Artificial Intelligence Consortium minutes – June 2026

Item 1: Welcome remarks

Co-chair David Geale welcomed attendees to the AI Consortium’s (AIC) fourth quarterly meeting, held at the FCA’s offices. Co-chair Sarah Breeden noted that the rapid pace of AI development in financial services underscores the importance of initiatives such as the AIC in identifying how regulators can best support responsible AI adoption.

The contributions during the session were under Chatham House Rule.

Item 2: Workshop presentations

Each group provided an update on their near-complete workshop analysis followed by discussion. 

Workshop 1: Explainability and transparency in generative AI

The workshop leads set out their approach to explainability and transparency in generative AI (GenAI), anchored in an outcomes-based definition focussed on understanding decision-making and whether systems are behaving as intended. Members outlined how the group has used a quality assurance case study to demonstrate how both explainability and transparency can be addressed in AI systems through features such as human in the loop (HiTL), adversarial testing, and accountability and auditable documentation of decisions.

The workshop identified recurring model risk challenges arising from GenAI such as system complexity and interconnectivity, limited transparency from third-party providers, and variability and unpredictability of outputs. The members noted these features can mean it is difficult to apply existing model risk frameworks proportionately. To address this, the workshop considered the value of managing model risk from an AI model system perspective, recognising that most use cases involve multiple interconnected models and components. Members emphasised the importance of governance across the full AI model system, transparency on how components interact, and testing both the AI model system as whole and its individual elements, noting that models can be updated independently.

Members also noted that, in practice, GenAI systems are often classified as high risk under frameworks such as SS1/23 – Model risk management principles for banks (“SS1/23”), which may limit proportionality in applying controls. The workshop suggested introducing mitigating factors to support more nuanced application of existing frameworks, supported by strong governance and clearly defined boundaries. Across the discussion, members highlighted the importance of outcome-based validation, strengthened oversight of third-party providers, and clearer governance across AI systems.

It was noted by a member that inconsistent definitions of AI use cases across firms may create challenges for applying governance frameworks, and agreeing on a definition could tackle this issue. Another member questioned whether complexity alone should drive risk classification, noting that firms have experience of complex models, but that GenAI raises distinct challenges in understanding and managing the unpredictability of model outputs.

It was suggested by a member that, in the context of explainability, it would be useful to develop different playbooks or approaches for developers and deployers of AI models, since the two might face different challenges. The workshop lead noted that solutions such as model scorecards could support consistent implementation.

Workshop 2: AI-accelerated contagion

The workshop lead outlined the group’s analysis of how AI adoption may alter contagion pathways across the financial system, with potential impacts such as price volatility, changes in participant behaviour, and system‑wide disruption during periods of stress. The workshop examined how increased automation, speed, and reliance on shared technical dependencies may amplify operational risks and create common points of failure across firms. The workshop proposed a firm-level response taxonomy as a starting point for understanding these risks.

The workshop highlighted the importance of distinguishing between visible and invisible risks. For example, fast-moving contagion could unfold within minutes following a provider or infrastructure outage, whereas slower-moving contagion may arise from correlated errors that propagate gradually across firms and may only become apparent through cross-firm comparison. Members also observed effective testing of contagion could require a cross-firm approach.

The workshop identified two practical considerations (1) ensuring agent actions remain within clearly defined limits, and (2) reframing the supervisory question from whether individual firms can deploy systems safely, to whether the wider system can also observe, test, and contain their behaviour.

Members discussed how AI-accelerated contagion could be identified in practice, noting that engagement with regulators would be important to enable the cross-firm visibility required to understand how contagion spreads. It was also suggested that frameworks developed in other workshops, including the AI Edge Cases workshop, could support this approach.

Members cautioned against attempting to forecast specific outcomes, instead highlighting the value of wargaming techniques (i.e. simulation-based exercises) to explore potential failure modes. Several members noted similarities with cybersecurity approaches, agreeing this could help inform standards and definitions.

