Has financial crime entered an AI arms race?

Artificial intelligence is transforming the fight against financial crime. However, it’s also giving criminals powerful new tools. As banks, regulators and technology providers race to harness AI for fraud detection, AML and risk management, cybercriminals are using the same technology to launch more convincing scams, automate attacks and evade detection.

The result is an escalating AI arms race, where success depends not just on adopting the technology, but on staying one step ahead of those using it for the opposite purpose.

How GenAI is changing financial crime  

How is Generative AI changing the sophistication and scale of financial crime? In the view of Bhavin and Pooja Shah, founder and CEO and director of product and client solutions at Sherlocq respectively, the tools that promised to transform compliance are now being weaponised by the very criminals they were designed to stop. Both sides of the fight are running the same technology, and across every major financial centre, the gap between attack and defence is widening.

They said, “We are no longer debating whether generative AI will reshape financial crime. It already has. The question confronting regulators, compliance leaders, and technology providers: from London and Frankfurt to Singapore, Dubai, and New York, is whether defences can evolve fast enough to keep pace, or whether we have entered a genuine arms race with no clear winner.”

For many decades, financial crime followed easy-to-follow patterns: fraud required effort, money laundering required human networks and sanctions evasion required specialist knowledge. “Each constraint acted as a natural brake on scale. Generative AI has removed the brakes,” said the Sherlocq pair.

Generative AI, they added, allows criminal networks to personalise attacks at scale, test them against live defences, iterate in real-time and deploy across jurisdictions simultaneously.

They added, “Large language models are crafting hyper-personalised phishing campaigns, generating synthetic identities that pass traditional KYC checks, and probing AML detection systems to identify thresholds and exploit blind spots. US fraud losses climbed to $12.5 billion in 2025, with AI-assisted attacks contributing significantly to the increase. The pattern is not uniquely American: authorised push payment fraud in the UK, invoice fraud across the EU’s single market, and trade-based money laundering in APAC corridors are all exhibiting similar AI-assisted acceleration.

GenAI also enables attackers to deploy AI agents to automate and scale social engineering attacks at volume, conduct deepfake video calls to authorise fraudulent transactions, generate forged documents indistinguishable from real ones and simulate legitimate transaction patterns to evade detection. “These capabilities mean that financial crime is no longer just more frequent, it is more adaptive, targeted, and difficult to identify,” said Bhavin and Pooja.

They added, “This is not an incremental evolution in fraud technique. It is a structural change. The cost of a convincing attack has collapsed. The cost of defending against one has not.”

Generative AI is not creating entirely new forms of financial crime. Instead, it is transforming the speed, scale and economics of existing criminal activity.

In the view of Elaine Mullan, head of marketing & business development at Corlytics, financial crime has always evolved alongside technology.

She remarked, “Innovation that improves speed, convenience, or connectivity has a flip side, it creates new vulnerabilities. Online banking accelerated cyber fraud. Social media has led to single platforms being used in large-scale impersonation scams. Cryptocurrency frequently makes headlines for the wrong reasons, for opening new pathways for illegal activity.”

However, Mullan believes Generative AI is different – not because it introduces entirely new forms of criminality, but because it fundamentally changes the economics of financial crime in her view.

“It makes sophisticated fraud cheaper to execute, faster to scale and significantly harder to detect, and it is for these reasons that the financial sector has entered into what increasingly looks like an AI arms race,” said Mullan.

The same technologies helping banks improve customer engagement, automate operations, and increase efficiency are also enabling criminals to operate in more nefarious ways, she added.

“Deepfake audio is frighteningly convincing,” explained Mullan. “AI-generated phishing attacks have lost the grammar mistakes and awkward phrasing that once exposed them. There is now fake ‘proof’ to support synthetic identities including realistic documentation, credible digital histories and even convincing behavioural patterns. In short, generative AI lowers the barrier to take part in financial crime. Sophisticated fraud no longer requires sophisticated technical expertise. The cyber specialists are no longer a must-have as there is now a wide range of inexpensive or free AI tools. That changes the threat landscape dramatically, because financial crime is no longer simply becoming more digital. It is becoming more adaptive. More personalised. And increasingly, more human.”

