Introduction
Traditional transaction monitoring was designed for a payment world that no longer exists. For firms operating African-corridor mobile money, the question is no longer whether rules-based transcation monitoring is adequate, it is not, but whether Agentic AI can fill the gap responsibly, explainably, and in a way regulators will accept. This article argues that Agentic AI represents a material advance over rules-based monitoring for African-corridor payments. But the argument is not primarily technological. It is regulatory, ethical, and operational. The technology only earns its place if it can be governed, explained, audited, and demonstrated to be fair to the communities it monitors.
Why the Rules-Based Model Fails Here
As discussed in my previous article (How Mobile Money Is Changing the AML Playbook), rules-based transaction monitoring was built for structured instruments moving through predictable rails. The typologies were known, transaction volumes were manageable, and customer behaviour in Western financial contexts was broadly predictable. That world is gone.
In its place is a mobile money ecosystem processing billions of low-value, high-frequency cross-border transfers across corridors that existing monitoring tools were never designed to understand. The UK–Nigeria, UK–Ghana, and UK–Kenya corridors alone serve millions of diaspora customers whose behaviours, irregular timing, mobile wallet networks, reliance on informal documentation, and seasonal remittance spikes tied to school fees, harvests, religious festivities, and family obligations, routinely satisfy standard alert conditions without any criminal intent whatsoever.
Adding more rules has not solved this. It has produced alert queues so overwhelmed that genuine risks are missed, compliance teams spending most of their time on data retrieval rather than judgement, and false positive rates that bear no relationship to actual financial crime prevalence. Three structural failures drive this and no amount of rule-writing resolves them.
Volume Misalignment: Mobile money operates at transaction frequencies never contemplated when alert queue architectures were designed. A single active corridor customer may generate dozens of transactions per week. Rules engines scale alert generation linearly with volume, producing alert fatigue so severe that compliance teams operate in permanent triage.
Behavioural Misclassification: The legitimate payment behaviours of diaspora mobile money users, multiple small transfers, use of wallets rather than named accounts, irregular timing, and routing through family members, satisfy many standard alert conditions. This is not a calibration problem that threshold adjustment can fix. It is a structural mismatch between the typologies encoded in rules and the actual financial lives of the communities being served.
Typology Lag: SIM-swap fraud, mobile wallet structuring across networks of informal agents, and account takeover exploiting multi-SIM environments bear no resemblance to the pattern’s rules were written to catch. Rules capture yesterday’s threats while today’s evolve faster than analysts can respond. As the FCA’s guidance on transaction monitoring under SYSC 6.3[i] makes clear, firms must have systems commensurate with the nature, scale, and complexity of their business. A monitoring system generating above 80% false positives is not a compliant system. It has substituted activity for effectiveness.
What Agentic AI Does Differently
Many firms already deploy artificial intelligence and machine learning in transaction monitoring and broader compliance processes, typically to improve alert scoring, reduce false positives, and enhance pattern recognition. These systems are reactive: they analyse data and surface outputs for human review. They are decision-support tools, not decision-making agents. The term “AI-powered compliance” has been applied so broadly that it has lost precision. Agentic AI is something meaningfully different. An agentic system pursues defined goals through sequential, autonomous action. It does not simply flag a condition and wait. It can be designed to plan an investigative sequence, execute it, incorporate the results, adapt its approach based on what it finds, and deliver a structured, evidenced output. In an AML compliance context, this is the difference between an alert and a case file.
Smarter Investigations and SAR Preparation
In a rules-based environment, compliance officers spend the majority of their working time gathering the information needed to make a decision: pulling transaction history, querying sanctions databases, reviewing onboarding records, and checking prior alert outcomes. The decision itself, the act of contextual judgement that requires a human, takes a fraction of total time.
An agentic system inverts this entirely. Retrieval is autonomous. The analyst receives a structured case file with risk factors identified and the specific signals driving them already documented. Their time goes entirely to contextual judgement, decision-making, and accountability. That is where human time should be spent.
