Rewiring Financial Crime Programs with AI: From Reactive to Relentless

Financial crime programmes were largely designed around a sequential operating model. Customers are assessed during onboarding, transactions are screened against predefined scenarios, alerts enter queues, investigators review cases and controls are recalibrated after weaknesses become visible. The process is structured and auditable, but often slow.

Criminal behaviour is moving in the opposite direction. Fraud networks test controls continuously, move across channels and change identities, devices, payment methods and laundering routes when friction appears. Artificial intelligence gives criminal actors additional speed, scale and personalisation, while institutions may still require weeks or months to translate a newly observed threat into a production control.

Rewiring financial crime programmes with AI does not mean replacing rules, investigators or accountable decision-makers with autonomous systems. It means redesigning how information is collected, connected, interpreted and converted into action. The objective is a control environment that learns from outcomes, adapts to emerging risks and reduces the distance between detection, investigation and prevention.

A relentless programme is not uncontrolled or indiscriminate. It is continuously informed, intelligence-led and capable of improving before the same criminal method causes repeated harm.

Listen the podcast

Watch the video

Why financial crime programmes remain reactive

Traditional controls are often organised around individual products, transactions and regulatory obligations. KYC teams assess customer identity and expected activity. Fraud teams focus on authentication and payment risk. AML teams monitor transactions. Cybersecurity teams investigate compromised credentials and devices. Sanctions teams screen names and payments.

Each function may operate effectively within its own perimeter while missing the wider criminal sequence.

A stolen credential may first appear as a cyber indicator. The subsequent login may be assessed by an identity system, the beneficiary creation by a fraud model and the movement of funds through connected accounts by AML monitoring. Unless these events are connected, the institution sees several weak signals rather than one coherent attack.

Legacy monitoring also relies heavily on predefined rules and thresholds. These controls remain useful for known risks, regulatory requirements and easily explainable patterns. However, they can become imprecise when used as catch-all detection mechanisms. Broad scenarios generate large alert populations, while sophisticated activity structured below thresholds or distributed across related entities may remain hidden.

The Wolfsberg Group has encouraged institutions to move beyond dependency on automated transaction monitoring as the dominant pillar of suspicious-activity detection. Its framework emphasises observed or crystallised risk, improved risk-indicator coverage, better-quality leads and effectiveness measures linked to investigative and law-enforcement value—not simply alert volumes.

The strategic problem is therefore not a complete absence of data or controls. It is the delay and fragmentation between them.

What rewiring with AI actually means

AI transformation is sometimes reduced to adding a machine-learning model to an existing monitoring system or introducing a generative assistant into case management. These interventions may produce local efficiencies, but they do not necessarily change the operating model.

A genuinely rewired programme begins with an integrated risk architecture. Customer information, transactions, counterparties, devices, authentication events, communications, adverse intelligence, prior alerts and investigation outcomes are made available through governed data services. Analytical tools then assess activity at customer, account, transaction, device and network level.

Machine learning can identify nonlinear relationships and behavioural deviations that fixed rules struggle to represent. Natural-language processing can extract information from documents, news, payment references and investigator notes. Graph analytics can reveal shared infrastructure and indirect relationships. Generative systems can summarise evidence, propose investigative steps and draft reviewable case narratives.

These capabilities become more valuable when connected through feedback loops. Confirmed fraud, suspicious activity reports, account closures, false positives, customer explanations and external intelligence should inform subsequent detection and prioritisation.

FATF notes that advanced analytics and data pooling can help institutions analyse structured and unstructured information, identify patterns more effectively and reduce false positives. It also stresses that collaborative analytics must respect privacy, data-protection and fundamental-rights frameworks.

The objective is not to create one all-knowing model. It is to ensure that each control contributes intelligence to the wider system.

How AI changes detection across the customer lifecycle

At onboarding, AI can strengthen identity and risk assessment by evaluating document integrity, device characteristics, behavioural biometrics, contact information, network relationships and consistency between the application and independent data.

The value lies in combination. A new customer using a recently issued identity document is not necessarily suspicious. The risk changes where the same device has opened several unrelated accounts, the address appears across a known mule network and the declared employment is inconsistent with external information.