Finally, members discussed potential contagion risks arising from the consumer side, including the possibility that widespread use of AI tools could influence customer behaviour at scale. It was suggested by a member that AI agents may actively drive such behaviour; this would represent a higher risk use case requiring stronger controls, including mechanisms to intervene or halt activity through kill switches.

Workshop 3: Concentration risk

Workshop leads shared how they have considered concentration of AI providers in UK financial services and the risks this may pose to financial stability and market integrity. The workshop supported their analysis with a survey conducted within their professional network which suggested that concentration arises from underlying characteristics of AI provision, particularly at the model and compute levels, with limited alternatives. This was noted as particularly relevant where such services support Important Business Services (IBS), that is, services whose disruption could threaten a firm’s safety and soundness or financial stability.

The workshop highlighted the need for a shared understanding of firms’ relationships with third‑party AI providers. Members noted that existing work, including by the Cross‑Market Operational Resilience Group (CMORG), provides a useful foundation for more standards and consistency in this area.

The workshop also identified talent concentration as a potential systemic issue, noting strong demand for AI‑related skills. Members emphasised that required capabilities extend beyond engineering to include governance and operational oversight. The workshop discussed opportunities to accelerate AI upskilling within financial services, including through targeted skill accelerator programmes focussed on areas such as large language model (LLM) operations and governance. Members also highlighted the importance of strengthening the domestic talent pipeline, including by developing capability in regions outside London.

The workshop was asked whether concentration risks differ depending on whether models are hosted in‑house or accessed via software‑as‑a‑service (SaaS), noting that SaaS models may increase third‑party dependency and reduce substitutability. Workshop members confirmed this had been considered, with their member-led survey indicating greater reliance on SaaS, which they intend to reflect in their final analysis.

Workshop 4: Evolution of AI edge cases

The workshop leads shared that the group’s aim was to highlight the range of controls required for novel AI systems, particularly in fast‑changing environments with frontier AI models. Rather than focussing on specific high‑risk use cases, the workshop tested six themes relating to AI edge cases to assess how far traditional risk management approaches remained applicable. Across these themes, similarities were identified in how systems failed, prompting the workshop to further explore the steps required when AI system outputs are no longer consistent with expectations. The workshop proposed a four‑step approach, which they suggested could be applied across firms, namely: (1) identifying what type of failure has occurred; (2) detecting, using observable signals, how the failure shows up in practice; (3) identifying the minimum evidence needed to diagnose the failure; and (4) implementing pre-defined controls to contain the failure.

The workshop was asked whether the workshop had defined the point at which a use case moves from mainstream to an edge case. Workshop members acknowledged that this was debated early on, noting that definitions may vary depending on business or technical perspectives. The workshop members explained that a precise definition was not required, and that they focussed on cases where novelty is introduced leading to traditional risk management controls not being readily applicable. Members added that adopting a system‑level approach extends risk management beyond the foundation model to other components of the AI system, including prompts and retrieval augmentation.

The workshop members noted how developments in frontier AI cyber capability reinforce the usefulness of the framework in addressing novel risks from AI systems. They concluded by encouraging greater cross-firm visibility, including by suggesting voluntary output reporting to support governance of these systems.

Item 3: Consortium discussion on key trends

The Co-chairs invited members to discuss implications for the financial sector, in light of recent breakthroughs in frontier AI models. Members initially reflected on the cyber implications, noting that recent developments (such as Anthropic’s Mythos model) highlight the importance of cross-sector and cross-border collaboration in strengthening defence capabilities within financial services.

Turning to non-cybersecurity implications, members discussed the potential impact of these developments on trust, adoption, and market dynamics. Members noted that, alongside cybersecurity considerations, it remains important to maintain attention on other factors underpinning trust, including ethics, bias, and the design of automated systems.

