Thailand’s Most Dangerous Bank Fraudster Is You

The Anantara Siam Bangkok hotel smells of jasmine and money. On the ground floor, at a corner table in a restaurant, Nutapone Apiluktoyanunt leans forward and says something that should concern every person with a bank account in Southeast Asia.

Apiluktoyanunt cites a figure from surveys his team has been tracking: 73% of Thai consumers are exposed to scam risk. “73,” he says. “That’s a very high number.”

Apiluktoyanunt is the managing director of SAS Thailand, and he has spent the better part of a decade watching fraud evolve from a blunt-force nuisance into a precision-engineered industry.

The figure he cites is not a prediction. It is the present state of play. Nearly half of Thai consumers, 47%, have already experienced some form of financial fraud. The gangs aren’t picking pockets; they’re picking psychology.

The authorized threats

To understand why threats like Authorized Push Payment (APP) fraud have broken traditional detection systems, you need to go back to the original logic of those systems. Legacy fraud engines operated on a simple premise: an unauthorized transaction is anomalous, and anomalies can be flagged. When a bot drains your account at 3 a.m. from an IP address in Moldova, a rules engine catches it.

APP fraud turns the entire model on its head. The transaction is no longer anomalous. Now, it is the customer making a payment they believe in completely. Ian Holmes, SAS’s director and global lead for enterprise fraud solutions, describes it with clinical precision: “Authorized means it’s actually me, the customer, making this — and I’m happy to do it at the time. It’s not until days, weeks, or months later that I realize I was incorrect.”

This is the architecture that romance scams, investment scams, and employment scams run on. The fraudster does not crack your account; they crack you. Through WhatsApp, through Facebook, through months of manufactured intimacy or manufactured urgency, they engineer a moment where you, the customer, become the vector of attack. When the bank sends a push notification asking “Is this transaction yours?” you tap yes, because you’re certain.

“I can no longer send you a push notification to say, is this transaction yours? Because you’re going to click ‘yes’,” Holmes says. “You’re so convinced, from a social engineering perspective.”

The bank is then trapped in an impossible position: decline a transaction the customer insists is legitimate, and risk a customer service flame war. Approve it, and become complicit in the fraud the customer will report in three weeks. Traditional rules-based systems have no play here.

The body has no poker face

The countermove is not to build smarter rules but deploy smarter sensors. SAS calls it Behavioral Telemetry, and the concept is rooted in a simple human truth: bots are precise and people are not.

Consider what happens when you type on your phone on the BTS Skytrain. You are probably using your thumb, tilting the device slightly and are making spelling error and correcting it. You hold the phone at the specific, slightly impractical angle that is uniquely yours. None of this is conscious and all of it is data.

Holmes describes the SAS Viya engine’s approach: “The angle that I hold the mobile phone, the speed at which I type and any spelling errors I make, makes it human-like. Seeing those little nuances make sure it looks like a human rather than a bot.”

The practical upshot is significant. A bot conducting a scripted attack leaves a signature of machine-like regularity — perfect keystrokes, consistent timing, and no hesitation. A social engineering victim, by contrast, is emotionally charged, possibly typing quickly, and may be slowly. The entropy of being human becomes, paradoxically, a form of authentication.

For banks that have layered this digital behavioral intelligence onto existing fraud models, the results have been measurable. In slides presented at the forum, SAS cited analytics deployments that reduced fraud growth by more than 50% and cut false positives by 40% — figures that matter far beyond operational efficiency. For fraud specialists, a false positive is not just a declined transaction. It is a legitimate customer who had their payment blocked and who may never fully trust the bank again. Reducing them is both a commercial and an ethical imperative.

SAS also works with third-party providers to go deeper into the hardware itself. Using relationships with security partners, the system can inspect whether a device has been jailbroken, running checks at the firmware level before a single transaction is even proposed. “There are permissions around that,” Holmes acknowledges, “but we have done studies where the customers are very willing to give some of that data” — particularly when the alternative is losing their savings to a scam farm in Cambodia.

Your money’s next stop

Even when fraud is detected at the point of transaction, the money still has to go somewhere. That somewhere is a mule account — known locally in Thailand as a “horse account” or Banchee Ma — and the ecosystem of mule infrastructure is where the real systemic damage is done.

