Insight

The Risk and Reg Edit: Autumn 2026 edition

From controls to outcomes: What FATF’s mutual evaluations mean for financial institutions

Faraaz Nakvi

The UK’s next FATF mutual evaluation is underway. Evidence gathering has commenced ahead of the onsite assessment in 2027. This fifth round is the UK’s first evaluation under FATF’s 2022 methodology, which places greater emphasis on effectiveness, material risks, and demonstrable outcomes. FATF describes effectiveness as a significant focus of the assessment and requires countries to demonstrate that their frameworks work in the context of the risks they face.

Although mutual evaluations assess countries rather than individual firms, financial institutions should not regard them as somebody else’s examination. Supervisors form part of the system being assessed, and the effectiveness of supervision depends, in part, on whether supervised firms are managing financial crime risk effectively.

There is an important lesson here for firms. Financial crime compliance is increasingly an evidence problem, not simply a control framework problem. Having the right policies, systems, and controls matters, but firms also need to be able to demonstrate why those controls address their most material risks and what they achieve.

Firms require an evidence chain connecting regulatory obligation, risk, control, data, decision and outcome – not just individually but end-to-end.

A firm may, for example, demonstrate that its Know Your Customer (KYC) policy reflects regulatory requirements, its transaction monitoring scenarios are operating, and its sanctions alerts are investigated. But can it demonstrate that the combination of those activities addresses its most significant financial crime risks? Can it explain why investment is concentrated in specific controls, how those controls respond to identified threats and whether they are changing the resulting risk?

FATF’s experience demonstrates why the distinction matters. Its analysis of previous mutual evaluations found substantial progress globally in implementing the technical requirements of its recommendations, while countries continued to experience significant challenges in taking effective action in line with the risks they face. The lesson for firms is that technical compliance and operational activity are not the same as effectiveness.

Financial crime rarely follows the firm’s organisational structure

Many institutions still manage financial crime through separate processes spanning customer due diligence, fraud, transaction monitoring, sanctions screening, and investigations. Organisationally this makes sense, but financial crime doesn’t respect those boundaries.

The same customer or network may create signals across fraud, money laundering, and sanctions controls. Assessing those signals independently can therefore leave a firm with individual controls that operate as designed, while still failing to establish a coherent view of the underlying risk.

Firms should apply the effectiveness principle at an institutional level. Can the organisation explain its financial crime risk as a whole, or only explain what each control separately decided?

This also changes the role of data. Data is not simply an input into individual controls, it also allows the firm to connect obligations, risks, and controls to the decisions taken and outcomes achieved. Fragmented customer information or an inability to reconstruct historic decisions, for example, therefore becomes much more than a technology problem. It weakens the firm’s ability to evidence that its framework is effective.

AI will make the evidence chain more important

The risk of assessing financial crime signals in isolation becomes particularly significant as firms introduce AI and agentic models into KYC, screening, investigation, and other financial crime processes.

The readiness test should be whether the evidence on which an agent acts is sufficiently orchestrated, governed, and explainable to support a defensible decision.

Firms will increasingly need to understand which data informed a decision, which obligation and policy applied, which risks were considered, and whether the basis for the decision can subsequently be reconstructed. An AI agent sitting on top of a poorly governed evidence architecture will just automate any weakness.

This is our interpretation of the implications for firms rather than a direct FATF requirement. But it follows naturally from a methodology increasingly focused on demonstrating effective outcomes against material risks.

Five questions firms should be asking now

As the UK progresses towards its mutual evaluation, senior financial crime leaders should ask:

What are our most material financial crime risks, and can we demonstrate that our investment and control intensity genuinely follow them?
Can we trace a material regulatory obligation through policy, risk assessment, control, and data to the outcome it is intended to achieve?
Where a customer presents signals across fraud, AML, or sanctions, can we see the aggregate risk or do our controls continue to assess it in isolation?
Could we reproduce and explain a material automated or AI-assisted decision months after it was made, including the information and policy basis used at the time?
Which controls appear effective according to operational metrics, but for which we can’t demonstrate an actual risk outcome?

These questions move the conversation beyond whether individual controls work, towards whether the financial crime framework works as a system.

The UK’s mutual evaluation will assess the effectiveness of the UK’s regime, not individual firms. But firms should use it as a catalyst to challenge their own evidence base.

The UK Anti-Money Laundering and Asset Recovery (AMLAR) Strategy

Tim Newman

This year saw the launch of the UK government’s Fraud Strategy 2026-2029, alongside the Anti-Money Laundering and Asset Recovery (AMLAR) Strategy 2026-2029. Together with the forthcoming release of Economic Crime Plan 3, these strategies indicate a continued commitment from the government to reform the economic crime ecosystem. The urgency of these reforms is underscored by the fact that just under 867,000 suspicious activity reports (SARs) were filed in 2024/25, with as much as 80-90 percent of those reports potentially classified as low value.