During account access, behavioural models can assess how the customer interacts with the service. Typing patterns, navigation, session velocity, device changes and authentication behaviour may indicate account takeover, remote-access manipulation or automated activity. These signals can influence payment controls before funds leave.

Transaction models can move beyond single-payment rules by evaluating sequences and networks. A transfer may appear ordinary in isolation but become higher risk when preceded by password recovery, a telephone-number change, internal consolidation of savings and creation of a first-time beneficiary.

Network analytics is particularly important for mule accounts, organised fraud and complex laundering. Criminal groups frequently change individual accounts while reusing devices, addresses, beneficiaries, merchants, companies, cryptocurrency wallets or cash-out routes. Graph models help institutions identify the operating structure rather than only its most visible transaction.

The Bank for International Settlements has highlighted the potential for machine learning and data integration to reduce false positives and improve detection of fraudulent and potentially illicit transactions. It also identifies a role for emerging AI co-pilots in supporting human financial-crime reporting activity.

From alert generation to intelligence-led investigation

AI can improve investigations before it changes the underlying detection model.

Investigators frequently spend time gathering information from separate systems, normalising transaction data, identifying linked parties and constructing timelines. An AI-enabled workspace can assemble relevant evidence, calculate totals, group counterparties and highlight contradictions for review.

Generative tools can then convert complex case information into a structured summary. They may explain which factors contributed to the alert, identify missing information and recommend proportionate next steps such as reviewing a related account, contacting the customer or escalating sanctions concerns.

The control boundary is critical. The system should distinguish verified facts from model-generated inference. Investigators need to see the transactions, documents or signals supporting each conclusion. A plausible narrative that cannot be traced to evidence creates more risk than an inefficient manual process.

AI can also improve prioritisation. Instead of treating every alert generated by a scenario as equally important, models can consider customer risk, network centrality, typology relevance, value at risk and the probability that immediate intervention will prevent loss.

The strongest performance measures therefore extend beyond the number of alerts closed. They include true-positive precision, coverage of priority threats, time to intervention, funds prevented or recovered, network accounts identified and the quality of suspicious activity reporting.

Wolfsberg’s transition framework similarly recommends evaluating priority-risk coverage, expanded risk indicators, precision, recall, report quality and downstream integration. It warns that new AI-based approaches should not be judged solely on whether they reproduce every output of the legacy system.

How generative and agentic systems fit

Generative AI is most defensible when used as a controlled co-pilot. It can summarise customer files, translate complex network activity, draft information requests, compare evidence against procedures and prepare first versions of investigation or reporting narratives.

Agentic systems extend this model by carrying out multi-stage tasks. A constrained agent might retrieve approved case data, calculate transaction flows, search internal intelligence, draft a chronology and present recommended actions to an investigator.

The distinction between assistance and authority remains fundamental. An agent may orchestrate workflow steps, but high-impact decisions—freezing funds, exiting customers, alleging first-party fraud or submitting regulatory reports—require accountable human oversight and defined approval controls.

AI adoption is already significant across financial services. A 2024 Bank of England and FCA survey found that 75% of responding firms were using AI and another 10% planned to do so within three years. Foundation models represented 17% of reported use cases, while one third of AI implementations involved third parties. However, 46% of respondents reported only a partial understanding of the technologies they used.

This creates an important warning: operational adoption may advance faster than institutional understanding.

Why AI transformation can fail

The first failure mode is poor data. Fragmented identifiers, missing transaction context, inconsistent typology labels and weak investigation outcomes limit what models can learn. AI can accelerate analysis, but it cannot reliably infer economic reality from incomplete or incorrectly joined records.

The second is automation bias. Investigators may accept model conclusions because they are presented confidently or because challenging them requires additional work. Human review becomes ineffective where the reviewer cannot understand the model’s reasoning or access the underlying evidence.

The third is model drift. Customer behaviour, products, criminal typologies and payment channels change over time. A model performing well during validation may deteriorate unless outcomes and population changes are monitored continuously.

