In early 2024, The Guardian reported that U.K. engineering firm Arup lost $25 million to a deepfake scam. Fraudsters impersonated company executives using AI-generated video and voice, convincing an employee to wire money in what became one of the most expensive synthetic fraud incidents to date. It was a warning to companies everywhere that fraud, as they once knew it, had evolved.
Unfortunately, the defense systems built to prevent attacks like this also come with unintended costs. In sectors like CBD, telehealth, gaming, crypto, alternative finance and nicotine — all of which are often considered high-risk — businesses operating legally are often flagged as too risky by automated fraud systems. Even mainstream processors like Stripe or PayPal abruptly freeze some businesses’ accounts, subjecting them to higher transaction fees without any appeal or explanation.
According to Kirk Fredrickson, founder of 2Accept, a California-based company that helps high-risk merchants stay compliant, the problem is that most fraud engines treat anything unfamiliar as dangerous. “We’ve seen companies lose accounts overnight for nothing more than a keyword in their product description. That kind of overreach doesn’t just hurt business; it undermines trust in the system,” he told me.
For companies already operating under regulatory scrutiny, being labelled “fraud” by mistake can be just as damaging as the real thing. And that’s exactly the challenge that companies like 2Accept are trying to solve.
When Al Gets It Wrong
While legacy fraud systems were designed to block fraud, a flagged keyword in a product description or a single transaction from the “wrong” region can trigger account freezes or permanent shutdowns.
A report from Fraud.com estimates that false positives cost merchants 2.8% of annual revenue. Recovering from a blacklisted business can be challenging. Many platforms don’t offer recourse or even reasons for such shutdowns. Once blacklisted, recovery is often nearly impossible, since many platforms provide no explanation or recourse.
That’s created a market for firms that help businesses navigate hostile compliance environments. “Most companies we work with aren’t trying to skirt the rules,” Fredrickson said. “They’re trying to play by them, and our job is to make sure AI can tell the difference.”
His company, 2Accept, solves that challenge using onboarding models that monitor patterns across transactions, chargebacks and merchant behavior to help businesses stay in good standing. According to Fredrickson, the company’s systems reduce account termination risk by up to 60%, helping thousands of merchants across CBD, telehealth and fintech to stay compliant.
A New Phase In Fraud Prevention
This dilemma isn’t just a small-business issue. Major players are also trying to cut down on wrongful declines. Mastercard now uses Decision Intelligence Pro, an AI system analyzing 160 billion transactions a year in real time, combining behavioral and device data to distinguish fraud from legitimate activity.
Another major U.S. company, Riskified, recently helped a U.S. ticketing platform recover $3 million in sales by deploying adaptive AI at checkout to reduce unnecessary blocks, according to Business Insider. HSBC also reported its AI models reduced false positives by 60% while detecting two to four times more real fraud. 2Accept reports that merchants on its platform see up to 48% fewer chargebacks and benefit from partnerships with tier-1 acquiring banks.
Across the industry, the big message now is that it’s no longer enough for AI to catch more fraud. It must also catch fewer honest mistakes. And while there’s so much left to be seen in terms of how much more effective these systems can be, current efforts are undoubtedly a step in the right direction.
Accountability Is No Longer Optional
As AI takes on more responsibility in fraud detection, the margin for error is shrinking, and systems tasked with determining who transacts and who doesn’t cannot conceal their decisions. “The tools we build have to be explainable,” said Fredrickson. “It’s not enough to flag a transaction. You have to be able to say why and what can be done about it.” He’s been vocal on that point since founding 2Accept in 2015, long before AI governance entered the regulatory mainstream. Over the past decade, Fredrickson has also contributed insights to U.S. financial policy discussions on risk-based access, including feedback used by the IRS on banking reform, briefing materials for Senate staff and input submitted to the Government Accountability Office.
Businesses are not the only ones expressing this expectation of accountability, as the law is now increasingly backing it. The EU AI Act and frameworks like the Digital Operational Resilience Act are requiring that automated systems used in high-risk domains like fraud detection offer transparency and accountability by design. In the U.S., agencies like the Consumer Financial Protection Bureau are investigating whether financial institutions’ AI tools are unfairly limiting access to credit or financial services, especially when there’s no clear explanation for a denial.
Across the industry, this is pushing a move toward explainable AI and hybrid systems that blend automation with human review. Experian’s recent report revealed that AI-powered fraud targeted 35% of U.K. businesses in Q1 alone. Over half are now investing in tools not just to catch more fraud but to avoid mistaking customers or companies for criminals.
The Need For Smarter Fraud Tools
Modern fraud is constantly adapting, but so are the systems built to fight it. The real challenge isn’t just about catching criminals but doing so without misjudging the honest businesses caught in the middle. When an AI model misreads a pattern or flags a company based on a misunderstood keyword, the consequences can be immediate and severe: Frozen accounts, lost revenue, reputational damage and no clear path to appeal.
Being treated like a scammer when you’re not one is more than an inconvenience. It can undo years of trust, growth and planning. That’s why the tools we rely on to detect fraud must also be smarter, fairer and learn to understand better.
Fredrickson believes the next phase of fraud prevention is not just about tighter controls but fairer systems. “You can’t build trust with one hand and take it away with the other,” he said. “If AI is going to govern access to financial infrastructure, then it has to work for everyone, especially those trying to do things right.” Tools that reduce wrongful termination rates by half can be crucial in sectors such as CBD or wellness, where up to 70% of merchants face closure within their first year. “Our job isn’t to judge an industry,” Fredrickson said. “It’s to make sure ethical businesses have a fair shot at thriving.”
Fraud prevention is shifting from defining strict boundaries to comprehending the true nature of the data. Instead of cutting off businesses at the first sign of risk, today’s systems are beginning to pause, assess and adapt. The goal is not only to prevent fraud but also to shield legitimate businesses from becoming entangled in the process. While AI still serves as the gatekeeper, it’s now evolving to distinguish between a potential threat and a legitimate business that adheres to the rules.