last refreshed - October 2026
The AI You Don't See: How Machine Learning Revolutionized Fraud Detection
Artificial intelligence has become ubiquitous, and with that ubiquity has come a change in how people think about it. For most, it's become an assistant that answers questions, writes documents, and synthesizes spreadsheets.
But chatbots are only the newest, most visible form of the technology. Machine learning has been monitoring the systems we use every day for much longer, and that silent, ever-present work goes underappreciated because it's unseen.
Fraud monitoring is one of the clearest examples. Banks began using neural networks to flag suspicious card activity in the early 1990s, long before anyone asked a chatbot to draft an email. In recent years, AI has taken a more involved role, scoring every card swipe, bank transfer, and portal login as it happens.
Because the technology has been around for a while, the changes don't seem all that revolutionary. But their impact will be resonant. Fraudsters are finding new, increasingly elaborate ways to steal hard-earned money, and better models mean billions of dollars each year are better safeguarded from them.
The stakes are growing. Americans reported losing $15.9 billion to fraud in 2025, according to the Federal Trade Commission. That's the highest figure on record, up from $12.5 billion a year earlier. And the FBI's Internet Crime Complaint Center counted another $20.9 billion in cyber-enabled crime losses, a 26% jump. Both agencies say the actual totals are likely much higher due to underreporting.
The standard in finance
Slash goes above with better controls, better rewards, and better support for your business.
Why Hard-Coded Rules Couldn't Keep Up
The old way of monitoring card and bank activity was by creating models around predetermined rules. Typical red flags included:
- Transactions that were larger than expected
- Transactions that occurred in an unfamiliar location
- Transactions that were sent to an unfamiliar party
- Transactions that were made in quick succession at unusual merchants
All of these conditions could be hard-coded and were generally more effective than no safeguard at all. Still, they came with two big problems.
The first was that, over time, fraudsters learned the rules. They became better at avoiding the hard-coded flags, either by making smaller, less frequent transactions or by spreading activity across many cards and merchants so no single account tripped a threshold.
Card testing shows how that works in practice. Bots submit thousands of small online charges, cycling through card numbers, security codes, and expiration dates until one goes through. Each charge is built to look too small and too ordinary to flag, and Visa has estimated these attacks cost about $1.1 billion a year.
The second problem was that many of these checks relied on physical purchasing to run properly. As card-not-present transactions have increased with online purchasing, it's harder to apply the same rules when there's no card, no terminal, and no cardholder at the register.
Location, for instance, means little when a purchase can be routed through any country. Banks could tighten the rules to compensate, but a rule strict enough to stop that kind of fraud starts rejecting real customers. In a 2026 PYMNTS Intelligence study, nearly half of merchants said wrongly declined orders had cost them sales.
The Weak Point in the Approval Chain
For businesses, internal protections have prioritized smarter ways to authorize transactions, and digital banking gave them the tools to do it.
Banking apps made it possible to upload receipts or deposit checks using online verification. Digital statements and transaction logs made it easier to catch when a transaction looked suspicious.
Approval rules on bank transfers followed. They became normalized as it got easier to send a manager within the company a digital form to sign when a larger transfer needed to be made.
But even still, those have their limitations. More fraudsters have begun using psychological attacks, not just technology ones.
The most common version is business email compromise, in which a fraudster poses as an executive or vendor and asks for payment to a new account. It hit 74% of organizations in 2025, according to the Association for Financial Professionals' annual payments fraud survey. It's also expensive. The FBI put reported BEC losses at $3.04 billion for the year. Only investment fraud cost more.
From Fixed Rules to Learned Behavior
Machine learning handles fraud detection in two ways: it first learns how you or your business spends money, and it knows how to predict suspicious activity based on large models and data sets tuned for financial protection.
The first element works by setting a behavioral baseline. The model determines whether a given transaction fits the account's own history: usual merchants, amounts, times, devices, and payees. A $9,000 wire can be routine for one business and alarming for another, so the model scores each business on its own terms.