Agentic AI can also assemble transaction narratives and draft SAR reports for analyst review, meaningfully reducing the administrative burden on already stretched compliance teams. But this must be stated clearly: the MLRO retains full personal accountability for every SAR decision under POCA 2002 and the Money Laundering Regulations 2017. AI assistance does not dilute that liability.
Under FCA DP5/22[ii], every AI-generated recommendation must be traceable and explainable. Firms must be able to demonstrate to a compliance officer or regulator precisely how a decision was reached, what data informed it, and who reviewed it. Modern agentic systems can produce structured reasoning chains that document the specific features driving a risk assessment, the weight given to each, and the basis for escalation or closure. This is not merely more useful than a raw alert. It is the foundation of a defensible audit trail.
Corridor-Specific Intelligence at Scale
A UK domestic payment model and a UK-Nigeria remittance model should look entirely different, different behavioural baselines, typologies, seasonal patterns, counterparty risk profiles, and documentation standards. In theory, rules engines can be segmented by corridor. In practice, maintaining separate and current rule sets for a dozen corridors, each with its own evolving threat landscape and customer behaviour profile, requires analytic capacity that few firms can sustain. The result is either lowest-common-denominator rules applied across all corridors, or corridor-specific rule sets that are months out of date.
AI models can be trained and continuously recalibrated at corridor level without proportional increases in analyst workload, so emerging patterns are reflected in the model within weeks rather than months. But this only works if the training data is right. A model trained primarily on Western bank account holder behaviour will systematically misclassify African-corridor payment behaviour, not through malice, but through ignorance. The effect on the customer is identical. Higher false positive rates mean higher rates of account review and transaction delay for communities that pose no genuine risk. An AI model that has never adequately learned the legitimate behaviour of the community being served is not a compliance outcome. It is a harm, one with real consequences for real households, and one that regulators are increasingly scrutinising.
FCA Consumer Duty[iii] requires firms to ensure AI-driven decisions deliver good outcomes across products and services, price and value, consumer understanding, and consumer support. Pre-deployment bias testing across corridors and customer segments, with ongoing false positive and false negative monitoring, and documented evidence that training data meaningfully represents African-corridor payment behaviour, is not optional. It is a governance prerequisite and for FCA- supervised firms, a component of their obligation to treat customers fairly.
Dynamic Sanctions and PEP Screening
PEP status in African markets is highly dynamic. Government transitions, ministerial appointments, and electoral cycles mean a customer’s status can change materially between onboarding and their next transaction. HM Treasury, OFAC, and the UN Security Council publish list changes without advance notice. An agentic system re-screens the entire customer base the moment those lists update and autonomously escalates matches for human review. A rules engine screens at execution. The lag in between, which in a rules-based environment can extend to days or weeks, is a genuine regulatory and reputational vulnerability that agentic AI can close.
The Opportunity and the Responsibility
Agentic AI has real potential to address the structural challenges facing African-focused payment firms: alert volume, corridor complexity, evolving typologies, real-time sanctions screening, regulatory horizon scanning and resource-intensive investigations. Firms that adopt it well will achieve faster onboarding, fairer outcomes, and the regulatory responsiveness that builds durable competitive advantage.
But the opportunity and the responsibility are inseparable. The MLRO’s personal liability does not transfer to an algorithm. The FCA’s expectations of explainability and human accountability do not diminish because a process is automated.
Firms that deploy agentic AI with strong governance, quality data, meaningful human oversight, and a clear accountability framework will be better positioned to scale safely across Africa’s rapidly evolving payments ecosystem. The technology is ready. The question is whether the governance is.
How we can help
Contact us via info@opselcompliance.com or +447950377849 if you need support with system review, selection, or implementation in financial crime compliance.
Open to RegTech partnerships too, if you’re looking to collaborate and reach relevant clients through our network, let’s connect.
References
[i] [i] https://handbook.fca.org.uk/handbook/sysc6
[ii] https://www.bankofengland.co.uk/prudential-regulation/publication/2022/october/artificial-intelligence
[iii] https://www.fca.org.uk/publication/corporate/ai-update.pdf