This, she states, presents a new and difficult challenge for traditional fraud and AML frameworks. Controls that used to be designed around static rules, historical typologies and identifiable anomalies were built to detect behaviour that looked suspicious. So, these controls, Mullan outlines, are no longer enough.

“AI-enabled financial crime increasingly mimics legitimate behaviour rather than behaviour that looks suspicious. That distinction matters,” she said. “The next generation of fraud is unlikely to appear obviously fraudulent. It will appear credible. Authorised push payment fraud is growing as criminals target victims use AI-assisted impersonation ploys to create urgency and trust very convincingly. Voice cloning technology and AI-generated documents are getting more and more difficult to identify as fake. For financial crime teams, the job is now less about spotting obvious red flags and more about understanding when behaviour does not feel right. That requires a different approach.”

Whilst rule-based systems still help, they are not enough on their own to catch smart AI-driven crime. Compliance teams already deal with too many alerts, messy data and many cases to investigate, explained Mullan, and GenAI could make all of this even harder with more data, more alerts. The answer, however, is not simply deploying more AI. It is deploying smarter AI, in her view, an important distinction.

Mullan stressed that the next phase of financial crime prevention will likely centre on behavioural intelligence. “So, financial crime teams will move beyond checks like account balances and ID documents. They now look at how people use their devices, such as how they type, swipe or move to spot signs that something is wrong. Instead of judging one action on its own, these systems look for patterns over time. If the behaviour does not match the real user, it may be a sign of fraud,” she remarked.

Behavioural analytics will become critical because intent is becoming harder to distinguish from normal activity, Mullan continued. “Biometric technologies will also play a growing role. Facial recognition, behavioural biometrics and voice authentication are evolving in response to deepfakes and identity manipulation. But there is a dilemma, a challenge, as these technologies introduce new governance questions around false positives and privacy. This tension between preventing financial crime and protecting privacy may become one of the biggest regulatory challenges of the next decade.”

Financial institutions require advanced AI systems to combat increasingly sophisticated AI-enabled crime. But regulators, Mullan makes clear, also expect those systems to be transparent, auditable and defensible, even as supervisory priorities differ across jurisdictions.

“Accuracy alone is no longer enough. Trust matters. Explainability matters. Governance matters,” she said. “The organisations most likely to succeed in this environment will not necessarily be those with the largest compliance budgets but the institutions capable of adapting fastest. Those willing to modernise legacy infrastructure. Invest in behavioural intelligence. Integrate fragmented data environments. And rethink financial crime prevention as a dynamic capability rather than a static control framework. Because the central issue is not simply technological. It is strategic resilience.”

It is clear that Generative AI is reshaping the scale, sophistication and accessibility of financial crime at extraordinary speed. But as Mullan outlines, perhaps the most important shift is that the future of fraud may look more trustworthy than genuine human interaction.

She explained, “For decades, fraud prevention relied upon a relatively simple assumption, that suspicious activity eventually reveals itself through inconsistency or obvious deception. Generative AI challenges that assumption entirely. AI-generated communication can be calmer, clearer, more persuasive and even more contextually accurate than genuine human communication. Deepfake voices do not panic under pressure. AI-generated phishing emails no longer contain spelling or grammar mistakes and can sound like a native speaker or mimic a specific person’s tone and style. Synthetic identities can maintain consistent behavioural patterns over time. The next generation of financial crime may not fail because it looks suspicious. It may succeed because it appears more credible than reality itself.”

This in Mullan’s view creates a profound challenge for the financial sector, as most AML and fraud systems were designed to detect abnormal behaviour. AI criminals are increasingly learning how to behave normally, and this from the perspective of Mullan changes the nature of detection.

She said, “The real disruption is not simply that AI enables new forms of fraud. It is that it enables criminal behaviour to blend seamlessly into legitimate financial activity. This is particularly visible in synthetic identity fraud, where AI-generated personas can build realistic digital footprints, transaction histories, and behavioural consistency over extended periods. The future of financial crime is unlikely to be obvious. Instead, it will hide within behaviour that looks entirely genuine.”

