

With its realistic data and its full and complete picture of these crimes and others, this enhanced version of IBM Synthetic Data Sets enables training a variety of standalone or integrated models to detect scams using P2P and other criminal behavior. With increased fraud detection model accuracy, more fraud accurately detected results in lowering business risk and saved costs lost to fraud. The amount of money saved could be reinvested into innovations and revenue for the company.

Next Steps

• Want to learn more about the instant payment use case on AI on Z and LinuxONE? Or get started with trying out our limited time Trial datasets by contacting aionz@us.ibm.com.
• View the data schemas for what type of information is included in the datasets in Github: https://github.com/IBM/IBM-Synthetic-Data-Sets

Sources:

1. Basu, Kaustuv. Bloomberg Law, “IBM Synthetic Data Tool Targets Fraud on Instant Pay Platforms.” October 16, 2025.
2. Gunther, Jilenne. AARP, “Fighting Financial Exploitation on Person-to-Person Payment Platforms: What Consumers Want” February 8, 2024.  
3. Federal Trade Comission “New FTC Data Show Big Jump in Reported Losses to Fraud to $12.5 Billion in 2024” March 10, 2025.
4. IPX 1031 American Fraud and Identity Theft Statistics 2024
5. Katov PhD, N. (April 2025), Mitigating Fraud in The AI Age: Supporting Transaction Fraud Detection at Scale on IBM z17, Celent.
6. Financial Crime Insights, Ashish Pradhan “Money Laundering: How Much Do We Really Identify?” March 29, 2025.