When Finance Thinks Alike, It Fails Together

Imagine that every risk committee receives excellent advice.

The models are current. The data is relevant. The recommendations are explainable. Each institution reviews the evidence and makes a defensible decision.

Then they all make the same decision at the same time.

They reduce exposure to the same sector. Tighten credit to the same customers. Sell the same assets. Flag the same transactions. Move liquidity toward the same safe destination.

No institution behaves irrationally. The system still becomes unstable.

Finance has always had herd behavior. AI changes its speed and uniformity. Common providers, data, objectives, and industry practices can turn independent firms into variations of the same decision engine.

The systemic risk is not only that AI will make a bad decision. It is that many institutions will make the same good decision and discover that it becomes bad when everyone acts on it together.

The next generation of financial resilience will require diversity by design.

Rational decisions can produce an irrational system

A bank manages its own risks. A supervisor watches the system created by all banks managing their risks simultaneously.

Those perspectives are not identical.

Selling an asset before its price falls may be prudent for one portfolio. Widespread selling can cause the fall. Restricting credit after a risk signal may protect one balance sheet. If lenders respond together, the restriction can weaken otherwise viable businesses and validate the signal. Moving deposits or collateral toward the safest institution can be logical for each customer and destabilizing for everyone else.

AI can intensify this gap between microprudence and macrofragility because it makes similar analysis cheaper and faster. The same capability that helps one firm notice a risk helps every other firm notice it too.

The Financial Stability Board's report on AI identifies third-party concentration, market correlations, cyber risk, and weaknesses in models, data, and governance as vulnerabilities with systemic potential. These risks interact. Common providers can create common blind spots; common blind spots can produce correlated behavior; faster execution can leave less time to interrupt it.

The problem is therefore larger than model accuracy. It concerns the architecture of the market around the model.

The same intelligence is already entering many institutions

Financial institutions do not need to build identical systems deliberately. Market economics can move them in that direction.

The strongest models require enormous investment. Infrastructure is concentrated. Vendors embed the same models in industry software, consultants replicate successful patterns, and benchmarking rewards convergence toward what peers accept.

The Bank of England and FCA survey of AI in UK financial services found that the three largest providers accounted for 44% of reported third-party model supply, while a third of use cases were third-party implementations. This does not prove firms will act alike. It shows how shared dependencies form.

As argued in When Every Bank Has the Same AI, What Still Matters?, common model access shifts competitive advantage toward context, integration, control, and learning. From a stability perspective, however, common access raises a second question: what happens when those institutions also receive similar outputs at a critical moment?

A model can be interchangeable commercially and highly consequential systemically.

Architecture changes behavior

It would be a mistake to assume that every type of AI creates the same collective dynamics.

An ECB research bulletin published in 2026 reports simulation-based experiments in a stylized redemption game. Q-learning agents coordinated strongly but could fall into extreme, bank-run-like redemption behavior. Large language model agents were less prone to those runs but produced more heterogeneous and unpredictable decisions.

This was an experiment, not a prediction of the next crisis. Its importance lies in the mechanism. Agents facing the same environment and goal behaved differently because their decision architectures differed.

That means a conventional stress test may be incomplete. Institutions usually test what happens when prices, defaults, volatility, or liquidity change. They will also need to ask what happens when multiple AI systems observe the shock, form expectations about one another, and act.

The IMF's October 2024 Global Financial Stability Report warned that wider AI adoption could increase market speed and volatility under stress, deepen dependence on a few providers, and create monitoring challenges. Value in normal conditions can coexist with uncertainty under collective pressure.

Monoculture extends beyond trading

Market selling is the easiest synchronized behavior to imagine. The same pattern can emerge elsewhere.

In credit, common models may reduce exposure to a region or industry simultaneously, turning a risk forecast into a funding shock.

In fraud, institutions may learn the same indicators and block the same category of legitimate customer, while criminals adapt to the shared rule.

In wealth management, recommendation engines may rebalance households toward similar products or react to volatility using comparable thresholds.

In operations, one shared model, cloud, identity layer, or agent framework may transmit a failure across institutions.

In supervision, common analytical tools could help authorities see risk earlier. They could also narrow attention toward what the shared system is good at detecting.

The bank can become one brain without becoming one mind. The financial system needs the same distinction. Connection can improve coordination, but independent judgment is part of resilience.

Build diversity capital

Banks hold financial capital against losses. They build liquidity buffers against funding stress. In an AI-mediated system, they will also need diversity capital: deliberate differences that reduce the probability of common failure.

I would begin with six components.

1. Provider diversity

Map material dependencies across foundation models, cloud platforms, data suppliers, and agent infrastructure. A second vendor is useful only if it does not depend on the same hidden layer.

2. Model diversity

Use different approaches for consequential decisions. Deterministic rules, statistical models, and language models may fail differently. Agreement informs; disagreement is a signal.

3. Information diversity

Preserve private, institution-specific evidence and frontline context. If every system consumes the same public narrative, markets lose independent signals. Proprietary context matters for resilience as well as competition.

4. Decision-path independence

A second approval is not independent when both reviewers receive the same generated analysis. Material decisions may require a separate team, source set, model, or scenario before the outputs meet.

5. Collective-behavior stress tests

Test how customers, counterparties, competitors, and agents respond to the same event. Include feedback loops between recommendations and market changes.

6. Temporal diversity

Not every system should act at maximum speed. Batches, randomized execution, thresholds, cooling periods, and circuit breakers can stop individually sensible actions from becoming a synchronized rush.

Diversity capital is not randomness. It is controlled plurality designed around material failure modes.

Procurement becomes a stability decision

The choice of vendor may look like a firm-level technology decision. At sufficient scale, it contributes to market architecture.

Procurement should ask who else depends on the service, which layers sit underneath it, how changes propagate, whether outputs can be independently compared, and what happens during a common stress event.

This extends the argument in Banks Are Buying AI With Contracts Written for 2015. Concentration cannot be managed through uptime commitments alone. A service may remain available and still transmit the same analytical error, behavioral bias, or change across much of the market.

The model may be the cheapest part of enterprise AI. The dependency created around it can be one of the most expensive.

Supervisors will need a system-level view

Individual model validation cannot observe a collective feedback loop that exists only when many firms deploy similar systems.

Supervisors will need information about concentration, architecture, autonomy, decision speed, and correlated use cases. Exercises should include several institutions and critical providers, with synthetic agents representing market participants reacting to one another.

A BIS-published address on responsible AI in finance specifically highlighted herding from widely used AI trading models and called for scenario stress testing, red teaming, explainability, and resilience. Those controls should not stop at the institutional boundary.

The purpose is not to make every firm use a different technology. Forced difference can create cost without resilience. The objective is to identify where similarity becomes material and ensure that the system retains alternative signals, routes, and response times.

Resilience needs independent thought

AI will help financial institutions see more, decide faster, and respond with greater precision. Those are real advantages.

They also change the unit of risk.

The institution can no longer evaluate only whether its model is accurate, fair, secure, and controlled. It must consider how the model behaves in a market filled with other systems trained on similar data, supplied by the same firms, pursuing comparable objectives, and reacting at machine speed.

Efficiency removes duplication. Resilience sometimes depends on it.

A second model, an independent analyst, a slower decision path, or a different source may look wasteful in normal conditions. During a common shock, that difference may be the only reason everyone does not run toward the same exit.

The most dangerous financial AI may not be the one that makes an obviously foolish decision.

It may be the one that gives every institution the same intelligent reason to move at once.