The implications from the perspective of Mullan go even further. “Deepfakes do not just weaken fraud controls. They also undermine trust in what institutions treat as evidence. Historically, voice recordings, video calls, documents and biometric verification carried implicit authority. They represented proof. That assumption is beginning to weaken – and fast! Audio can be cloned convincingly. Video can be manipulated seamlessly. Identities can be synthetically generated at scale. So financial crime is not just attacking systems, it is attacking authenticity. ”

From here, perhaps the most uncomfortable question for the industry is whether compliance teams are still preparing for the wrong threats, she said.

Mullan detailed, “For years, people mostly talked about deepfakes in terms of fake politicians, altered celebrity videos and misinformation. But criminals were already using the technology in simpler, more profitable ways. They used fake executives on calls. They created fake documents for account opening and onboarding. They also used AI to trick people into trusting them. In the long run, this kind of trust-based deception could do more damage to financial institutions than the more visible kinds of AI fakery.”

Whilst the technology changes, the underlying mechanisms remain remarkably consistent, she said.

Mullan concluded, “Manipulation, pressure, credibility, urgency and trust. Generative AI simply allows those psychological tactics to operate at unprecedented speed and scale. In many ways, financial crime is becoming emotionally intelligent. For years, cybercrime focused primarily on hacking systems. GenAI allows criminals to hack human psychology instead. And that may be the defining financial crime challenge of the next decade.”

According to Janet Bastiman, Chief Data Scientist at Napier AI, tasks that once demanded specialist knowledge and significant effort can now be carried out far more efficiently. “Techniques that previously required time, expertise, and coordination—such as creating synthetic identities or producing convincing documentation—can now be executed rapidly, at scale, and with a high degree of realism,” she says. Unlike regulated financial institutions, criminals also face none of the governance or supervisory checks that constrain legitimate AI use.

The result, Bastiman argues, is an “amplification effect”. Financial crime is becoming less reliant on isolated, labour-intensive attacks and more akin to an industrialised operation, where criminals can rapidly test, refine and repeat successful tactics. For financial institutions, this exposes a fundamental weakness: many traditional controls were designed for static threats, not adaptive adversaries capable of evolving at machine speed.

Scott Nice, CRO at Label, believes generative AI is reshaping financial crime by making deception both cheaper and more scalable. “Generative AI is changing financial crime because it lowers the cost and increases the scale of deception,” he says. Criminals can now create convincing phishing messages, synthetic identities, forged documents and even deepfake audio and video with far less effort than before.

The real shift, however, is not just the quality of the content but its ability to be personalised. “The sophistication is not only in the content itself. It is in the ability to personalise attacks,” Nice explains. By using publicly available information, fraudsters can craft messages that appear “relevant, timely and credible” to their intended victims, making traditional warning signs increasingly unreliable.

For financial institutions, the challenge stretches across the entire customer lifecycle. “AI-enabled threats can appear at onboarding, during authentication, in payment instructions, in account takeover attempts and in mule account activity,” Nice says. As the volume and quality of attacks continue to rise, he argues that firms can no longer rely on controls designed for a more manual threat landscape. Instead, “the response has to be stronger identity verification, better behavioural analytics and faster linkage between fraud, KYC and AML intelligence.”

While many see generative AI as the primary driver of increasingly sophisticated financial crime, Taavi Tamkivi, CEO of Salv, argues the reality is more nuanced. The biggest frauds today, he says, are still overwhelmingly carried out by people rather than autonomous AI.

“The highest-value crimes right now are still predominantly human-driven,” Tamkivi explains, pointing to a recent case in Estonia in which a nonprofit organisation lost €700,000 to fraudsters. “The people who did this spoke fluent Estonian. They were not agents. They were human beings who had done detailed preparation work and then executed.” For the most damaging frauds—those involving six-figure sums—he believes “humans supported by agents” remains a far more accurate description than AI replacing human criminals.

Where AI is making a difference, Tamkivi says, is in the groundwork. Agentic AI is well suited to identifying attractive targets, researching organisations and pinpointing the right individuals to approach. “Think of it like top-of-funnel sales,” he says. AI can narrow down the list of potential victims, leaving “human beings… to do the hard part at the end.”

In his view, criminals have been slow to adopt more advanced AI capabilities not because they lack access to them, but because they simply do not need them. “The protection layers are still weak enough that human-driven methods work,” he says. While deepfake video calls and fully synthetic identities are already being tested, the most profitable attacks still rely on established techniques. “Once the defences get stronger, that will change. For now, the most accurate picture is humans supported by agents, not agents replacing humans.”

