AI, neobanks, and new futures in fraud prevention

Fraud and scammers are as old as finance and bankers. In other words, where there’s money, there’s bad actors, and fraud detection is as essential a process to finance now as it was a century ago. Fraud detection has continually evolved much like any other aspect of the finance world, and it’s also evolved in tandem with technology itself. The development of artificial intelligence brought with it obvious applications to banking, and now we’re living in a finance world replete with AI, including and especially within the realm of fraud detection.

Granted, AI’s applications go both ways. Fraudsters aren’t stuck in the past, meaning that new fraud attempts utilising AI technology are everywhere. For banks, this means that their fraud-detecting AI technology needs to be better than the fraudsters’, and it needs to be incredibly comprehensive.

Among the most ambitious adopters of AI in finance are neobanks, which are agile enough to adapt quickly to the changing landscape and adopt the very latest technology. Just five years ago, large language models basically didn’t exist in any way that was beneficial to banks. Now these highly functional and powerful AI tools are crucial to fraud prevention efforts, especially at neobanks.

AI detection’s semi-recent leaps and bounds

In the not-too-distant past, neobanks were mostly focused on developing code to automate specific processes - sorting, categorising, and recognising data - and automating translation services to better communicate with clients across the world. The entry of a more sophisticated AI (especially, it bears mentioning, OpenAI’s GPT-3.5 Turbo), changed neobanks’ entire approach to automation. No longer did specific processes have to be painstakingly automated. The speed and accuracy of a newly sophisticated AI allowed neobanks to analyse data in real-time without disrupting operations. This changed everything for fraud detection.

Because AI can rapidly analyse huge volumes of information, it can be depended on to categorise and identify subtle discrepancies in transactions and invoices. Traditional fraud detection relied on predefined rules and a comparatively simple decision tree, flagging transactions that exceed a certain amount, for instance. AI liberates fraud detection from this decision tree and the associated rigid systematic processes that fraudsters can sidestep. It understands patterns and anomalies, such as an invoice that lists a single line item at three times its usual price. It’s capable of making educated guesses based on incredible amounts of data and verifying legitimacy by suggesting follow-up questions to clients. The likelihood of a human monitor seeing this is slim, but AI catches it every time.

Fraud detection is always going to be somewhat reactive - in order to detect fraud, the fraud has to happen first. That said, AI-based fraud detection is balancing the scales between reactive and proactive, which is also to say that it’s balancing the scales between detection and prevention. 

Pitfalls and promise

AI fraud detection is not some perfect system that has rendered fraud a thing of the past. Accuracy is not guaranteed. Left unchecked, AI prevention systems become overly aggressive, blocking legitimate transactions and flagging false positives. Fine-tuning is a continual process. 

We sample and validate its outputs to ensure accuracy, adjusting prompts and refining the AI’s direction as needed. If left unchecked, AI can be overly aggressive—blocking transactions unnecessarily or flagging too many false positives. 

Any given AI system is also limited by its data. Hypothetically, this wouldn’t have to be the case, so long as financial institutions could develop a single, massive context for all operations across a reasonably large region, with a single AI monitoring all transactions across these institutions. Since fraudsters move money among accounts at multiple institutions to cover their tracks, an AI that wasn’t limited to a single institution could detect this, thanks to full visibility into how money flows. This is a possible future that neobanks, which tend to be more agile and open to cross-institution collaboration, are poised to pioneer. Not only that, this would upend fraud prevention in the best possible way.

Monitored and adjusted correctly, AI systems are infinitely faster and more reliable than a human monitor. IT is not limited to hardcoded strategies and scenarios, it's this agility, however, that can be hard to convey to auditors. There are no pre-set scenarios or predefined rules. This is an advantage, but only if it can be communicated effectively to regulators. As of now, this is as important a task as pursuing multi-institution detection. The future of fraud prevention may very well depend on it.