The fourth is bias. Historical investigations and reports may reflect uneven detection, customer selection or escalation practices. Training models on those outcomes without adjustment can reproduce and amplify previous weaknesses.

The fifth is adversarial manipulation. Criminals can test onboarding, authentication and transaction controls, then alter their behaviour to reduce risk scores. Generative models may also be exposed to prompt injection, manipulated documents or contaminated external information.

The sixth is third-party dependency. Institutions may rely on externally hosted models, datasets and infrastructure that they do not fully understand. Concentration, data location, model updates, service interruption and access to validation evidence require careful governance.

NIST’s AI Risk Management Framework and its generative-AI profile emphasise lifecycle risk management, testing and alignment between AI use, organisational objectives and risk tolerance.

What a resilient control stack looks like

The first layer is clear use-case accountability. Each AI system should have a defined purpose, owner, users, data boundary and level of permitted autonomy. Institutions should specify which decisions the system may recommend, support or execute.

The second layer is governed data lineage. Teams must understand where information originated, how it was transformed, which identifiers were used and whether the data remains suitable for the control objective.

The third layer is model validation proportionate to risk. Validation should assess methodology, data quality, precision, recall, stability, explainability, bias, security and operational performance. Generative applications require additional testing for hallucination, omission and unsupported inference.

The fourth layer is human oversight designed around the actual decision. A reviewer needs sufficient time, expertise and evidence to challenge the output. Placing an approval button after an opaque automated process does not create meaningful control.

The fifth layer is continuous monitoring. Institutions should track model drift, changes in alert populations, customer impact, overrides, false negatives and emerging typologies. Confirmed incidents should become structured feedback rather than remaining in narrative case files.

The sixth layer is operational resilience. Firms need fallbacks for unavailable models, corrupted data, vendor outages and material performance failures. High-risk decisions should not depend on a single external technology service.

The FCA’s current approach relies on existing frameworks, including governance expectations, the Consumer Duty and the Senior Managers and Certification Regime. It is focusing on how firms oversee AI, test models, monitor outcomes, explain decisions and protect vulnerable customers.

How institutions should implement the transition

Implementation should begin with clearly defined financial-crime outcomes rather than a general mandate to “use AI”. Suitable early use cases include evidence retrieval, alert enrichment, entity resolution, network visualisation, adverse-information analysis and investigator summarisation.

Systems should initially operate in controlled test or shadow environments. Historical cases can be used to measure whether the new approach identifies priority threats, reduces unproductive work and produces investigator-ready intelligence.

Parallel comparison with legacy controls may provide useful evidence, but the objective should not be exact duplication. A new model may legitimately stop generating low-value alerts while identifying risks that the previous system missed.

Investigators, model developers, data engineers, compliance specialists and model-risk teams should work together from design through deployment. Technology teams need to understand the typology and regulatory decision; financial-crime teams need enough technical knowledge to challenge the model.

Deployment should be incremental, with defined exit criteria, rollback arrangements and independent assurance. The programme should measure not only efficiency but whether AI improves prevention, investigation and intelligence outcomes.

What this means for financial crime leaders

Rewiring a financial crime programme with AI is not primarily a procurement exercise. It is a redesign of data, controls, decision rights, investigative work and institutional learning.

Leaders should ask whether the organisation can connect risk across the customer lifecycle, whether confirmed cases improve future controls and whether investigators receive intelligence rather than unstructured alert volume.

They should also resist two unhelpful extremes. The first is assuming that AI will solve weaknesses in data and operating processes automatically. The second is retaining ineffective controls indefinitely because innovative models do not reproduce every historical alert.

The transition from reactive to relentless depends on disciplined adaptation. Rules, machine learning, network analytics, generative tools and human expertise should operate as complementary components of one control environment.

Criminal organisations already learn from every failed attempt. Financial institutions need programmes capable of doing the same—without sacrificing explainability, proportionality or accountability.

The strongest financial crime defences will not be those that automate the most decisions. They will be those that convert the widest range of reliable signals into timely action, learn continuously from outcomes and adapt faster than the threats they are designed to disrupt.