Andrew Davies, global head of FCC at ComplyAdvantage, remarked, “The financial crime landscape is undergoing a massive shift as GenAI fundamentally alters the speed, scale and sophistication of the threats we face. By lowering the barrier to generate and deploy new attack vectors, GenAI empowers bad actors to industrialize highly effective financial crime attacks.”

Davies also stressed that GenAI is generating a myriad of attack types, from realistic deepfakes and synthetic identities to automated, hyper-targeted phishing campaigns at a volume that legacy systems were never built to handle.

Can existing controls keep up?

Can existing fraud and AML controls still detect AI-enabled threats effectively? For Sherlocq, the honest answers is simple – not in their current form.

Bhavin and Pooja remarked, “Most AML and fraud detection systems were architected for a world of rules-based screening, predictable patterns, relatively static threats, and retrospective transaction monitoring. They look for known patterns: structuring, round-dollar transfers, high-risk jurisdictions and flag anomalies against historical baselines. These systems were never designed to detect an adversary that learns, adapts, and optimises in real time.”

This does not mean, however, that existing controls and worthless. Sanctions screening, transaction monitoring, and CDD processes remain essential foundations. But they are necessary and no longer sufficient, said the Shah’s.

Furthermore, the Sherlocq team detail that financial institutions are not standing still, as today, 90% of financial institutions use AI in fraud detection, and many have integrated it within the past two years.

They said, “The shift represents a move from reactive to proactive risk management. AI-powered systems can analyse vast datasets to uncover hidden patterns, detect anomalies in real time, and continuously learn from new fraud tactics. Unlike rule-based systems, AI models can identify subtle behavioural deviations that may indicate fraud or money laundering, allowing institutions to anticipate threats rather than simply respond to them.”

Nice believes many existing fraud and AML controls still have value, but on their own they are no longer sufficient to deal with AI-enabled threats. “Some existing controls will still work, but many will need to be upgraded,” he says. Rules-based monitoring, manual reviews and document verification can continue to identify certain forms of financial crime, but they struggle against attacks that are “faster, more personalised and more adaptive.”

According to Nice, the biggest weakness lies in static control frameworks that assess events in isolation. “AI-enabled financial crime exposes the weakness of static controls,” he explains. A synthetic identity, for example, may appear legitimate during onboarding and only reveal itself when viewed alongside other signals such as device behaviour, payment activity, network connections or relationships with other customers.

Rather than replacing existing control frameworks entirely, Nice argues that firms should focus on strengthening the intelligence that sits around them. “Existing AML and fraud typologies remain relevant,” he says, “but detection needs to become more contextual and more dynamic.” That means moving beyond individual alerts and incorporating real-time data, behavioural analysis and broader contextual information into decision-making.

Ultimately, he believes the strongest defence combines technology with human expertise. “The most effective firms will combine traditional controls with real-time data, anomaly detection, network analytics and human investigation,” Nice says. “Technology should enhance judgement, not replace it.”

Meanwhile, for Davies, he believes the answer, similar to Bhavin and Pooja, is no. “Legacy, static scenario-based, and highly siloed systems simply cannot keep pace with automated adversaries, leaving loss prevention and compliance teams drowning in false positives and blind to the full picture of risk.”

In their current form, many cannot—because they are rooted in assumptions that no longer hold. Traditional controls are built on historical patterns, predefined rules, and relatively stable behaviours. Generative AI, by contrast, enables continuous variation, meaning that patterns are deliberately designed to evade detection. Modern detection techniques – including dynamic segmentation of customers to ensure accurate risk scoring – are essential to evolving defences at the pace of financial crime.

And the issue is not simply that models need to be more advanced, but that the underlying architecture of many AML systems is no longer fit for purpose. Layering AI for post-alert triage on top of fragmented data and rigid workflows does not solve the underlying risk exposure, it often obscures it.

To respond effectively, institutions need systems that have AI embedded from the ground-up and architected to meet the demands of AI at the scale of payments. AI is only as effective as the data environment it sits on. Where legacy AML systems rely on fragmented, siloed data and batch processing, AI-ready AML requires clean, connected, real-time data pipelines. Without this purpose-built compliance-first AI approach, it can amplify inefficiencies.