The second element relies on network knowledge. Card networks and large processors watch millions of accounts at once, so an attack spotted at one bank can shape scoring everywhere. Visa has said its models weigh more than 500 attributes of each transaction.
Just as a language model reads a sentence word by word, the models powering card networks and processors read transactions in sequence. That lets them pick out patterns that look harmless one at a time, like the card-testing attacks from the rulebook era, buried in a merchant's normal traffic.
There's another payoff that gets less attention. A model that understands normal behavior can approve more good transactions, not just block more bad ones. For merchants and cardholders, fewer false declines matter nearly as much as fewer fraudulent charges.
The Protection Most Businesses Never See
For all the progress at the network level, most businesses aren't implementing these tools themselves. The AFP survey found just 17% of organizations use AI to fight payments fraud.
The ones that do are seeing results. Among them, 49% reported more efficient fraud reporting, 45% said they were better at detecting deepfakes, and 43% said they could identify fraud in real time.
The implementation matters most for smaller firms. They're targeted less often, the survey found, but when an attack succeeds, they're more likely to absorb the entire loss because they lack the recovery resources of larger companies.
For a small or midsize business, then, most of the gains from machine learning come down to choosing the right bank, card issuer, and network with security in mind.
What the FBI's first AI cybercrime section means for payment approvals
Banks have used machine learning on card traffic for decades. Criminals are now using generative tools on the social layer that sits in front of those models: emails that sound like a CEO, voice clones that request a wire, and synthetic profiles that make a vendor change look routine.
The FBI Internet Crime Complaint Center made that shift explicit in its 2025 Annual Report, adding a dedicated section on artificial intelligence used in cybercrime. IC3 received more than 22,000 AI-related complaints that year, with adjusted losses exceeding $893 million. Inside that set, businesses reported more than $30 million in losses on business email compromise cases that involved AI, including chat-generated messages and voice cloning used to request payment.
Those figures do not replace network-level card scoring. They explain why approval workflows still matter. A model can correctly score a wire as unusual, but the payment still goes out if a human believes a deepfake call or a polished email. Dual approval on new payees, callback verification on account changes, and banks that can hold suspicious outbound RTP or FedNow transfers before settlement close the gap the complaint data is describing.
For operators, the takeaway is layered defense: choose issuers and networks with strong behavioral models, and keep human verification on the payment instructions AI is best at faking.
Keep card and payment monitoring on by default
Use Slash controls and background monitoring across cards, payouts, and logins.
How Slash Monitors for Fraud in the Background
At Slash, automated monitoring is built into the platform and runs in the background across cards, payments, and account access.¹
Card testing, the attack that slipped past the old rulebook, is one example. Slash worked with Visa to develop monitoring designed to detect patterns consistent with PAN enumeration on the Slash Visa Platinum Card. The system looks for repeated declines that share a BIN, response codes that suggest an attacker is closing in on valid card details, and merchants seeing a sudden wave of failed payments across many different cards.
When a merchant looks like a testing ground, Slash can check whether any Slash cards were recently used there and treat them as exposed before a fraudulent charge goes through. From there, it can block the merchant or deactivate the affected cards.
The same approach was applied to transactions made off of cards, too. Outbound transactions on Slash are monitored, and suspicious activity can be flagged or held for review before it settles. That matters on instant rails like RTP and FedNow, where a payment is generally final once it's sent. Real-time alerts keep teams informed as payments are initiated, approved, and completed.
Monitoring also covers activity that stays within a card's limits. Slash can surface patterns like repeated max-limit purchases or bursts of small charges, which can signal misuse even when no rule is broken. At login, device fingerprinting and new-device detection add a verification step when an account is accessed from an unfamiliar device or IP address.
When something is flagged, Twin, Slash's AI financial assistant, lets teams investigate it using live account data and act from the same conversation, whether that means checking recent activity or freezing a card from Slack. The controls a team sets still matter, but they no longer have to catch everything on their own.
Apply in less than 10 minutes today
Join the 10,000+ businesses already using Slash.
