NextFin News - The Bank of England is warning that AI is moving from a productivity story to a market-structure story. In a letter to lawmakers, Deputy Governor Sarah Breeden said the central bank is working with international counterparts to understand when AI agents trading in financial markets could “demonstrate correlated behaviour or ‘herding’” and potentially amplify a stress scenario. The warning is not that AI is already breaking markets. It is that autonomous systems could eventually make an ordinary selloff move faster, farther, and with less human visibility than investors are used to.
That matters because the next phase of AI adoption is not about chatbots sitting at the edge of the business. It is about systems moving deeper into execution, risk management, and payments. The Bank’s concern is a familiar one in modern market regulation: when many participants use similar tools, similar data, and similar objectives, they can end up taking the same action at the same time. In calm conditions that can look efficient. In a shock, it can look like a stampede.
The Bank said financial system participants have not yet adopted advanced generative or agentic AI in a way that would present systemic risk. It also said risks are likely to rise, potentially rapidly, as deployment grows. That is the key nuance. The regulator is not claiming an AI-triggered flash event is imminent. It is saying the market may not get much warning once autonomous tools become common enough to matter.
The warning also reflects how central banks are now thinking about fragility. The issue is no longer only whether a model is accurate. It is whether a model’s behavior becomes correlated across firms when markets turn. If one AI system reduces risk, many others may do the same. If one system exits liquidity, others may follow. That can create a feedback loop in which prices fall, volatility rises, more systems de-risk, and prices fall again.
Breeden’s comments fit into a broader 2026 policy shift in which the Bank is treating AI as a potential source of financial-stability risk rather than just a compliance issue. The institution said it is incorporating AI scenarios into cyber and operational testing and studying how agent design might be shaped to reflect public policy objectives. That is a strong signal that supervisors are already thinking about the structure of machine behavior, not only the performance of individual models.
“We are working with international counterparts to better understand the conditions under which AI agents trading in financial markets could demonstrate correlated behaviour or ‘herding’ and so potentially exacerbate procyclical dynamics to amplify a stress scenario.”
The phrase “procyclical dynamics” is the heart of the warning. It refers to behavior that reinforces the cycle already underway. In markets, that usually means selling into weakness and buying into strength. AI can intensify that pattern by compressing decision times and making similar responses more likely. A human trader might hesitate when a market starts to unravel. An agentic system may have already executed the de-risking before the hesitation even occurs.
That is one reason the Bank’s concern is broader than model quality or software bugs. A perfectly functioning system can still be destabilizing if it behaves like every other system at the same time. Markets have always had herding. What changes with AI is the speed, scale, and opacity of that herding. If the same training data, risk constraints, and reward functions dominate, then a fast-moving market may end up with fewer genuinely independent decision-makers than participants assume.
The Bank’s own assessment is still measured. It said current adoption of advanced forms of AI does not yet present systemic risk. But it added that risks are likely to increase as use expands. That makes the near-term question one of threshold rather than certainty. A market can appear stable right up until enough autonomy is concentrated in enough important firms that a common response becomes self-reinforcing.
That threshold matters because the financial system is already built around feedback loops. Leverage, margin, collateral calls, execution algorithms, passive flows, and volatility targeting all push in the same direction under stress. Agentic AI could add a layer that not only reacts to those forces, but reacts to them faster. If the system begins to process the same shock in near real time across many firms, liquidity can disappear before human oversight can slow the process.
The concern is not that AI has to be “wrong” to be dangerous. It is that AI may be right in the same way at the same time. That is a subtle but important distinction. A market full of smart agents can still be fragile if those agents share the same blind spots.
Why Regulators Are Focused on Agentic AI
Agentic AI is the relevant frontier because it is not just generative text or pattern recognition. It is software that can act with more autonomy across a workflow. That makes it useful for trading, execution, customer service, monitoring, and operational tasks. It also makes supervision harder. If a firm cannot clearly explain why the model changed course, then management may not realize how quickly a local optimization problem can become a systemwide one.
The Bank’s 1 April response to lawmakers makes clear that it is not waiting for a crisis before studying the issue. It said the institution is proactively promoting responsible adoption of new technologies and is evaluating and monitoring the financial-stability implications of AI investment, development, and adoption. It also said it is working to incorporate AI scenarios into cyber and operational testing. Those are not the actions of a regulator treating AI as a distant hypothetical.
“Risks are likely to increase, potentially rapidly, amid growing intent among financial firms to grow their deployment of advanced AI.”
That line suggests the Bank sees a nonlinear risk path. Adoption may look gradual, then move quickly once use cases prove profitable. That is often how market infrastructure changes happen. A tool starts at the edge, becomes routine, and then becomes embedded in the most sensitive parts of the system before supervisors have enough experience to judge how it behaves in stress.
