Trustworthy AI in Financial Services: Managing Risks and Opportunities

As financial institutions (FIs) increasingly adopt artificial intelligence, the question is no longer only whether AI makes individual firms more efficient. It is also whether shared AI dependencies could create new sources of correlation and vulnerability across the financial system, even when firms act independently.

Researchers from Work Package 4 at the University of East Anglia’s Norwich Business School are exploring this question. In a new policy brief, Sean Ennis, Raphael Markellos, Nikolaos Vlastakis, Nabhoneel Deb and Amir Jafarzadeh identify three structural conditions that could potentially amplify systemic risks in financial institutions:

High concentration across the AI stack. 

• FIs increasingly source AI capabilities from the same providers across cloud infrastructure, foundation models and specialised platforms. The three largest cloud providers account for around two-thirds of the market. A single outage or attack could therefore affect several layers of many institutions’ AI capabilities at once.

Internal controls only partially offset shared dependencies. 

• Governance maturity varies widely, automation is moving from advice towards autonomous execution, and many institutions rely on standardised external models. A 2025 Alan Turing Institute survey found that 66% of FIs deploying LLMs used external providers without internal customisation.

Supervision has structural blind spots. 

• AI vendors and cloud providers largely sit outside direct financial supervision. Existing rules do not fully address tacit algorithmic convergence, while current oversight remains focused primarily on individual institutions and is fragmented across borders. According to an IOSCO survey, only around a third of supervisors collaborate with overseas authorities on AI.

Across the industry and regulatory sources reviewed, 56% flag infrastructure concentration and 60% flag supervisory opacity as emerging concerns.

Policy recommendations

To reduce the risk that shared dependencies become systemic vulnerabilities, the policy brief proposes five areas for supervisory action:

1. Extend oversight of critical third parties. 

2. Supervisors should explore extending critical third-party oversight regimes to AI vendors, foundation-model providers and cloud hosts, with designation lists that are updated continuously.

3. Map AI dependencies. Supervisors should improve their understanding of FIs’ dependencies across vendor, model and cloud layers, for example through dependency registers.

4. Strengthen AI governance. Supervisors should require minimum governance standards at the point of AI adoption, covering model ownership, accountability and escalation procedures.

5. Establish fallback mechanisms. Supervisors should consider pre-approved fallback mechanisms for highly automated AI applications.

6. Strengthen cross-border cooperation. Supervisors should develop joint frameworks between national jurisdictions to address gaps created by cross-border AI development.