Khetan Gajjar, Field CTO EMEA, Mimecast
Sitting at the centre of lending, payments, insurance, pensions and investment, society is especially vulnerable in the event of any disruption to the financial services industry. A systemic system outage or cyber incident wouldn’t stay contained for long and would reverberate widely.
It’s therefore no surprise that global regulators are giving particular attention to the risks AI introduces for the industry. In the UK earlier this year, the FCA’s Mills Review – which examined how AI will transform retail financial services – called for an evolved regulatory perimeter and oversight in response to the ways that AI can amplify fraud, cyber risk and operational vulnerabilities.
The review stopped short, however, of calling for entirely new AI-specific financial regulation. It argued instead that existing frameworks and obligations, such as the EU AI Act, remain broadly workable but will need adaptation as autonomy increases.
This interesting conclusion contrasts with a still widely held perception that AI, and especially the introduction of AI agents, creates brand new risks that have to be managed with bespoke controls and isolation. In reality, financial services organisations already govern data access, insider risks and compliance more rigorously than almost any other sector. All of these functions make up the solution to AI risks. After all, every AI agent is acting on behalf of a human as an extension of their access credentials.
So, the suggestion that the FCA needn’t rip up the financial regulation playbook in response to AI risks should be taken seriously. The sentiment holds lessons for how individual financial services firms think about and manage those risks themselves.
Tackling the Shadow AI blocker to managing AI risk
While the existing foundations of mature governance across credentials, access, data resilience and third-party risk seen in financial services firms are transferable – there’s a crucial sticking point. Existing controls are built around an alert-first, investigate-second logic that’s designed for human-speed events. Agents, of course, move at machine speed.
Adapting to that speed is vital, but the cart can’t come before the horse. A staggering 98% of organisations have unsanctioned AI tools running in their environment. Attempts to adapt security and governance controls to AI risk are a non-starter if firms don’t have full visibility over the AI tools they’re using.
And, standalone, the issue of ‘shadow AI’ poses serious compliance risks for UK firms. Under the Senior Managers and Certification Regime, accountability for a system rests with a named individual. Under SYSC 10A, a firm must be able to produce a complete record of client communications on request. Use of an unsanctioned tool to draft client correspondence undermines both at once.
GDPR reaches just as far. The data these tools interact with, and in some cases are trained on, often includes personal data. Prompts, outputs and conversation logs containing that personal data may therefore fall within the scope of a data subject access request (DSAR). If a firm cannot identify, retrieve and disclose that information appropriately, the individual may complain to the ICO, creating the potential for regulatory scrutiny. Using an unsanctioned tool to draft client correspondence can therefore undermine several compliance frameworks at once.
This is why adapting to AI threats has to start from a point of knowing what AI is already running across an organisation. In practice, this means maintaining a complete, continuously updated inventory of every AI system, agent and integration with access to organisational data, including embedded capabilities inside approved applications as well as standalone AI applications.
The case for active enforcement of AI usage policies
Most UK financial services firms will already have an AI usage policy in place, and the stipulations tend to look something like an approved tool list with anything outside it prohibited, a bar on entering client data or inside information into non-company models, mandatory human review of AI output in advice and customer-facing contexts and mandatory due diligence before deploying an AI system.
Reasonable expectations, but the true test of their usefulness is how policies are followed in reality. Total visibility of AI deployment doesn’t automatically lead to better enforcement of policy; it just sets the right foundations.
Organisations need to build on those foundations with continuous behaviour monitoring and mechanisms to intervene at the moment of risk. This means that policy drift and input into unsanctioned AI tools is acted on, in the moment. These are the practical changes that will shift financial services away from the alert-first, investigate-second model towards a remediate-first, alert-second model. And doing so is fundamentally rooted in managing human risk, rather than a reinvention of the wheel.
Active management to adapt to AI risk also extends to identity governance. Regular audits of network and data access need to encompass whether AI systems and agents have more permissions than they need. Offboarding processes should ensure agents linked to employees, teams or vendors do not keep running long after ownership has changed.
Redefining the bottom line
AI adoption in financial services will not slow down, and nor should it. Used well, AI systems improve efficiency, customer experience and help teams work through complex information faster. The risk comes when adoption and implementation move faster than governance controls.
For the financial services sector, AI governance belongs inside operational resilience rather than beside it, treated as an extension of data and human risks. The firms that put in the work to establish a complete view of AI running inside their own business, and manage risks actively, will be in a great security and compliance standing no matter where the FCA, the ICO and the EU AI Act land on AI in the coming months.