Over the last few years, a significant proportion of the intended reform has focused on embedding and expanding public-private partnerships. This began with the Joint Money Laundering Intelligence Taskforce (JMLIT),extending into more ambitious areas of intelligence and information sharing, such as data fusion and the new Online Crime Centre. Much of this has relied on additional investment from the private sector, which has long argued for a dialling down of what it sees as low-value ‘box-ticking’ activity (as seen in the SARs regime) so it can dial up investments into more strategic areas.

While previous Economic Crime Plans have embraced the concept of improving the ‘effectiveness’ of the AML regime, the volume of SARs (and, by implication, the amount of money spent on low-value activity) has stayed at stubbornly high levels, indicating that this ambition has not translated into reality. However, the AMLAR Strategy changes the language from ‘effectiveness’ to reducing low-value activity; a subtle but significant shift to a less abstract ambition. The government’s challenge will be to show that this is not letting banks off the hook by allowing overall AML spend to fall, but that investment is being rebalanced into more useful areas.

To support government, particularly on the review of the suspicion threshold, firms will need to provide a solid base of evidence for how rebalancing can work. That means they will need to consider the implications for their onshore UK workforce and, if they are a multinational with global AML spend, the implications for areas of spending such as IT systems, which are difficult to apportion by country. To support effective dialogue with the government, banks will need to think about how this is calculated and presented back.

Key questions firms need to ask themselves around their financial crime function and spend include:

Can we show how our financial crime spend is split by customer journey and by cost category?
How would we apportion shared global costs, such as technology platforms, to the UK?
If our existing AML spend is targeted towards higher value activity, what would our financial crime function need to look like?

Firms will also need to take the following steps to work effectively with the government on its AMLAR strategy:

Engage in early dialogue with the government. This will help firms jointly shape the evidence base for review of the suspicion threshold, including how costs are split by customer journey (such as KYC at onboarding, transaction monitoring and alert remediation) and by cost category (including people, technology and third-party costs).
Commission data capture. Firms will need to take a single, joint approach to calculate their compliance costs, such as how global AML costs (e.g. technology platforms) are apportioned, and keep calculations simple, so that data is easy to capture and the numbers are clear when shared with the government. This will support more effective engagement and help the government build the business case for change.
Define the future function. Work out what the financial crime function needs to look like as activity pivots from lower-value, higher-volume work to lower-volume, higher-value work. This needs to be done in the context of other actions within the AMLAR strategy, particularly the establishment of the National Financial Intelligence Service (NFIS).

Firms that start evaluating what the financial crime function of the future needs to look like now will be better placed to manage this shift.

Financial crime control in the digital asset era

Madhav Manek

Digital assets are no longer a niche or future-facing topic. As an already established part of the financial system, digital assets are not just creating a new risk category, they are exposing where existing control models no longer work. For banks, neobanks, and payment providers, the question is whether existing control frameworks can still work effectively in environments where money can move in seconds and may be difficult to recover once it leaves the firm. Where once firms could engage in blanket de-risking, they will now have to move towards more targeted, risk-based financial crime control.

As they do so, they will be forced to reckon with a plethora of new challenges. One of the most prominent is the ‘interdiction’ gap. Where traditional payments embedded natural delay through clearing, settlement, and reconciliation, distributed ledger-based settlement collapses those stages into a single step, creating a structural mismatch between the speed of value transfer and firms’ ability to detect and intervene in time. Criminal networks exploit this through peel chains, rapid address rotation, and cross-chain swaps. On occasions where fraudulent activity with digital assets has occurred, banks can no longer rely on traditional methods of loss mitigation, such as chargebacks, and correspondent bank interventions. In authorised push payment (APP) fraud, for example, victim funds can move from a bank account to an exchange on-ramp and be dispersed across wallets or converted into stablecoins before manual review begins. Supervisory expectations are therefore shifting from recovery towards demonstrable prevention and pre-execution risk assessment.

Control burdens also vary, falling into four clusters: reserve and governance integrity for issuers; onboarding, wallet, and counterparty controls for exchanges and custodians; transaction interception and sanctions exposure for payment firms; and third-party, reputational, and selective participation risk for banks. While digital asset activity is inherently cross-border, it’s not the case for regulation. Regimes diverge in timing, scope, and supervisory intensity, creating duplicated onboarding, inconsistent wallet screening and differing Travel Rule expectations when one customer journey spans several jurisdictions.

