Responsible Deployment of Frontier AI in Financial Services
Workshop Summary
Overview
The workshop brought together financial services practitioners, academics, industry bodies, and technical specialists to explore how frontier AI, particularly large and small language models (LLMs and SLMs), generative AI and agentic systems, can be deployed safely in financial services.
The discussion centred on a collaborative research project between the University of Warwick, Coventry University and Keele University, in partnership with FinTech West. Running until October 2026, the project will inform technical experiments, case studies and final recommendations for policymakers and industry, with the final report scheduled for publication in October 2026.
The project is structured around three work packages:
technical methods for hallucination detection in LLM outputs;
comparative analysis of AI-related regulation across the UK, EU and US; and
business adoption frameworks grounded in principal–agent theory, AI epistemology and systemic risk thinking.
Technical Focus: Hallucination Detection and Model Controls
Early-stage work includes developing a detection pipeline for hallucinations in financial services contexts, building an industry-informed prompt library, and exploring dynamic controls that go beyond traditional static model validation.
Participants discussion highlighted:
The role of dynamic controls for interactive models
The importance of clean data and retrieval-augmented generation
When SLMs are preferable to general-purpose LLMs for tightly scoped, regulated tasks
Regulatory Landscape: UK, EU and US Approaches
The regulatory segment contrasted three trajectories.
The EU AI Act was characterised as comprehensive but complex, with a potentially high compliance burden.
The US approach remains largely post-incident and enforcement-driven, with strong penalties but limited ex ante consumer protection.
In the UK, sectoral regulators and organisations such as the AI Safety Institute and the Bletchley process provide diagnostic and advisory capabilities but currently operate without dedicated AI legislation or extensive statutory powers.
A recurring theme was the gap between regulations focused on outcomes and technical work addressing root causes of risk.
Business Adoption and Implementation Challenges
From a business and implementation perspective, participants reported that current deployment remains focused on back-office and decision-support applications, including custom SLMs, with humans retaining final decision authority, rather than fully automated consumer-facing AI.
Key challenges identified included:
Monitoring model drift in interactive systems
Understanding AI embedded within intermediary processes
Navigating immature AI liability insurance markets, particularly for SMEs
Consumer Behaviour and Trust Dynamics
Several contributors noted the contrast between institutional caution and the reality that consumers already use public models such as ChatGPT for financial advice and complaints.
This raised questions around:
Trust and explainability
Appropriate human oversight
Designing user experiences that combine conversational interfaces with structured decision flows
The discussion emphasised orchestration, context engineering and integration of non-AI tools, rather than reliance on unconstrained LLM outputs.
Why This Matters for Financial Services Leaders
Frontier AI is no longer experimental. It is moving into live financial services environments, often faster than governance frameworks can evolve.
For senior leaders, this raises urgent questions:
How do we evidence responsible deployment to boards and regulators?
Where do traditional model risk frameworks fall short for generative systems?
How do we manage hallucination risk and automation bias in interactive tools?
What level of human oversight is proportionate and defensible?
How do we innovate confidently within existing UK regulatory expectations?
The workshop reinforced that the challenge is not whether to use AI, but how to operationalise trust, assurance and accountability in practice.
Responsible AI is becoming a strategic leadership issue, not just a technical one.
Next Steps and Outputs
Outputs from the workshop, including a refined hallucination taxonomy, prompt scenarios and mapped risk-mitigation mechanisms, will feed directly into the project’s technical, legal and business work packages.
Public-facing materials, including a summary of key themes and selected artefacts from the emerging framework, will be shared as the project progresses toward publication of the final report towards the end of the project.
Interested in participating in future workshops or receiving project updates? Get in touch.