What problem does it solve?
Scams and most fraud need a place to land the money. Mule accounts, opened by fraudsters or run by recruited account holders, receive the proceeds and move them on quickly, for example to other banks, crypto exchanges, cash machines or remittance services. By the time the victim reports, the money has often left.
A single bank looking at one account at a time sees little: a new account with some incoming transfers. The pattern only shows in the network, such as many senders who are scam victims, a fan in and fan out shape, shared devices and addresses, or accounts that were opened in a burst. Regulators are also shifting scam losses onto firms. Under the UK reimbursement rules for authorised push payment scams, the sending and receiving firms split the cost of reimbursing victims equally, so for a receiving bank detecting mules is now a loss and compliance issue, not only a crime prevention one.
- The US Federal Trade Commission reports that in 2024 consumers reported losing more money to scams paid by bank transfer or cryptocurrency than through all other payment methods combined.New FTC Data Show a Big Jump in Reported Losses to Fraud to $12.5 Billion in 2024 (2025)
How does it work?
- Build the graph. Accounts, customers, devices, IP addresses, addresses, phone numbers and payment flows become nodes and edges, updated continuously from onboarding and payment data.
- Score accounts and communities. Behavioural models score each account for mule like activity (rapid in and out, pass through balances, sudden change after dormancy), and graph algorithms find clusters and chains that share signals.
- Bring in external signals. Scam reports from other banks, confirmation of payee mismatches, industry or central bank mule lists and law enforcement requests enrich the scores.
- Assemble the case. An agent drafts a fund flow timeline and case narrative for each cluster: who received what from whom, where it went next, and which signals link the accounts.
- Decide and act. An investigator decides on restrictions, exits, recall requests to peer banks and reporting, and the decision and reason are recorded against every account touched.
- Audience
- Back office
- Autonomy
- Copilot
- Adoption
- Early adopters
- Channels
- Internal tools, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Interactions handled | Not pooled | at least 180 million | 1 | 1 vendor |
Value drivers: Risk and loss reduction, Compliance quality, Speed and cycle time.
Indicative value
A retail bank receiving scam proceeds in 2,000 reported cases a year
USD 150,000 to USD 1.2 million
Scam proceeds frozen or recovered per year
How this is calculated
Formula: cases * averageLoss * extraRecovery. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Scam cases per year where the bank received the funds cases, cases per year | 2,000 | 2,000 | The reference bank. Replace with your own count of inbound scam reports. |
| Average amount received per case averageLoss, USD per case | 1,500 | 4,000 | Editorial assumption. Replace with your own data. |
| Additional share of funds frozen or recovered through earlier detection extraRecovery, fraction of funds | 0.05 | 0.15 | Editorial assumption. Public deployments rarely disclose recovery rates, so keep this low until you have your own results. |
What it leaves out: Covers recovered funds only. It leaves out reimbursement liabilities avoided under schemes that share losses with the receiving bank, the investigation time saved, regulatory benefits, and the cost of false restrictions on genuine customers.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Reserve Bank Innovation Hub (Reserve Bank of India)
India · Government and public sector · 2025
The Reserve Bank Innovation Hub, a subsidiary of the Reserve Bank of India, built MuleHunter.AI, a machine learning model that helps banks detect mule accounts used to move fraud proceeds. The RBI announced the pilot with two large public sector banks in December 2024, and the Governor said in October 2025 that it had been scaled from about 5 banks to 21 banks, using system wide learning. The RBI has declined to disclose how many mule accounts it has identified.
- Users served: 21, banks using the model, October 2025
"MuleHunter.ai, developed by the Reserve Bank Innovation Hub has been scaled up from about 5 banks at the beginning of this year to 21 banks."
Claimed by: organization
ANZ, Commonwealth Bank, NAB, Suncorp Bank and Westpac (BioCatch Trust Australia)
Australia · Banking · 2025
In November 2024 ANZ, Commonwealth Bank, NAB, Suncorp Bank and Westpac joined BioCatch Trust Australia, launched as a pilot of an interbank network that shares behavioural and device intelligence about receiving accounts, so the sending bank can review a payment to a likely mule account before money leaves. BioCatch reports that in the third quarter of 2025 the network analysed more than 180 million payments and revealed more than $60 million in attempted fraud (currency not stated in the release), and that two more institutions, including Macquarie Bank, have since joined.
- Interactions handled: at least 180 million, payments analysed in the third quarter of 2025
"in the third quarter of 2025 alone, analyzed more than 180 million payments totaling more than $330 billion, revealing more than $60 million in attempted fraud."