The real issue is that the market may not know when it has crossed from efficiency into fragility. If a trading model improves execution quality in normal conditions, firms will want more of it. If several firms buy the same kind of improvement at the same time, the market can become more correlated, not less. In a quiet tape that correlation is invisible. In a volatile tape it can become the dominant fact.
That is why the Bank is studying how AI agents might “herd” and why that herding might worsen procyclical dynamics. Herding is not new, but AI can make it more machine-like. It can compress the time between signal and action, and it can do so across multiple functions at once. If the same shock hits funding markets, cash equities, and derivatives simultaneously, autonomous systems may react in the same direction across those venues before humans can intervene.
The policy response, for now, is to learn faster than the risk spreads. The Bank said it is working with international counterparts because the problem is not confined to one market or one jurisdiction. If model developers, trading firms, exchanges, and supervisors are all using similar technologies, then the risk of synchronized behavior becomes cross-border as well as cross-asset.
That makes the current stage of the debate important. Regulators are not yet writing a crisis playbook for AI-induced market breaks. They are still trying to identify the failure mode. In systemic-risk terms, that is the right order. The market does not need a finished apocalypse to justify preparation; it needs a plausible mechanism.
How AI Could Amplify A Selloff
The clearest mechanism is correlation. Markets can absorb many different mistakes, but they struggle when too many participants make the same mistake at once. If AI agents are trained on similar data and optimized against similar risk limits, their behavior can become highly correlated under stress. The first wave of selling can trigger the second, not because every model is irrational, but because every model sees the same deterioration and responds according to the same playbook.
Speed is the second mechanism. Human traders can pause, override, or interpret a fast-moving market. Autonomous systems compress that reaction time. When volatility spikes, that speed can become dangerous if many firms are using similar models to cut exposure, widen spreads, or pull orders at the same moment. A market that loses liquidity for a few seconds can recover. A market that loses liquidity across many venues at once can cascade.
Leverage is the third mechanism. When prices fall, margin and collateral pressures rise. If AI is helping manage risk in leveraged books, then it may de-risk just as collateral systems are tightening. That creates a loop: weaker prices force more defensive behavior, and more defensive behavior weakens prices further. The structure is familiar. What changes is the rate at which the loop can spin.
This is why the warning should be read as a market-structure issue, not a science-fiction story about machines taking over trading desks. The danger is not autonomy in the abstract. It is autonomy interacting with the existing procyclical machinery of modern finance. A small shock can become a larger one if systems designed to protect firms collectively intensify the move.
There is also a governance challenge inside firms. If supervisors cannot easily explain what the model is doing, then the model can become a source of hidden concentration. A firm may think it is diversifying decision-making by using advanced AI, while in practice it is embedding the same logic in multiple places. That can create a false sense of resilience.
One implication is that future stress tests may need to move beyond balance sheets and into behavior. The Bank is already signaling that it wants to understand not just whether firms use AI, but how those systems act under stress. That could eventually mean testing whether multiple AI systems reach similar conclusions under the same shock, and whether those conclusions lead to destabilizing market-wide responses.
For now, the most important point is that the regulator sees a window for prevention. It said current adoption does not yet present systemic risk. It also said risk may rise quickly. Those two statements together imply that the system is still early in the curve, but not for long if deployment expands across trading and payments.
What Happens Next
The immediate market implication is not a direct tradeable signal. It is a regulatory one. The Bank of England is telegraphing that AI is likely to appear more often in supervisory discussions, stress testing, and cross-border coordination. Firms that rely on advanced AI will probably face more questions about governance, explainability, and fallback procedures.
For investors, the wider takeaway is that AI risk now cuts in two directions. It can support productivity and margins, but it can also create a new source of systemic correlation if it becomes deeply embedded in market plumbing. That does not mean the technology is destabilizing by nature. It means the market will need guardrails if autonomy keeps moving deeper into the financial system.
The next catalysts are likely to be policy documents, supervisory reviews, and further central-bank research rather than a single market event. The Bank said it is working on simulation methods and AI scenarios, and that suggests more evidence will emerge before any hard rule changes do. That process may be slow in public, but it is moving faster in the background than many market participants may expect.
The paradox behind Breeden’s warning is straightforward. The most efficient systems are not always the most resilient. If AI agents all learn from the same past and react to the same present, the market can look more advanced right up until it looks more fragile.
In that sense, the central bank’s message is less about fear than about timing. The danger is not that AI will suddenly invent a new kind of panic. It is that it may make old panics arrive faster than the market can absorb them.
Explore more exclusive insights at nextfin.ai.
