The financial markets sector is accelerating its engagement with generative AI (GenAI), yet moving from proof of concept (POC) to production remains a complex challenge. Key questions continue to dominate industry conversations: What use cases are mature enough for deployment? How can firms embed the necessary controls to manage reputational and regulatory risk? What organisational and cultural barriers impede adoption? And how should firms measure success and optimise costs?
These were among the central themes explored during a panel discussion at the A-Team Group’s recent AI in Capital Markets Summit, featuring Nirav Shah, Senior Executive Director, Asset Management Research Technology at J.P Morgan Asset Management; Jacek Wieclawski, Head of Innovation at Rabobank London; and Malavika Solanki, Strategic Change Specialist, formerly at NatWest Markets. The session was moderated by Priyank Patwa, Director AI, Data & Analytics Lead at Deloitte.
Building Familiarity
Initially, firms have adopted GenAI through internally focused, low-risk applications to build familiarity with the technology and develop foundational competencies. It has allowed firms to better understand their models, data quality, and the challenges inherent in building small POCs. Early applications include enhanced chatbots capable of leveraging transaction and interaction histories to provide proactive client service, together with more novel initiatives – one panelist described the automated generation of research scripts combined with deepfake technology, designed to create personalised analyst videos with full analyst oversight and consent.
This incremental approach, starting with internal pilots and gradually moving to client-facing innovation, underscores the importance of managing risk and cultivating user trust.
Challenges
Transitioning from pilot to production presents two primary challenges: embedding robust controls and aligning use cases with the technology’s inherent strengths. The difficulty lies not in ‘solutionising’ with GenAI, but in implementing appropriate controls for reputational and regulatory risk.
Another critical consideration is the nature of the use case itself. There is a frequent misalignment when GenAI is applied to quantitative problems where deterministic outputs are required. Given GenAI’s probabilistic nature, one panelist suggested it is better suited to qualitative analysis, semantic search, and content generation. Successful production cases have included semantic search tools for internal knowledge management and coding co-pilot systems enhancing developer productivity.
Persistent barriers to scaling GenAI include data quality and accessibility, regulatory and model validation challenges, and cultural resistance. Despite rich datasets, inconsistencies and governance issues can impede effectiveness. Regulatory scrutiny demands thorough model validation, often extending timelines and complicating deployment. Cultural resistance, stemming from mistrust and fear surrounding AI, continues to pose significant obstacles. Many organisations are focusing on demystifying AI and fostering broader literacy to ease these transitions.
Applying MLOps Techniques
Addressing the technical complexities of scaling GenAI requires robust machine learning operations (MLOps) practices. Traditional MLOps principles around model performance monitoring, drift detection, and retraining remain vital, but GenAI’s non-deterministic outputs introduce additional layers of complexity. Organisations must focus on auditability of the entire prompt-to-response chain and ensure factual accuracy, particularly through citation mechanisms. Furthermore, data infrastructure must evolve to accommodate unstructured data for effective retrieval-augmented generation (RAG) systems.
As GenAI adoption scales, building the right teams becomes crucial. Organisations are prioritising AI-focused IT architects and security experts to manage integration within complex, legacy systems. Regulatory and risk specialists are being brought in early to navigate compliance challenges, and AI product owners or translators are playing a critical role in bridging technical and business requirements. More than technical skills, mindset and the ability to champion AI initiatives across the organisation are being recognised as decisive factors in successful adoption.
Metrics
While cloud compute costs are a major consideration, a broader set of metrics is needed. These include model training costs, user engagement, and the tangible value delivered by AI outputs. Early-stage POCs also benefit from having metrics that track scalability and approval times, ensuring a more structured path to production. One effective KPI framework focuses on calculating the cash return on capital invested, factoring in time saved, accuracy improvements, and business impact. Additionally, concentrating adoption efforts on a core 20% of engaged users often delivers a disproportionate share of benefits.
Strategically, firms are focusing internal efforts on the top layers of the GenAI stack: application integration and control frameworks, while leveraging third-party solutions for foundational models and infrastructure. This targeted investment approach ensures that resources are directed towards areas delivering differentiated value.
The discussion underscored that while GenAI offers transformative potential, realising that promise requires careful navigation of technical, organisational, and cultural challenges. Early successes are built on robust governance, a strategic focus on suitable use cases, and ongoing education to drive adoption.
In their concluding remarks, panelists agreed that with AI literacy becoming a fundamental requirement for career longevity, building a culture of informed, responsible innovation has now become essential for financial services firms aiming to stay competitive.
Subscribe to our newsletter