The Bank of Thailand has established a working classification for mule accounts in colors: from black, dark grey and light grey to dark brown and light brown. Brown accounts are where a bank is internally suspicious and gray is where that suspicion has been proposed to the regulator as a potential mule account. Black is where that designation has been validated. The BOT has also moved to restrict cash withdrawals above five million baht, requiring declaration of source and destination. Apiluktoyanunt notes, “These are some of the things our government is trying to do to reduce some of this scam landscape.”

But restricting cash movement is playing defense. The offensive play is Entity Resolution, reconstructing a true, unified identity for every actor in the financial network, regardless of which channel, branch, LINE account, or SMS thread they used to interact with the bank. Once that unified identity layer exists, Graph Analytics can map the hidden architecture: which accounts cluster together, which beneficiaries are shared across hundreds of seemingly unrelated transactions, which nodes in the network are the real chokepoints.

That AI-driven network view is where the systemic picture of cross-border scam syndicates becomes visible. Dismantling the region’s scam-origination hub rep requires the same systemic intelligence that built it.

The rules aren’t ready for you rither

In September 2025, the Bank of Thailand published its AI risk guidelines for financial institutions. It is a framework modeled on global standards, covering governance, model life cycles, data quality, and explainability. Most major Thai banks have AI-powered fraud models. The gap, Apiluktoyanunt observes, is not in the models.

“We are moving towards a lot of AI-driven supervision,” Apiluktoyanunt says, pointing to the BOT’s Banking 5.0 initiative. “But there is uneven execution.” The fracture lines are consistent: fragmented ownership between fraud teams and IT departments, models that are deployed but never stress-tested against adversarial scenarios, and post-incident feedback loops that are too slow to retrain models before the fraudsters have already adapted.

Holmes frames the structural problem in terms that any fraud specialist will recognize: “Banks have heavily invested in building their own models for risk for many years. From a fraud perspective, they were either using rules or vendor models like SAS. Now they’re wanting to utilize some of that data science expertise in building their own fraud models.” But without the governance infrastructure that risk management teams built over decades. The result is sophisticated detection capability housed inside brittle institutional plumbing.

Teaching the machine to investigate you

Consider what a fraud investigation actually looks like at ground level. When a specialist receives an alert, he or she opens five different internal databases. They query a sanctions list, check a third-party address verification service and read three years of transaction history and attempt to manually construct a narrative of what happened. They then write a report. This process, in a busy compliance team, can take hours. The money has already moved.

Agentic AI, as Holmes describes it, is the attempt to compress that process to minutes. “An AI agent can gather data, assess that data, extract entities from that data — and compile a report which is ready for QA” by a human reviewer. The human remains in the loop, but their role shifts from data gatherer to validator, which is a far more defensible use of specialist judgment.

Holmes is precise about why agentic AI has achieved traction where earlier AI implementations stalled: “It’s applicable and understandable by the person who needs to utilize it on a daily basis. AI is kind of in the background — it’s like a magic black box, and it becomes difficult for people to understand and therefore rely upon. Whereas we all know that we use ChatGPT to design a holiday. And therefore we can understand it.”

Simply put, fraud specialists are not going to trust a system they cannot interrogate. Agentic AI gives them something they can actually see working.

The next burglar has a quantum key

The longer-term threat landscape sits at the intersection of two emerging technologies: quantum computing and generative AI. Apiluktoyanunt describes plainly: “We need to have AI able to cope with all this huge transaction in a real-time manner. Quantum AI is coming. We start to look at quantum right now.” The corollary threat that adversaries could use the same quantum and generative AI capabilities to inject malicious prompts and manipulate detection models from within is one Holmes takes seriously.

For Holmes, the defence runs through containment. Generative AI components are deployed inside the bank’s own infrastructure perimeter, tuned on the bank’s own data without having to call out to external APIs. The attack surface shrinks and output stays explainable to a regulator. And the system keeps learning, feeding every fraud event back into the next model iteration and tightening the loop between detection and adaptation.

“All this is built into SAS Viya,” Apiluktoyanunt says. From his vantage point in Bangkok, the destination is a platform capable of processing millions of PromptPay transactions in real time, resolving identity across every channel simultaneously and staying ready for whatever the next generation of adversarial technology brings.

Image credit: iStockphoto/think4photop