.svg)
Across financial crime and anti-money laundering (AML) efforts, artificial intelligence has demonstrably sharpened detection, reduced false positives and uncovered patterns that remain invisible to traditional, rules-based systems. Yet despite the undeniably clear and constantly increasing potential, many institutions find themselves stuck in pilot programmes, unable to move beyond controlled experimentation into full deployment.
The question isn’t whether AI can transform AML, but whether it can be trusted to do so at scale, and the real barrier is not technological capability, but confidence.
As regulators increasingly embed explainability, auditability and governance into supervisory expectations, trust is emerging as the threshold for AI adoption. It’s inevitable that institutions will need to embed AI capabilities to keep pace with innovation. It is not enough to instil these platforms blindly. The future of AML will not be shaped by how powerful AI models become but by how transparent, accountable and understandable they are in practice.
The AI trust test
For AI to scale meaningfully within AML frameworks, it must not be defined solely on performance, but by accountability, clarity and control, in other words, it must prove more than it can deliver strong outcomes, it must also prove it can be trusted.
Regulators are no longer satisfied with outputs alone and increasingly require institutions to demonstrate a clear understanding of how every decision is reached, whether using AI or rules. Flagging suspicious activity is not enough, firms must be able to clearly evidence why the activity was flagged, what data informed the decision and how the outcome aligns with regulatory expectations. This is designed to both improve financial crime compliance from regulated firms and discourage the historical over-reporting that swamps law enforcement with false positives.
The regulatory landscape is shifting to take a more outcomes-focused approach. Historically, compliance has been evaluated by the existence of processes or systems but, today, scrutiny shifting to both the accuracy of the outcomes and the decision-making processes themselves, requiring that a risk-based approach be evidenced and placing greater emphasis on transparency and governance at every stage.
This evolving approach is already reflected in regulatory initiatives such as the Financial Conduct Authority’s innovation programmes (such as Supercharged Sandbox and AI Live Testing), which enable firms to test AI models within controlled environments while prioritising safety, oversight and explainability. These programmes signal a clear direction of travel: innovation is encouraged, but only where it can be properly understood and governed.
The principle is simple. Systems that cannot be explained cannot be confidently deployed, meaning models that cannot be interrogated cannot be approved. Trust is no longer a by-product of compliance but has become the fundamental pre-requisite to it and must be at the core of regulatory processes and the platforms supporting them.
From opacity to glass box transparency
Hesitation around AI in AML stems, largely, from the perception of ‘black box’ or ‘opaque’ models, systems that generate outputs without providing sufficient visibility into their processes and underlying logic. Within regulatory environments, opacity doesn’t just breed discomfort, but distrust in any outputs it generates and, therefore, its implementation.
The move towards ‘glass box’ transparency is an approach defined by embedding explainability into the system by design rather than applying it retrospectively. In practice, meaningful transparency requires systems to provide clear, contextual reasoning at the point of decision making to enable investigators, auditors and regulators to understand not only what has been flagged, but why it seems suspicious.
Behavioural insights and contextual signals become directly integrated into workflows, enabling compliance professionals to interpret risk in real time rather than reconstruct it after the fact which not only improves efficiency, but strengthens the defensibility of decisions under scrutiny.
At their core, glass box models must be useable. Regulators can’t seek to interrogate every technical draft of a model’s architecture, but what they can do is demand a clear and coherent narrative that links data inputs, decision logic and outcomes. Without implementing transparent models, AI models in AML, even extremely sophisticated ones, will always be constrained by a lack of trust.
It is important to understand the critical nature of the set-up of the underlying AI that supports transparency such as Large Language Models (LLMs) or agents that provide explanatory narratives. These AI tools can provide inaccurate reasoning for recommendations even when alerts are correct, if not correctly designed, trained and implemented. Trust in explanations must not be blind, it must be spot-checked and validated regularly to ensure that the reasoning for AML alerts is sound, and therefore the analyst decisions are based on solid insights.
Global signals of trust
Globally, there are growing indications that trust in AI for AML is beginning to be prioritised, particularly within leading financial markets such as the United Kingdom.
In the UK, regulatory engagement is evolving beyond high-level guidance and is transforming into active collaboration, largely due to initiatives such as sandbox environments and synthetic data programmes within which investigators can test, validate and refine AI systems within realistic yet carefully controlled conditions. FCA programmes include the Synthetic Data AML Solution Sprint (SAMLS). These efforts are significantly helping to establish clear expectations around explainability, governance and model performance, reducing uncertainty and enabling more confident adoption.
Rather than acting solely as overseers, regulators are increasingly positioning themselves as facilitators of responsible innovation, signalling a broader shift not only in regulator behaviour, but philosophy. It is not just the FCA leading the charge for collaboration in innovation. The European Union’s (EU) new Anti-Money Laundering Authority (AMLA) has launched an industry-wide data collection exercise to test risk assessment models ahead of direct supervision starting in 2028. This is a clear move by the regulator to facilitate fincrime controls.
Compliance professionals are recognising that well-governed AI has the potential to strengthen, rather than undermine, financial crime oversight when built on foundations of explainability and trust.
This isn’t to say reservations don’t continue to persist. Many institutions approach AI adoption with caution, often due to a lack of clear, regulatory-ready frameworks that can support explainability and governance at scale. The resulting divides are not technological, but structural as trust is being built in environments where the priority is clarity, collaboration and accountability, leaving uncertainty and hesitation persisting in markets where these elements are less defined.
Building accountable AI
In order to bridge the gaps and unlock the full potential of AI in AML, financial institutions must move beyond viewing AI as a standalone technological solution and instead, embed it within a broader, well-defined, transparent compliance ecosystem, at the centre of which must be accountability.
It’s essential that AI systems operate within robust governance frameworks that clearly define AI roles, responsibilities and AI mechanisms. The introduction of AI, crucially, should not diminish human involvement but rather enhance it as human expertise remains and will always continue to be essential to provide context, judgement and ethical guidance that cannot be replicated by automated systems.
The human-in-the-loop approach ensures that compliance professionals always remain responsible for outcomes, supported by AI systems that improve operational efficiency and provide clarity and insight rather than complexity or ambiguity.
Auditability is equally important as every decision generated by an AI system must be traceable, with clear documentation and evidence trails of how inputs were processed, conclusions were reached and how those conclusions withstand scrutiny.
When explainability, governance and oversight are embedded from the outset, AI evolves from a mere technical capability to a trusted compliance asset to be relied upon for more than just detection, but defensible decision making.
Earning the right to scale
The path forward for AI in AML will not be defined solely by advances in technology, but by the frameworks that make those advances usable, governable and trustworthy.
Despite heavy financial investment in compliance infrastructure, broader industry analysis, as discerned in the Napier AI / AML Index 25-26, consistently shows increased spend alone has not translated into greater effectiveness. It’s evident that AI can deliver significant cost savings, estimated at around $183 billion per annum globally, but these benefits are only achievable if models are explainable, decisions can be audited, and regulators are satisfied.
AI offers a clear opportunity to shift from volume-driven processes to precision-led intelligence, but only if it can meet the heightened expectations of modern regulation.
Glass box AI therefore is not and cannot be a future aspiration for compliance, but an immediate requirement, because scaling AI is not simply about demonstrating that it works in principle, it is about proving, unequivocally and continuously, that it can be trusted in practice.