Firms should therefore take the following steps to develop more targeted and effective digital asset financial crime controls:

Rigorously test governance, monitoring, and escalation frameworks. Test whether your frameworks can support action before or during execution, not only after the fact. Effective prevention increasingly depends on four capabilities: screening counterparties and wallets at initiation, dynamically scoring transactions using behavioural and blockchain indicators, introducing targeted friction where risk escalates, and applying interdiction rules before value is released.
Close the control gaps supervisors are already flagging. European and UK reviews consistently highlight weaknesses in the ability to act on emerging risks, with more than half of EU competent authorities identifying material shortcomings in monitoring and customer verification with enforcement following.
Map controls to entity type. Compliance does not require every firm to have the same digital asset controls. It requires firms to map obligations and control priorities by entity type rather than adopt a single enterprise-wide model. Controls should follow risk exposure, not firm type, so neobanks, incumbents, and hybrid firms should each define the gateway risks they create and align onboarding, monitoring, governance, and response accordingly. Neobanks, for example, should prioritise detecting industrialised mule account opening ('mule-as-a-service').
Design controls to tackle cross-border fragmentation. Set a single enterprise baseline for wallet screening and Travel Rule data, with documented exceptions by jurisdiction, and give unhosted-wallet exposure a named owner. Senior leaders should be able to show how these exceptions are governed within one coherent enterprise framework.
Build an intelligent risk management model across three layers. The more durable alternative to blanket de-risking is targeted participation based on business model, control maturity, and risk appetite. This must be built across upstream risk gating before payment initiation, real-time interdiction as transaction risk evolves, and disciplined escalation and recovery once suspicious activity is detected.

Success in this space will be determined not by whether firms adopt the same controls, but by whether each can demonstrate effective, preventive controls aligned to its role in the digital assets ecosystem. In an environment defined by speed and irreversibility, a control framework is judged by what it can stop before value moves, not by its ability to investigate after the fact.

When AI becomes the customer: Managing financial crime risk in an agentic economy

Adam Pearce

Agentic AI is the latest step in AI’s evolution in financial services, with many firms already shifting from early pilots and iteration to enterprise-wide deployment. This new capability is bringing real opportunity, supporting firms to better meet financial crime compliance directives through improved understanding and interpretation of new directives, automating assembly of evidence, shifting onboarding from periodic checks to real-time due diligence, and reducing false positives. In scam cases, agents can validate behavioural and transaction data in real time against additional information emerging from the customer conversation, surfacing relevant points to help the human operator guide the discussion as the context evolves.

However, this same technology also lowers the cost of crime. Criminals can run many conversations at once, in multiple languages and vernacular, generating dozens of tailored lures at the press of a button, with deepfakes adding further sophistication. As agents increasingly act on customers’ behalf, a possible next step is initiating payments and, in time, conversing with other agents in agent-to-agent flows. This would raise the risk that an authorised agent acts outside the authority it has been given by the customer, alongside new questions about how banks verify who, or what, they are dealing with. If high-velocity agent payments become routine, transaction monitoring, traceability, and accountability could become harder to manage.

For now, the account holder, whether an individual or a business, remains the customer for KYC and due diligence purposes, and an agent acts only as their delegate. Although AI is not yet the customer, it may increasingly become the one acting for them. A bank-provided agent can be verified, trusted, and tracked; a third-party registered agent may be permitted with additional guardrails and friction; and unknown or unverified agents, such as home-built or unregistered ones, may not be permitted at all.

To minimise the risk from wayward agents, customers may have to enrol specific agents as a delegated authority under their accountability, with the customer governing the controls and limits. With visibility and traceability agreed by all parties, banks can then validate agents, or even offer or partner with them directly. Further down the line, the providers of agents may need to certify them to a clear standard of control and oversight, and agree to cover liabilities where an agent strays from its agreed bounds. Where customers enlist agents outside of an agreement with their bank, banks will need to monitor and likely restrict access, in much the same way that they restrict some payments to certain crypto exchanges.

Meanwhile, a dedicated, agentic-only channel with delegated subordinate user access, similar to business users or power-of-attorney arrangements, could help banks separate legitimate agent-initiated payments from criminal automation. Monitoring can then be built around that specific use case, while existing channels can identify and refuse agent access if necessary. This will not necessarily make banking cheaper; the required controls may cost more, and agentic channels could end up as a premium extra.

These developments change the risk picture in three ways. First, accountability cannot currently be handed off to an agent. As AI gives human staff greater support, their role will increasingly shift towards mediating, arbitrating, and controlling automated processes. Firms therefore need governance that puts a human in the loop at critical junctures. As the volume and pace of agentic review grows, that review must be genuine rather than a nod-through of what the AI has generated.

Secondly, criminal agents can harvest information at scale. Seemingly unconnected conversations across different channels could all feed back to one point, concentrating risk on people susceptible to a particular type of scam.

Thirdly, because the pace of agent-driven payments leaves little room for review after the fact, firms must be able to interrogate large, disparate sources of data repeatedly, and with potentially new context. For instance, if a customer reveals new information in the middle of a phone call, such as a new beneficiary or investment approach, that information needs to be immediately checked. Doing so after the conversation makes for a poor customer experience. Firms therefore need to ask themselves whether they could show, for any automated decision, what it was based on, and where an agent hands over to a human when it can’t answer.

