Mule-Account Detection: What Rules Miss
Mule account detection AI looks at behavior and account networks, not just single transactions. Here is what rules miss, what ML and graphs add, and how to run it.
Mule account detection AI finds accounts that are being used to receive and pass on money from scams and other crimes. Rules look at one transaction or one account at a time. Machine learning and graph analysis look at behavior over time and at how accounts connect, which is where mule networks show up. The best setups use both, with investigators making the final call.
This guide is for fraud, risk and compliance leads at Thai banks, e-wallets and fintechs. It describes the approach, not outcomes: how the detection works, what rules miss, and how to run it responsibly. For the wider context, see our page on AI for banking and financial services.
What a mule account is
A mule account is a bank or e-wallet account that moves money on behalf of someone else, often to hide where it came from. The account holder may be a willing participant, a person who was paid or tricked, or a victim whose identity was misused. That distinction matters for how you respond, since a mule may also be a victim.
Typical activity involves money arriving from several sources, then leaving quickly to other accounts, cash-out points or crypto platforms. The behavior is designed to look ordinary, one step at a time.
Why rules alone fall short
Rules are fast, clear and easy to explain. They are a sound foundation, and you should keep them. Their limits are structural.
- They see one event. A rule like "flag transfers above a set amount" sees a transaction. A mule network splits activity across many small, individually normal transfers.
- They are visible. Once criminals learn a threshold, they stay just below it.
- They age. Each rule reflects a pattern someone already saw. New scam methods pass until a person writes a new rule.
- They flag the wrong people. Broad rules catch many honest customers, which burdens investigators and frustrates customers.
None of this means rules are wrong. They handle known, clear cases and regulatory requirements well. The point is that they work best as one layer.
What machine learning adds to mule account detection AI
A machine learning model looks at many signals together and weighs them. Instead of "was this transfer large," it asks whether the whole pattern resembles known mule behavior.
Useful signals fall into several groups:
| Signal group | Examples |
|---|---|
| Account history | Age of account, sudden change from dormant to active, change of device or phone number |
| Flow pattern | Money in and out within a short period, balance rarely held, many new payees |
| Counterparties | Number of distinct senders, whether receivers are new or already flagged |
| Device and session | Shared devices, new device just before activity, remote-access tools |
| Customer profile | Activity that does not fit the declared occupation or usual behavior |
Models are trained on cases your investigators have already confirmed. This is also the weak point: confirmed cases are a small subset of reality, and labels reflect what you caught in the past. A model can inherit blind spots, so you should test it against new patterns too.
What graph analysis adds
Graph analysis treats accounts as points and transfers, shared devices and shared contact details as links. It then looks at the structure of the network.
This reveals things no single-account view can show: a group of new accounts all paying into one account, many accounts logging in from the same device, or a chain of accounts that passes money onward within minutes. A mule account may look unremarkable alone, while its position in the network is not.
Graph methods also help investigators. Rather than a single alert, they see the connected accounts and can act on the group. Our data and analytics practice covers this kind of connected-data work.
From score to decision
A score alone does not stop anything. Decide what each level of risk triggers, and make the response proportionate.
- Low risk: let the transaction proceed and keep the score for later review.
- Medium risk: add friction, such as a confirmation step, a delay or a call to the customer.
- High risk: hold the transaction or restrict the account and send it to an investigator with the reasons.
Every action should be reversible and logged, because holding an honest customer's money harms them. Set the thresholds against the number of cases your investigation team can review in a day. Thresholds that exceed capacity create backlogs, and delay is itself a risk.
Explainability for investigators and regulators
Investigators need reasons they can act on: which signals contributed, which linked accounts matter, and what changed recently. A bare score invites distrust or blind acceptance, and neither is healthy.
Supervisors also expect governed models. Document what the model uses, how it was tested, how it is monitored and who approves changes. Check current Bank of Thailand guidance on AI and model risk for your license type, and consult the Anti-Money Laundering Office (AMLO) requirements on suspicious transaction reporting. Your compliance and legal counsel should confirm what applies to you.
Feedback loop and monitoring
Fraud patterns change as criminals respond, so a static model decays. Plan for:
- Investigator feedback. Each case outcome, confirmed or cleared, flows back as a label.
- Drift monitoring. Watch score distributions and alert rates for sudden shifts.
- Regular retraining and challenger models. Test a new model alongside the live one before replacing it.
- Red-team thinking. Ask how someone would avoid your detection and test those cases.
Data, privacy and hosting
Detection uses sensitive financial and device data, so privacy and security design matter. Process data in your own controlled environment, limit who can see linked-account views, and follow the PDPA principles of purpose limitation and data minimization. Check guidance from the PDPC and confirm your approach with counsel. Sharing signals between institutions raises further legal questions, so treat it as a separate project. Our AI infrastructure practice covers in-country and on-premises deployment.
Running a pilot in shadow mode
Shadow mode lets you learn without risk to customers. The model scores live traffic, but only your existing rules act. Investigators then review a sample of what the model flagged and what the rules flagged.
Compare the two sets in three groups: cases both caught, cases only the rules caught, and cases only the model caught. The last group shows what the model adds, and the second shows what it misses. Ask investigators to label the model-only cases as useful, not useful or unclear, and record why. After several weeks you will know whether the model earns a role, and which thresholds fit your team's capacity.
What can go wrong
Too many false positives. Honest customers get blocked, such as small businesses that receive many payments. Measure alert quality and review who is flagged.
Biased labels. If past investigations focused on certain groups, the model learns that focus. Review outcomes across customer segments.
Slow data. A signal that arrives hours late cannot stop a transfer made in seconds. Check data latency early.
Overconfidence. No model catches everything. Keep manual reporting channels and staff awareness.
Poor customer handling. A held account needs a clear, respectful path to resolution.
When not to do this: if your transaction data is fragmented across systems or your team cannot review the alerts, fix those foundations first. A model that produces alerts nobody can review adds noise.
How to start
Begin by listing the fraud typologies your investigators see, the data you have on each, and how long it takes to reach an investigator today. Choose one flow, such as new-payee transfers, and run a model in shadow mode next to your rules. Compare what each catches and what each misses, and let investigators judge the quality of the alerts.
For the onboarding side of the same problem, read Thai KYC With Document AI: How It Works. To talk through your detection setup, book a consultation.
Frequently asked questions
What is a mule account?
A mule account is a bank or e-wallet account used to receive and pass on money for someone else, often to hide its origin. The holder may be complicit, tricked or a victim of identity misuse, so responses should be proportionate.
Why do rules miss mule accounts?
Rules usually evaluate one transaction or account at a time and use visible thresholds. Mule activity is spread across many small transfers and linked accounts, and it changes once criminals learn the limits.
What does graph analysis add?
It treats accounts, devices and transfers as a network, so it can show many accounts paying into one, shared devices or fast chains of transfers. Investigators then act on the connected group, not a single alert.
Should AI block accounts automatically?
Keep a person in the loop for serious actions. A common pattern lets low-risk activity pass, adds friction at medium risk and sends high-risk cases to an investigator with reasons logged. Every action should be reversible.
Ready to put this to work?
Turn this playbook into a working system
From a first agent to a national platform, the same team plans it, builds it and runs it with you.