AI must adapt in near real-time, incorporate a broader range of signals, and provide clear, explainable outputs. Without this, firms risk both missing emerging threats and creating additional complexity and regulatory exposure.

Bastiman believes many existing fraud and AML controls are struggling to keep pace with AI-enabled financial crime because they are built on assumptions that no longer reflect today’s threat landscape. “In their current form, many cannot,” she says. Traditional controls rely on historical patterns, predefined rules and relatively stable customer behaviour, whereas generative AI enables criminals to constantly adapt their tactics to avoid detection.

According to Bastiman, the problem extends beyond simply deploying more advanced models. “The underlying architecture of many AML systems is no longer fit for purpose,” she explains. Adding AI to existing workflows—such as using it to prioritise alerts after they have already been generated—does little to address the underlying weaknesses. In many cases, she argues, it simply “obscures” the real risk created by fragmented data and rigid legacy systems.

Instead, Bastiman believes firms need to rethink the foundations of their compliance technology. AI should be embedded into systems from the ground up, supported by clean, connected, real-time data rather than siloed datasets and batch processing. “AI is only as effective as the data environment it sits on,” she says, warning that without a purpose-built, compliance-first approach, organisations risk amplifying existing inefficiencies instead of eliminating them.

Looking ahead, Bastiman argues that effective detection will depend on AI systems that can adapt in near real time, draw on a much broader range of signals and provide clear, explainable outcomes for investigators and regulators alike. Without those capabilities, she warns, firms risk not only missing emerging threats but also creating additional operational complexity and regulatory exposure.

Technologies defining the next generation

A key question for many when debating this topic has been what technologies will define the next generation of financial crime prevention?

For the Shah pair, if the threat has fundamentally shifted, the defensive toolkit must shift with it. Several approaches will likely define financial crime prevention over the next three to five years, in their view.

Firstly, real-time adaptive AI detection. They detailed, “The most critical shift is from retrospective monitoring to systems that learn continuously from live data. Singapore’s MAS launched a proof-of-value initiative in May 2026 with banks, GovTech, and the Singapore Police Force to explore AI/ML techniques for pre-emptive scam detection, a powerful signal that regulators are endorsing this direction.

“The UK’s FCA has similarly flagged AI-assisted transaction monitoring as a supervisory priority for 2026, and VARA in the UAE has embedded technology governance requirements into its rulebook for virtual asset service providers.”

Following this is AI-powered evidence claims. AI, they state, can build verifiable evidence chains linking financial documents to actions, creating auditable trails that substantiate legitimate transactions and expose fraudulent ones. “The World Economic Forum spotlighted this approach in early 2026 as a way to build trust in digital financial systems without overreach, a balance that regulators in both common law and civil law jurisdictions are actively seeking,” Bhavin and Pooja remarked.

Following that is agentic AI for investigations. “Blockchain analytics firms are deploying AI agents capable of executing full compliance workflows: tracing funds across chains, identifying connected wallets, flagging behavioural patterns, with analysts shifting from performing the work to supervising it. This shift has material implications for how compliance teams are structured and resourced,” they added.

The final area is cross-border intelligence networks. As the Shah pair stated, criminal networks operate across institutions and borders. “Defensive systems, for the most part, do not.”

They added however, “Progress is being made: FATF has called for enhanced public-private information sharing, and the Egmont Group’s network of financial intelligence units is exploring AI-assisted typology sharing. But regulatory friction (data sovereignty rules, differing AML standards, and inconsistent legal gateways) continues to limit the speed at which cross-border intelligence flows. Closing this gap is arguably the single greatest structural challenge in the fight against AI-enabled financial crime.”

For Sherlocq, three areas demand instant attention from compliance leaders across every jurisdiction.

“First, control architecture,” said Bhavin and Pooja. “If your AML and fraud controls rely primarily on static rules and batch processing, they are already behind the threat curve. The question is not whether to modernise; it is how quickly you can do so without creating new operational or model risk in the transition.”

