Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintech Is Using AI to Detect Mule Accounts in Real Time
Transcript
- Lucas: Luna, I want to talk about something that keeps compliance officers up at night — but it's also a fascinating AI use case. Mule accounts. Luna: Mule accounts — so, legitimate-looking bank accounts that criminals use to layer and move dirty money? Lucas: Exactly. And historically, banks have been terrible at catching them until it's way too late. But a handful of fintechs are now using graph-based AI and behavioral analytics to flag suspicious account networks in hours — not months. Luna: So we're moving from reactive to near real time detection. How does the graph part work? Lucas: Think of it like a social network map for money movement. Traditional anti-money-laundering systems look at each account in isolation — they check transaction amounts, frequency, maybe country of origin. But mule accounts almost never operate alone. They form clusters. Luna: Right — one account receives a deposit, then fans out payments to ten others, which then consolidate and move funds upstream. Lucas: Exactly. Graph models capture that structure. They look at who's paying whom, how quickly money flows through the network, and whether the pattern resembles known mule rings. There was a really stark example in the UK last year — 2025 — where a ring used about four hundred mule accounts to move roughly twenty-eight million pounds before detection. Luna: And that was caught after the fact. Today's systems would flag the network when the second or third layer of payments hits? Lucas: In theory, yes. Revolut and Monzo have both been pretty public about building these capabilities. Revolut, for instance, now uses a combination of graph analytics and behavioral biometrics — things like how fast someone types their password, or the angle they hold their phone. If an account was opened with one behavior profile and suddenly all transactions are happening from a different device posture, the system can freeze the account pending review. Luna: That's a lot of data to process in real time. Are these models running on-device or in the cloud? Lucas: Mostly cloud-based with edge inference on the mobile app. The heavy graph computation happens server-side because you need to compare against the entire network. But the behavioral biometrics can be scored locally on the device and just send a risk score — that keeps latency low. Luna: I can see why regulators are interested. The EU's Anti-Money Laundering Authority, AMLA, has been pushing for mandatory real-time screening of all cross-border transactions by 2027. The UK's FCA is experimenting with a similar sandbox. Lucas: Yeah, and that creates a really interesting tension. Because onboarding has to be fast and frictionless — especially for neobanks competing on user experience. But if you gate every new account with intense scrutiny, you lose customers. So fintechs are trying to thread the needle: let people in quickly, but monitor aggressively from day one. Luna: And mule accounts are often opened with stolen or synthetic identities anyway, so the onboarding check might not catch them. The real detection happens after they start moving money. Lucas: That's the key insight. Mule accounts often pass KYC because the identity document is real — it's just not being used by the actual person. So you need the behavioral layer post-onboarding. One approach I find clever is what some fintechs call 'account velocity monitoring.' If an account that's been dormant for six months suddenly receives five thousand dollars and within an hour sends it to three separate accounts, that's a red flag. Luna: And the graph model would see that those three accounts also received money from other dormant accounts — so it's a pattern, not a one-off. Lucas: Right. And the models get better over time. They learn new typologies. For example, during the pandemic, a lot of mule networks started using smaller, repeated transactions to fly under the reporting threshold — structuring, basically. The graph models adapted by looking at aggregate flow between account clusters rather than individual transaction sizes. Luna: So it's an arms race. Criminals change tactics, the models update, criminals find new gaps. Lucas: Exactly. And one gap that's emerging is the use of cryptocurrency on-ramps. Mule accounts receive fiat, buy crypto, then move it through mixers. Some fintechs are now linking their graph models to blockchain analytics to trace those flows. Luna: That's a whole other layer of complexity. But I want to come back to the cost. Building these systems isn't cheap — engineering talent, cloud compute, data storage. How do fintechs justify the investment? Lucas: Two words: regulatory fines. In 2024, fines for anti-money-laundering failures across Europe totaled over four billion euros. That's a huge line item. Plus, if your brand gets associated with money laundering, depositors leave. So the ROI is actually pretty clear when you frame it as risk mitigation. Luna: And there's a competitive angle too. If you can offer real-time monitoring and faster suspicious activity reports to regulators, you build trust. That can translate into cheaper correspondent banking relationships, faster license approvals. Lucas: Speaking of trust — this is actually something I think about a lot. These conversations about how fintech is rebuilding financial infrastructure, from payments to lending to fraud detection — they're useful if you're building or running a business in this space. And if you find them useful, listener support is what keeps the show ad-free and independent. There's a page where that works, if you're inclined: buy me a coffee dot com slash fexingo. Luna: Yeah, and it genuinely makes a difference. We don't have sponsorship commitments, so we can drill into the stuff that actually matters — like how graph models detect mule networks — without pulling punches. Lucas: Exactly. So, back to the detection side — one thing that's still hard is false positives. If your model is too sensitive, you freeze legitimate accounts and anger customers. If it's too loose, you miss the mules. Luna: What's the current false positive rate look like for these graph-based systems? Lucas: The better fintechs claim around two to three percent — so out of a hundred flagged accounts, only two or three are actually mules. That sounds low, but when you're monitoring millions of accounts, that still means thousands of false alarms per day. So they layer on secondary checks: manual review, or in some cases, a second AI model that looks at transaction narratives and counterparty risk. Luna: So it's an ensemble of models — graph, behavioral, NLP on transaction memos. Lucas: Exactly. And the NLP piece is interesting because mules sometimes leave hints in payment references — things like 'thanks for the loan' or 'rent payment' when the amounts and timing don't match. The model can flag mismatches. Luna: It's almost like a forensic accounting AI. What about smaller fintechs that can't afford a whole data science team? Lucas: There are now several regtech as a service platforms — companies like ComplyAdvantage and Napier — that offer graph-based AML monitoring as an API. You pipe in your transaction data, they return risk scores. So the barrier to entry is dropping. Luna: That's good because mule networks don't discriminate between big banks and small fintechs. They'll target whichever has weaker defenses. Lucas: Right. And the regulators are starting to expect this level of sophistication. The FCA's latest guidance on financial crime explicitly mentions 'network analysis' as a best practice. So it's moving from nice to have to table stakes. Luna: Where do you see this going in the next two years? More regulation, more data sharing between institutions? Lucas: I think we'll see the emergence of industry-wide graph databases — a shared map of accounts and transactions across multiple fintechs, governed by strict privacy protocols. The EU's AMLA is already piloting something similar for cross-border flows. If that scales, mule networks will have a much harder time hiding. Luna: But that also raises privacy concerns. A shared graph of everyone's transactions — even anonymized — is a pretty tempting target. Lucas: Absolutely. The technical challenge is to do the graph computation without centralizing the data. Techniques like federated learning and differential privacy are being explored. It's early, but it's the next frontier. Luna: So the same AI that catches the mules also has to protect the innocent. That's a good note to end on. Lucas: It really is. Thanks, Luna.