Global Advice Network
| Borgergade 111 | DK - 1300 Copenhagen K
|
|
|
|
Machine learning for real-time anti-corruption controlsPayments now clear faster than many compliance teams can investigate them. Account-to-account transfers, card authorisations, digital wallets and buy-now-pay-later services can settle in seconds, giving machine learning a role in identifying suspicious activity before funds become difficult to recover. The technology should prioritise transactions for review, not declare that a customer or business is corrupt. Its value lies in detecting meaningful patterns across high-volume payment flows. Australia’s payment environment combines established banks, fast-growing fintech companies, card networks and buy-now-pay-later providers. PayID and the New Payments Platform support rapid account transfers, while Sydney and Melbourne remain major centres for corporate activity and cross-border commerce. For entities covered by Australia’s anti-money laundering and counter-terrorism financing framework, AUSTRAC obligations add another layer to transaction monitoring. A payment may be technically legitimate yet still conceal trade-based laundering, invoice manipulation or proceeds routed through corrupted supply chains. A useful model combines country, customer, counterparty and transaction-level information. It can consider payment frequency, sudden changes in value, unusual beneficiary relationships, repeated round amounts and deviations from established behaviour. Because a single unusual feature can have an innocent explanation, effective detection depends on context and explainable reason codes. Separate models may also be needed for payment fraud and anti-corruption risk, since fraudulent payments and concealed corrupt activity can produce different signals. Real-time prediction should support proportionate intervention rather than replace professional judgment. Australian teams must also consider privacy obligations, customer transparency, investigation capacity and the risk of unfairly blocking legitimate commerce. In practical terms, the process should be “sorted properly”: documented, repeatable and designed to give customers and compliance staff a fair go. This requires governance that treats model performance, data quality and human review as connected controls rather than separate technical work. Why transaction-level risk scoring mattersPeriodic reviews can identify suspicious patterns eventually, but that delay may allow money to move through several accounts before anyone examines the activity. A transaction-level model evaluates risk as payments occur and can recommend enhanced checks before settlement. It also reduces the volume of alerts by ranking transactions according to their likelihood of warranting investigation, which helps teams focus on cases with the greatest potential impact. Risk is strongly connected to geography, industry and payment corridor. Exposure to bribery, sanctions, organised crime or weak institutional controls may vary considerably between markets, although country-level indicators must never become automatic assumptions about an individual customer. The Business Anti-Corruption Portal’s country risk profiles can help compliance teams understand those differences and align due diligence with the jurisdictions in which they operate. Transaction scoring also captures risks that a customer-level review alone may miss. A supplier with a long-standing clean history may suddenly change bank accounts, begin sending round-dollar payments to unrelated entities or route invoices through a new intermediary. Conversely, an unfamiliar counterparty may be assessed more fairly when evidence is combined across the payment, rather than treating every first-time transaction as suspicious. The objective is to connect individual events to a coherent pattern. Building a reliable high-risk detection modelThe model should begin with a clearly defined question, such as identifying payments likely to require enhanced due diligence. Relevant features can include payment velocity, amount deviations, counterparty networks, device and IP changes, beneficiary modifications, transaction sequencing and inconsistencies between invoice data and payment behaviour. For organisations exposed to trade-based corruption, invoice values, freight details, customs patterns and supplier relationships may be especially informative. Different techniques suit different problems. Gradient-boosted trees often provide a strong baseline and useful reason codes, while sequence models can detect changes in payment order. Graph-based methods are valuable when evidence is spread across customers, accounts and suppliers. Unsupervised anomaly detection can expose previously unknown patterns, although it needs careful interpretation because statistical rarity does not automatically indicate corruption. A practical ensemble may combine these approaches with transparent rules covering sanctions screening and mandatory reporting thresholds. Training data requires particular care. Confirmed corruption cases are rare, while labels based on investigations may be delayed, incomplete or influenced by which analysts happened to review an alert. Weak supervision, historical payment data and synthetic examples can help, but the team should test whether the model is merely reproducing existing investigation gaps. Time-based validation is usually more realistic than random train-and-test splits because fraud patterns and business relationships change over time. Concept drift should be monitored once the model is live. Turning scores into defensible payment decisionsA live score should normally combine inherent risk with behavioural change. An otherwise ordinary transaction can become more concerning when it combines a high-risk corridor, a recently altered beneficiary and an amount inconsistent with the customer’s history. The system should produce a risk band, relevant reason codes and a recommended level of review rather than an unexplained binary result. This makes decisions easier for investigators, auditors and internal assurance teams to understand. Possible interventions include immediate approval, additional authentication, a request for supporting documentation, enhanced due diligence or temporary review by a specialist. Thresholds should reflect the organisation’s risk appetite, payment value, customer type and available investigation capacity. A high score should not itself be treated as proof of criminal conduct or as the sole basis for a suspicious matter report. Analysts must assess the underlying facts and document why the recommended action was accepted, changed or rejected. That decision trail must remain available throughout the payment lifecycle. The Business Anti-Corruption Portal’s guidance on retaining compliance documentation reinforces the need to preserve model versions, input data, reason codes, analyst notes and final outcomes. Records should show which rules or features influenced a decision at the time, while also meeting Australian privacy and retention obligations. Without this evidence, even an accurate model may produce outcomes that are difficult to defend. Running the model in live operationsReal-time architecture must balance speed, resilience and control. Payment events should pass through validated data pipelines and receive a score within the service-level window required by the payment provider. Models need fallback rules if a data feed fails, an inference service becomes unavailable or confidence falls below an approved threshold. Security testing should cover data poisoning, model extraction, adversarial payment patterns and unauthorised changes to thresholds or feature definitions. Alert workload is an operational constraint that requires direct attention. If every borderline transaction becomes a full investigation, analysts may begin approving common alerts simply to keep up. Teams should measure review capacity, alert ageing, override rates, case quality and the proportion of alerts producing useful outcomes. A non-compliance setting offers a useful operational analogy in guidance about short play rounds: brief, purposeful activity alternates with calmer periods rather than sustained intensity. Similarly, real-time scoring can identify candidates rapidly while leaving quieter periods for investigation, calibration and model improvement. Customer treatment should be included in performance testing. Legitimate Australian payments must not be delayed disproportionately because of location, language, occupation or another proxy for risk. Clear escalation paths should allow a customer to correct inaccurate information, and dual controls should govern threshold changes. For high-value or sensitive flows, a second reviewer or stronger authentication may be appropriate before funds are released or a transaction is reported. Practical recommendations for controlled deploymentImplementation should proceed as a controlled risk-reduction programme rather than an unrestricted automation project. Teams can use the Business Anti-Corruption Portal’s anti-corruption resources to compare policies, terminology and due diligence expectations across markets before designing the system. Before production deployment, consider the following controls:
The first pilot should focus on a narrow payment stream where reliable labels, clean data and experienced reviewers already exist. Its purpose is to test whether machine learning surfaces relevant suspicious activity earlier without creating excessive customer friction or investigative work. Results should then be compared with the existing control baseline, including the cases that would have been missed and the legitimate transactions that were delayed. The practical takeaway is to begin with explainable signals, proportionate human oversight and a controlled payment scope, then expand only when evidence shows that suspicious activity is identified earlier and decisions remain defensible. |