Firms should therefore take the following steps to ensure agents are deployed safely and responsibly:

Put traceability at the core of your AI management strategy. Ensure there is a record of every decision made and the reason behind it, which also helps show that customers have not been treated unfairly.
Keep agents in tight ‘swim lanes’, separated by function. Employ different classes of agents for different purposes, whether data collection, interpretation, or decision-making, with a supervisor where work crosses functions so that no agent can wander beyond its remit.
Design the exception route. An agent’s inability to provide an answer should be treated as an acceptable response, with the issue escalated to a human who can make a decision. Outliers should be expected, not designed out.
Determine your position on customer-side agents now. Set out whether and how customers can enrol agents as delegated authorities, what limits apply to bank-provided, third-party and unverified agents, and whether a dedicated agentic channel is needed.
Maintain rigorous controls and testing. Existing governance and testing frameworks still apply, however new the technology. Firms should be prepared to test agents in different mock scenarios, such as customer complaints and ombudsman escalations.

Over the next 12 to 18 months, firms should expect pressure for agent providers to be certified and to carry liability, for banks to formalise delegated-authority enrolment, and for agent-to-agent payment conversations to move from pilot to practice.

People fight financial crime, not technology

Alexander McGill and Sonny Singh

Financial crime has grown over the past 10-15 years, becoming one of the biggest risks facing banks. As such, it has received increasing attention at board level, prompting many global financial institutions to rapidly scale up financial crime teams and capability. Most recently, however, some firms have started to slim down large operational centres built over the years, making them more streamlined and cost effective.

Behind all the growth in technology, data and, more recently, AI, lies a function where the hardest calls are still judgement-based. Simply put, the value of technology lies in the quality of the people designing and operating it behind the scenes. Without them, firms risk underutilising the tools they have invested in, or worse, overlooking key financial crime risks altogether.

What, therefore, might an appropriate balance between technology and people look like within a financial crime risk function? When it comes to KYC or onboarding processes, technology can help collect, validate, and structure information, but people need to understand context and recognise when a customer story does not make sense. Similarly, with transaction monitoring, technology can identify anomalies and prioritise signals, but human investigators bring curiosity, contextual judgement and an ability to connect seemingly unrelated information.

A recent Europol investigation into a cryptocurrency fraud network demonstrates why technology and human expertise need to work together. Investigators uncovered more than €700 million laundered through cryptocurrency exchanges, with Europol providing crypto-analysis intelligence alongside dedicated analysts and cryptocurrency specialists. Technology helped illuminate the movement of illicit funds and the structure of the network, while specialist human judgement helped turn those insights into intelligence that could support coordinated enforcement action across multiple countries.

The key questions firms should be asking themselves are practical ones. Where are our people spending their time: genuinely investigating, or on repetitive data gathering and administration? Where does human judgement materially change the outcome? Are we developing the skills the future function will need, such as analytical thinking, investigation, intelligence, and data/AI literacy? Are people empowered to challenge technology outputs, or are we swapping process dependency for automation dependency? And do our KPIs measure activity or effectiveness, since processing more alerts or closing cases faster is not the same as better outcomes?

To answer these questions and build a risk function that balances technology with human judgement, firms should:

Identify the human moments that matter. Pinpoint where judgement, investigation, challenge, and context materially affect decisions.
Automate around those moments, rather than through them. Target repetitive work, information gathering, and other friction around the decision.
Reinvest in capability. Use some of the capacity released by technology to strengthen analytical and AI/data skills.

As AI takes on more routine work and regulators look for outcomes and clear accountability, such as the UK’s anti-money laundering and asset recovery strategy published in September, the gap will widen between firms that have invested in their people’s judgement and those that haven’t.

Can firms prove their controls work, not just that they exist?

Whether it is an assessor testing a country’s regime, a government asking banks to show where their spend goes, or an AI agent moving money in seconds, the question firms must answer is unchanged. Can they show that their controls change outcomes, rather than just that they exist? Detecting financial crime after the event is no longer enough when value moves instantly and evidence is demanded up front.

As such, firms need to be able to demonstrate that their effort is aimed at their most material risks. Accountability must clearly remain with people, even as agents and automation take on more of the work, while judgement must be protected where it matters most. And controls have to operate before the payment is made, not only after it.

Firms that wait for assessors, regulators or incidents to test these questions will be answering them on someone else’s terms. Those that test their own controls now will be able to show, with evidence, that they work in practice.

People trust us because of our deep knowledge of the regulatory system. Our experience working with regulators, banks, insurers, building societies, and others means we’ll give you advice that works in the real world. If you’d like to discuss any of the issues below in depth with our experts, you can do so here.