Claimed by: vendor
BigPay
Malaysia · Payments and cards · 2025
BigPay, a Malaysian electronic money platform that was receiving over 1,000 reported mule cases a month, worked with Feedzai to turn mule patterns into detection rules on its existing Feedzai platform: alert logic for suspicious inbound payments, decline rules based on funding velocity and beneficiary risk, and rules that block cash out channels such as crypto, ATMs and remittances. The source attributes BigPay's result to this analyst built rule set and presents Feedzai's AI assisted alert prioritisation as the platform's next layer of defense, not as the cause of the outcome. The vendor reports that BigPay neutralised a new mule network surge within 72 hours and reduced overall mule activity by more than 90 percent within 60 days.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Payment flows with counterparty identifiers, including inbound instant payments
- Onboarding data, device and session data linked to accounts
- Confirmed mule and scam case outcomes for training and evaluation
- Access to industry or central bank mule intelligence where it exists
Systems to integrate
- Payments hub and core banking system
- Onboarding and identity verification systems
- Fraud and AML case management
- Industry data sharing schemes and peer bank recall processes
- Account restriction and exit workflows
Complexity: High
Graph analytics needs clean entity resolution across customers, devices and counterparties, and the best signals come from outside the bank. Restricting accounts is high impact, so the decision process and customer remediation need as much work as the model.
- 1
Start from confirmed cases
Collect every confirmed mule account and inbound scam case from the last two years and map the signals they shared. This becomes the training set and the benchmark.
- 2
Resolve entities before modelling
Link customers, accounts, devices and contact details reliably; poor entity resolution produces false networks that lead to wrong restrictions.
- 3
Combine rules, behaviour and graph
Start with known typologies as rules, add behavioural scoring, then graph features and community detection, and measure the lift each layer adds on the benchmark set.
- 4
Give investigators the network view
Provide a visual network and a drafted fund flow narrative per cluster, so investigators can act on a whole network at once instead of account by account.
- 5
Connect to the outside
Join industry intelligence sharing and agree recall and freeze procedures with peer banks, so detection turns into recovered funds.
Guardrails
- Account restrictions and exits only by a trained investigator, with the reason recorded
- Fast review and remediation route for customers restricted in error
- Graph links shown with the evidence behind them, never as an unexplained score
- Regular testing for disparate impact across customer groups, for example by age, nationality and student status
- Data sharing with peer banks only under the legal gateway that allows it
KPIs to instrument
- Mule accounts identified per month and the share confirmed on investigation
- Time from first inbound scam payment to restriction
- Value of funds frozen or recovered
- Share of restricted customers released after review
- Inbound scam reports from peer banks per million accounts
Human in the loop
The models and agent find and assemble; investigators decide. Every restriction, exit, recall request and report is a documented human decision, and a second line reviews samples of both actioned and dismissed clusters.
Common failure modes
- Networks built on bad links
- Shared addresses in student housing or shared devices in families create false clusters. Weight links by strength and require investigator review of the evidence.
- Detection without recovery
- Mules are found after the money has moved on. Measure time to restriction, not only detection counts, and connect to recall processes.
- Targeting the recruited, missing the organisers
- Restricting individual mules without mapping the network leaves the organisers active. Work at cluster level and share intelligence.
What are the risks and rules?
EU AI Act
Minimal risk
Detecting mule accounts is fraud and AML detection by a private firm, which Annex III does not list; point 5(b) explicitly excludes systems used to detect financial fraud from the credit scoring category. Restricting an account based solely on an automated score can be a decision with similarly significant effects under GDPR Article 22, so keep a human decision and a route to challenge.
Rules that apply
Guidance
- APP scams (Payment Systems Regulator, Europe). UK reimbursement for authorised push payment scams over Faster Payments and CHAPS is split 50:50 between sending and receiving firms, which puts mule detection on the receiving bank's balance sheet.
- Guidelines on Shared Responsibility Framework (Monetary Authority of Singapore, Asia Pacific). Singapore's framework, in force since 16 December 2024, that assigns anti phishing duties to financial institutions and telcos and requires payouts to scam victims where those duties are breached.
- COSMIC, Collaborative Sharing of ML/TF Information and Cases (Monetary Authority of Singapore, Asia Pacific). Platform launched by MAS with six major banks in April 2024 for sharing red flag information on customers across institutions. It currently covers misuse of legal persons, trade finance and proliferation financing, not retail mule accounts.
- Framework for Responsible and Ethical Enablement of Artificial Intelligence (FREE-AI) Committee report (Reserve Bank of India, Asia Pacific). India's framework for responsible AI in the financial sector (August 2025), which names fraud detection as a high stakes use and recommends that AI models are validated and tested periodically, including for drift and bias.
Controls to put in place
- Documented decision process for restrictions and exits, with recorded reasons
- Customer remediation route with a service level for review
- Model inventory entry, validation and fairness testing
- Legal basis documented for every external data sharing arrangement
- Audit trail linking each restriction to the network evidence and the investigator
Frequently asked questions
- Why does mule detection need graph analytics?
- A mule account often looks ordinary on its own. The signal is in the links: many victims paying in, money leaving quickly to the same onward accounts, and devices or contact details shared across accounts opened around the same time.
- Can banks detect mules together?
- Increasingly yes. In Australia, five large banks joined BioCatch Trust Australia in November 2024 to share intelligence on receiving accounts before a payment leaves. In India, the central bank's innovation hub offers MuleHunter.AI to banks; the Governor said in October 2025 that it had scaled to 21 banks.
- Should an AI model freeze accounts automatically?
- No. Freezing or exiting an account is high impact for the customer and often irreversible in practice. Let the model find and prioritise, and let a trained investigator decide with the evidence in front of them.
How to cite this page
Blits.ai AI Use Case Library, "AI for money mule account and network detection", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/mule-network-detection. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published