Some members reflected that the pace of advances in AI capability appears to have accelerated significantly. Members emphasised the importance of reassessing previous assumptions about the speed of change, including through stress‑testing scenarios where capabilities advance more quickly than expected, such as in agentic payments.

In this context, members highlighted risks relating to governance and capability. Some members expressed concern that the pace of development could outstrip existing governance frameworks and exacerbate domestic skills gaps. A member noted that, in some cases, the focus should be less on expanding capabilities and more on implementing effective safeguards. In particular, it was suggested that greater standardisation of AI incident reporting could support cross‑firm learning and improve visibility of failures, recognising that incidents may continue to occur despite the presence of safety mechanisms.

Finally, members noted that while frontier models are demonstrating increasingly advanced capabilities, including in highly technical domains, they may also enable the rapid generation of research and analysis at scale. Members highlighted that this may increase the volume of outputs containing errors or flaws, with potential implications for decision‑making and knowledge quality.

The Co‑chairs invited members to discuss approaches to the safe deployment of agentic tools within firms, including the design of model harnesses (i.e., the software layer around an AI model that manages its execution and interaction with other systems) and execution boundaries (i.e., controls that separate LLM probabilistic reasoning from the deterministic system actions). The Co‑chairs also asked whether there was sufficient clarity on what constitutes good practice in the design of harnesses. Members noted that while definitions of “harnesses” may vary, the underlying technical capabilities to implement guardrails are increasingly well established, and firms are not starting from a blank slate when developing these controls.

Members discussed the role of HiTL as a safeguard – it was observed that while HiTL is widely relied upon, it may not be practical in all contexts. As an alternative, members suggested greater use of testing approaches such as wargaming, focussed on anticipating adverse outcomes including assessing how systems respond to novel or unexpected data inputs. One observer highlighted the UK’s use of regulatory sandboxes as an effective mechanism for enabling the testing of AI systems in controlled environments.

Wrap up & closing remarks

The Co‑chairs thanked members for their contributions over the past year, including their work and analysis across the AIC’s workshops. They noted that the outputs from the workshops will be consolidated into a final publication later in the year.

The Co‑chairs concluded that members’ feedback and contributions will also inform the second phase of the AIC.

Attendees

Co-chairs & Moderators

Breeden, Sarah – Bank of England

Geale, David – Financial Conduct Authority

Members

Ahmed, Ratul – Commerzbank AG 

Beliossi, Giovanni – Axyon AI SRL 

Bhatti, Tanveer – Independent

Brink, Suzanne – Lloyds Banking Group

Daley, Sue – techUK 

Dunmur, Alan – Allica Bank 

Heffron, Sarah – JP Morgan

Hughes, Clara – Pension Insurance Corp 

Jefferson, Michael – Amazon Web Services 

Jones, Matthew – Nationwide Building Society

Kazantsev, Gary – Bloomberg LP

Patel, Parimal – Independent 

Pearce, Christopher – esure Group

Pearce, Luke – Santander 

Rees, Harriet – Starling Bank Limited 

Szpruch, Lukasz – The Alan Turing Institute 

Taylor, Neil – Mastercard

Valane, Jeffrey – HSBC (virtual)

Wade, David – Goldman Sachs 

Xu, Justin – MillTech

Apologies

Buchanan, Bonnie Gai – University of Surrey

Kazim, Emre – Holistic AI

Li, Feng – Bayes Business School 

Mullins, Inga – Fluency

Prince, Emily – LSEG

Rosenshine, Kate – Microsoft

Observers

Azid, Dominique – ICO

Chai, Soo Yun – ICO

Payne, Chris – HMT

Underhill, Michael – Ofcom

Bank of England

Gharbawi, Mohammed

Graham, Georgette

Lee, Amy

Mutton, Tom

Yoganayagam, Michael (virtual)

Financial Conduct Authority

Hegarty, Robert

Levett, Freddie

Simon, Christopher

Thorman, Libby