Deep Dive AI governance & financial stability
Who owns AI risk? The micro/macroprudential split.
AI risk spans firm-level controls, market-wide vulnerabilities and the interactions between agents. What the latest regulatory signals mean for EMEA operational risk teams.
The Bank of England’s Financial Policy Committee records system-wide AI funding and valuation concerns.
The Basel Committee agrees to review operational-risk event-type loss categories for cyber and AI.
ECB President Christine Lagarde calls for macroprudential monitoring and coordination across mandates.
The recent statements from Basel, the Bank of England and the ECB do not describe one settled regulatory home for AI risk. They expose three connected problems: a firm’s own controls and losses, financial exposures that build across markets, and instability that can emerge when many firms’ AI agents react together. EMEA operational risk teams can own the first layer and prepare useful evidence for the others; they cannot see or manage the whole system from inside one firm.
That distinction matters now. Basel is reviewing operational-risk loss categories, while the FPC and ECB are describing risks that cross firms and markets. The immediate task for a firm is to make its own controls and incident data legible, while preparing to coordinate on shared dependencies and correlated behaviour.
What the institutions actually said.
Basel Committee, 1 October. The Committee said frontier AI could amplify operational vulnerabilities through cyberattacks and correlated dependencies in the financial system. Its specific decision was to review the sufficiency and adequacy of existing operational-risk “event type” loss categories, with a focus on cyber risk and AI developments. That is a taxonomy review, not a decision to create AI-specific capital rules.
Bank of England Financial Policy Committee, record published 1 October. The FPC cited Morgan Stanley’s estimate that global AI-related debt issuance totalled about $450 billion as of early September, more than twice the total for 2025. It also reported that issuance by AI hyperscalers in sterling credit markets accounted for 47% of GBP corporate bond issuance so far this year. The Committee flagged opaque and sometimes circular financing arrangements and the possibility that weaker-than-expected AI productivity gains could affect both AI valuations and sovereign debt markets.
ECB President Christine Lagarde, 1 October. Her speech covered agentic trading, cyber resilience and geopolitical dependence on frontier models. She said the interaction of these risks calls for “careful and encompassing macroprudential monitoring” and that authorities should use their respective mandates and coordinate where risks cross them. The speech describes a need for coordination; the claim that no single body owns every layer is an interpretation of that mandate split.
Basel Committee, 1 October 2026 · Bank of England FPC Record, September 2026 · Christine Lagarde, “Where AI risks meet”, 1 October 2026
Why the split exists—and where it strains.
Microprudential supervision asks whether an individual firm is safe and sound. Macroprudential policy looks at whether the system as a whole is stable: individually prudent firms can still create fragility when they share exposures or respond in similar ways to the same shock.
In October 2024, Bank of England Deputy Governor Sarah Breeden argued that existing technology-neutral rules could remain useful at firm level if firms understand their models, while common model providers and correlated trading responses could create vulnerabilities that require a system-wide view. She pointed to stress testing and the AI Consortium as ways to build that understanding. The BIS Financial Stability Institute reached a more conservative conclusion: most AI risks in finance resemble familiar risks, though governance, model risk, data governance and third-party oversight may need further attention.
A third concern is the stability that can emerge from agents interacting, even when each follows its own risk-reduction objective. UNCTAD economist Piergiuseppe Fortunato describes how many agents reacting to volatility at once could amplify the instability they were designed to avoid. This is a useful extension of the micro/macro lens: look at the system behaviour created by the interactions, not only each agent or each firm.
Sarah Breeden, Bank of England, October 2024 · BIS FSI Insights No. 63, December 2024 · Piergiuseppe Fortunato, ProMarket, 18 September 2026
A taxonomy question with capital-adjacent consequences.
Under Basel 3.1 as implemented by the PRA, Pillar 1 operational-risk capital uses a Standardised Approach. Its Business Indicator Component is based on financial-statement measures, and the UK has set the Internal Loss Multiplier (ILM) to 1. As a result, a firm’s historical losses do not mechanically change that Pillar 1 capital number.
That does not make event data irrelevant. The taxonomy helps firms and supervisors classify and compare operational-risk losses, and loss information can inform supervisory assessment outside the Pillar 1 calculation. The Basel Committee’s announced review concerns those event-type categories; it does not itself decide how capital should change. Any eventual capital treatment would require separate policy work.
For an operational risk team, the near-term implication is a data one. If cyber and AI become more clearly distinguished in the taxonomy, firms will need reliable, comparable incident records. It is sensible to prepare for that possibility, while treating any future reporting or capital change as uncertain until Basel or national authorities publish details.
PRA, Implementation of the Basel 3.1 standards: Operational risk
A clean firm-level control set cannot remove system-wide exposure.
The FPC’s AI-debt and valuation concerns are tracked across markets. A bank’s operational risk function can assess its own exposures to AI-related borrowers, vendors and infrastructure, but it cannot observe the full financing network or control how other firms respond to a correction.
Agent interaction creates a related visibility problem. The Bank of England and FCA’s AI Consortium is already discussing concentration risk from third-party model providers and AI-accelerated contagion in financial markets. That work is active, but discussion and research are not the same as a shared, cross-firm regime that stress-tests correlated AI-agent behaviour. The reviewed sources do not show such a regime in operation.
MAS and industry partners’ Safeguards for Agentic Finance at Runtime (SAFR) is a useful operational reference for runtime safeguards. It is an industry-developed white paper under MAS’s BuildFin.ai initiative, not a dedicated capital regime. Its focus is how agent actions are authorised, checked and recorded at the point of action.
Bank of England / FCA Artificial Intelligence Consortium · MAS and industry partners, SAFR white paper, 3 July 2026
Prepare the firm’s evidence and escalation routes.
Day-to-day ownership of AI governance, model risk, data governance and supervisory dialogue is likely to remain within firm-level functions. The macroprudential questions call for coordination with regulators and the wider sector, but a firm can make its own contribution more useful by keeping event and exposure information structured.
Start a consistent AI-related loss-event log and a separate exposure / near-miss register. Keep realised losses distinct from failures caught before loss, vendor exposures and correlated-behaviour findings.
The loss-event log can use the existing Basel event-type category, with an explicit cyber/AI flag, trigger, realised financial impact, business line and detection method. A separate forward-looking register can capture what was caught, how close it came to a loss, which dependency or agent was involved, and which control detected or stopped it. Keeping the registers distinct helps preserve loss attribution while building a wider view of emerging exposure.
Three useful questions for a risk committee follow:
- 01
Can we distinguish AI-related operational losses from near misses and exposures in our internal records today?
- 02
Which shared model, cloud or data dependencies could make our response correlated with that of other firms?
- 03
Who owns escalation when an AI-related event crosses from firm-level control into a market-wide or multi-firm concern?
The next signals to track.
Watch for Basel’s follow-up on operational-risk event-type categories; any PRA or ECB implementation choices; further FPC analysis of AI financing and sovereign-market exposure; and whether the AI Consortium or other authorities move from discussion toward shared testing of correlated agent behaviour. MAS’s SAFR is worth following as an industry runtime-safeguards reference, while keeping its status distinct from prudential capital rules.
Sources