They went on, “Second, regulatory readiness. Regulators across the Gulf, Europe, and Asia-Pacific are moving toward prescriptive, technology-aware requirements. The EU’s AI Act introduces risk classification obligations that touch compliance AI directly. DORA imposes resilience standards on financial sector technology.

“Gulf regulators, including the CBUAE and VARA, are embedding AI governance into supervisory frameworks with increasing specificity. Gap analyses consistently reveal the same pattern: organisations have the right principles but lack the operational specificity regulators now demand. Tracking this regulatory evolution across jurisdictions in real time, not quarterly, is becoming a baseline requirement for compliance functions, not a premium capability.”

Third, organisational capability. “Investigators need to understand adversarial AI. Compliance officers need to evaluate model risk. Boards need reporting that captures the velocity of threat evolution, not just the volume of SARs filed. Seventy-five percent of anti-financial crime professionals plan to increase their use of AI for detection. The gap between recognising the threat and fielding the response is where real risk lives.”

Tamkivi believes the next major breakthrough in financial crime prevention will not come from replacing legacy systems, but from making them significantly smarter. Rather than embarking on costly, multi-year transformation programmes, he sees the biggest near-term opportunity in deploying agentic AI on top of existing infrastructure.

“One thing I find genuinely exciting is applying agentic intelligence on top of existing legacy infrastructure, rather than trying to replace it,” Tamkivi says. Most large financial institutions, he notes, rely on monitoring platforms that are deeply embedded into their operations and too expensive – and too risky – to replace overnight. “You are not going to tear that out and replace it in under three years. The costs and risks are too high.”

Instead, agentic AI can address one of the biggest bottlenecks in financial crime investigations: the investigator’s workflow. Analysts often spend much of their time switching between multiple systems, manually gathering information and trying to piece together relationships across fragmented datasets. “The data is not missing,” Tamkivi explains. “It is that the analyst has ten browser windows open, is copying and pasting between systems, trying to find chain patterns in real time, and the cognitive load is enormous.”

By acting as an intelligent layer across existing platforms, AI agents can surface relevant information, connect disparate data points and accelerate decision-making without disrupting the underlying technology stack. Tamkivi believes this pragmatic approach offers a much faster route to improved detection capabilities than wholesale system replacement. “Agents layered on top of existing infrastructure is something that could deliver real results very soon,” he says. “The full system replacement can come later, and probably will.”

Nice argues that the next generation of financial crime prevention will be defined less by any single breakthrough technology and more by the ability to connect and operationalise a broad set of capabilities. These include identity verification tools, biometric and liveness checks, device intelligence, behavioural analytics, network analysis, graph technology, machine learning and explainable AI.

However, he stresses that the real differentiator will not be the tools themselves, but how effectively they are integrated into day-to-day decision-making. “A powerful model is of limited value if investigators cannot understand the output, if data quality is poor or if case management remains fragmented,” he explains. In other words, technology alone cannot compensate for weak operating models or disconnected systems.

Nice also highlights the growing importance of reusable customer intelligence. The same underlying data used for onboarding can increasingly support AML monitoring, fraud detection, sanctions screening and tax compliance obligations such as FATCA and CRS. The goal, he says, is not unchecked data expansion, but a “governed, permissioned and auditable view of the customer” that can be safely reused across compliance functions.

Ultimately, he believes the most successful institutions will be those that balance automation with explainability and innovation with governance. Financial crime prevention is becoming more data-driven, but trust in these systems will continue to depend on evidence, accountability and human oversight.

Bastiman argues that the next generation of financial crime prevention will not be defined by any single technology, but by how effectively different capabilities are brought together within a coherent, intelligence-led framework.

She highlights a shift toward real-time, context-aware analytics that move beyond static thresholds to detect subtle behavioural changes, alongside more holistic, entity-centric models that connect data across customers, accounts and networks to build richer contextual understanding. In parallel, she points to the rise of continuous monitoring architectures that replace periodic reviews with persistent, always-on risk assessment.

Explainability, she stresses, will also become a core requirement rather than an optional enhancement. “Explainable AI as a foundational requirement – not an enhancement – ensuring that every decision can be understood, tested, and defended,” she notes, reflecting the growing regulatory and operational demand for transparency in automated decision-making.

