From Pilot to Practice: Making AI work for financial services, responsibly

UK financial institutions are moving from exploratory AI pilots to embedding models in live, regulated workflows. The pace is encouraging, but it raises an obvious question: Are organisations moving too quickly to ensure that AI adoption is safe, responsible and compliant? The conversation needs to shift from adoption to governance. 

The Financial Conduct Authority (FCA) has made its position clear. Its ongoing work on AI in financial services reflects a genuine desire to support innovation, but within a framework that protects consumers, promotes fairness and preserves market stability. The regulator’s message is one of enablement through accountability, and firms that treat governance as an afterthought should expect that to be reflected in supervisory engagement and, ultimately, enforcement action.

How AI is already reshaping financial services

AI is no longer knocking at the door of financial services, it’s firmly embedded within it. Across banking, insurance and asset management, firms are leveraging AI not as a novelty but as a fundamental operational tool. The question is no longer whether to adopt AI, but how to scale what is already in place while maintaining the standards regulators and customers expect.

LLMs are now being deployed to accelerate research workflows, automate some interactions and extract insights from vast quantities of unstructured data. Machine-learning models are being trained on proprietary datasets to sharpen fraud detection, improve credit decisioning, refine risk assessments and enhance market surveillance capabilities. Meanwhile, AI-powered automation is compressing the timelines of labour-intensive processes. Equally, the productivity case is also compelling. But acceleration creates its own risks. As organisations race to embed AI more deeply into their operations, compliance considerations can be deprioritised in favour of speed, and that’s precisely where the danger lies. The capabilities of AI are considerable, but so are the questions they raise around transparency, bias and regulatory accountability.

Explainability remains one of the most significant unresolved challenges. Generative AI models in particular can produce different outputs in response to the same prompts, undermining the consistency that regulated decision-making demands. Bias, often embedded in the underlying training data, can produce outcomes that disadvantage particular groups of customers without any deliberate intent. And when AI is layered onto legacy infrastructure without adequate care, the interaction can generate unpredictable results – compounding operational risk and creating exposure under regimes such as the Senior Managers and Certification Regime. The FCA’s April 2024 AI Update was explicit: the transformative potential of AI cannot come at the expense of consumer fairness or financial stability.

Governance is not the enemy of innovation

There is a persistent misconception in some quarters that compliance and regulation are obstacles to AI progress. The opposite is true. Regulation provides the certainty that enables firms to invest with confidence. Where the rules are clear, coherent and consistently applied, innovation flourishes, and where they are ambiguous or fragmented, firms hesitate and opportunities are lost.

Alignment between the FCA, the Prudential Regulation Authority, the Bank of England and the Information Commissioner’s Office is vital. Firms operating across the regulatory perimeter need consistent, coordinated guidance. A patchwork of competing or contradictory signals from different bodies does not serve the market. Getting that coordination right is as much a priority for regulators as it is for the firms they oversee.

Within firms, closing the explainability gap requires a fundamental shift in mindset. Transparency cannot be retrospectively applied to an AI system once it’s been deployed; it must be built in from the outset. That means selecting interpretable models where the stakes are high, maintaining comprehensive audit trails across AI-assisted decisions and subjecting systems to rigorous stress testing to identify the conditions under which outputs may disadvantage customers. Under Consumer Duty, this is not optional, firms must be able to demonstrate that their AI-driven processes are delivering genuinely fair outcomes.

The EU AI Act introduces additional considerations for organisations with cross-border exposure, establishing more prescriptive requirements for high-risk AI applications and prohibiting uses that are manipulative or discriminatory in nature. Firms that begin to addressthese standards now, rather than waiting for regulatory enforcement to force their hand, will be materially better positioned as the compliance landscape continues to evolve.

Firms must also think carefully about how their AI systems interact with the full range of customers they serve. The Equality Act 2010 prohibits indirect discrimination regardless of the mechanism that produces it, which means algorithmic decision-making falls squarely within scope. This includes a duty to identify and address instances where vulnerable customers may not be receiving the outcomes they are entitled to. It requires continuous monitoring, quality assurance and the operational capacity to identify and remediate unfair treatment in something close to real time.

Finally, firms should resist the temptation to force AI into existing infrastructure without adequate preparation. Legacy technology was not designed with modern AI in mind, and poorly managed integration creates a category of risk that is both difficult to detect and expensive to fix. A phased, expert-guided approach to integration that stress-tests compatibility at each stage before proceeding is not a counsel of excessive caution, it’ssimply good engineering, and it pays dividends in resilience and regulatory defensibility.

What’s next for AI in financial services? 

The trajectory of AI in financial services is heading in one direction – deeper integration, broader scope and greater regulatory scrutiny. The next phase of AI adoption will extend itsreach beyond operational efficiency into the very heart of compliance and conduct monitoring. Increasingly, AI will be the mechanism through which firms detect emerging risks, identify conduct concerns and fulfil their obligations to protect consumers.

Regulators will likewise sharpen their focus. Expect growing emphasis on explainability, not just as a technical standard but as an expectation embedded in supervisory conversations. Requirements to demonstrate consistent, fair and accountable outcomes will only intensify, and firms that have invested in getting their governance foundations right will find themselves significantly better placed to meet any potential scrutiny.

AI will not remain a supporting function operating alongside compliance, it will become inseparable from what compliance means in practice. Firms that recognise this now, and that are building the governance infrastructure to match, will not simply avoid regulatory censure, they will have turned compliance into a genuine competitive differentiator, and that’s where the real long-term advantages lie.