Underpinning all of this is the need for modern, flexible data architectures. These must allow institutions to evolve detection strategies without repeatedly rebuilding core systems, enabling faster adaptation as threats evolve.

For Bastiman, the critical distinction is that success will depend less on the sophistication of individual models and more on whether organisations have built the right foundations to support them. Ultimately, she argues, the fight against AI-enabled financial crime requires a shift from simply deploying more technology to deploying it correctly – with transparency, adaptability and strong architectural design at its core.

Davies states on this matter that the next  generation of financial crime prevention will be defined by AI-native, unified platforms that leverage AI-acquired and curated data, predictive and agentic AI to connect the dots in real time.

He detailed, “Yet, implementing these technologies is where the real pressure lies for many firms. Compliance leaders find themselves caught in a daunting squeeze: they are told from all sides they must adopt AI while simultaneously facing internal freezes on headcount spend.

“Navigating this transition while managing daily operations is incredibly challenging, but building a non-siloed, AI-native infrastructure is the only way firms can see the full picture of risk, automate routine tasks, and tip the scales back in their favor.”

Can compliance teams keep pace?

For RelyComply, the debate is no longer about whether AI will transform financial crime – it already has. The real question, the company argues, is whether compliance teams can adapt quickly enough to keep pace. “The question is whether compliance functions are willing to close the gap before the cost of doing nothing becomes irreversible.”

According to RelyComply, criminal networks are already using generative AI to industrialise their operations. Stolen identities obtained through phishing campaigns, data breaches and dark web marketplaces are being combined with AI-generated videos, cloned voices and fabricated transaction patterns to evade authentication controls and move illicit funds. These techniques, it argues, are no longer isolated experiments but “operational playbooks, replicable and scalable.”

The company believes one of the biggest weaknesses lies within financial institutions themselves. Organisations that continue to treat AML, fraud prevention and prudential risk as separate disciplines are creating blind spots that criminals can exploit. “Siloed systems produce siloed intelligence,” RelyComply says, “and siloed intelligence misses the cross-typology signals that define AI-enabled financial crime today.” Instead, it advocates a unified data architecture that connects compliance teams to real-time watchlists, sanctions data, adverse media and politically exposed person (PEP) databases through a single source of truth.

That foundation, RelyComply argues, enables institutions to deploy the technologies needed to match increasingly automated threats. These include liveness detection, behavioural biometrics and facial recognition to identify synthetic identities during onboarding, alongside agentic AI capable of gathering official records, tracking emerging typologies and supporting perpetual KYC throughout the customer lifecycle.

Crucially, the company stresses that AI should augment, not replace, experienced investigators. Human analysts remain essential for validating alerts, refining models and exercising judgement. What AI does replace, RelyComply argues, is the belief that manual compliance operations remain an effective baseline. “The criminals automated first. Institutions that haven’t responded in kind are operating without a defence.”

Lawrence Hamilton, CCO at Consilient, also remarked, “Our view is that AI is a critical part of the arms race but unlike criminal networks organizations are still reliant on silo’d approaches and guidance that causes detection and operational challenges.”

The road ahead

So, what is the road ahead? In the view of Bhavin and Pooja, the evidence suggests that financial crime has indeed entered an AI arms race.

They said, “Criminals operate without regulatory constraints, allowing them to adopt new technologies faster. Financial institutions must balance innovation with compliance, ethics, and customer trust. This asymmetry means the race is not just about technology; it is about speed, governance, and collaboration.”

Human oversight remains non-negotiable in this environment, they added.

“The EU AI Act, MAS guidance, and emerging VARA frameworks all grapple with the same question: where does the human sit in an AI-assisted compliance process? The answer matters not just for regulatory compliance but for the integrity of the decisions being made. AI systems that flag, prioritise, and recommend are valuable. AI systems that decide, without meaningful human review, introduce a category of risk that no regulator has yet sanctioned, and that no institution should accept,” the Sherlocq pair said.

They concluded, “Looking ahead, success in financial crime prevention will depend on continuous innovation in AI capabilities, cross-industry data sharing and intelligence collaboration, strong regulatory frameworks for AI governance that are consistent across jurisdictions, and the preservation of meaningful human oversight at every critical decision point. The arms race is real. The only losing move is to stand still.”