B2L Market Focus: Bank AI Concentration Risk Is Becoming Systemic

Bank AI concentration risk is moving from a technology governance problem to a question about the structure of financial markets. Banks are not merely buying software that helps employees work faster. They are embedding models, data services and cloud infrastructure into fraud detection, cyber defence, customer operations, compliance, credit assessment and market workflows. When many institutions rely on the same small group of providers, an error that once would have remained inside one bank can become a common shock.

That is the deeper meaning of the latest regulatory attention. On 12 October, the Financial Stability Board said it had finalised sound practices for responsible AI adoption and warned that artificial intelligence now belongs inside the financial stability agenda. A day later, Reuters reported that regulators in the United States and Europe were asking banks more detailed questions about autonomous agents, cyber controls, outside technology providers and recovery plans. The important issue is not whether AI is useful. It plainly is. The issue is whether efficiency is being purchased by concentrating operational dependence.

The market tends to value AI adoption through visible benefits: lower costs, faster decisions, better fraud detection and more scalable service. Those benefits are real. The less visible side is that banks may become more similar in the tools they use, the data they ingest and the actions their systems recommend. Similarity can reduce idiosyncratic waste while increasing correlated failure. That trade is becoming one of the most important financial infrastructure questions of the next phase of AI adoption.

Why bank AI concentration risk matters now

Three developments have converged. First, adoption has reached a scale where AI is no longer confined to isolated experiments. The Bank of England and Financial Conduct Authority survey found that 75% of responding firms were already using AI and another 10% planned to use it within three years. The median number of use cases was expected to rise from nine to 21. Large British and international banks were already far above the median.

Second, a growing share of those systems comes from outside the institution. One third of the AI use cases in the British survey were third party implementations, up from 17% in 2022. The top three providers accounted for 73% of named cloud providers, 44% of model providers and 33% of data providers. Those figures do not prove that the financial system is fragile. They do show that diversification at the bank level may coexist with concentration at the system level.

Third, AI is becoming more autonomous. The same survey found that 55% of use cases had some degree of automated decision making, although only 2% were fully autonomous. That distinction matters because agentic systems can sequence actions, use tools and respond to new information with less direct human intervention. A conventional model produces an output for a person to review. An agent can observe, decide, act and repeat. The operational boundary therefore moves from prediction quality toward authority, control and recoverability.

The FSB did not suddenly discover these risks this week. Its June consultation report set out 12 sound practices covering board oversight, risk assessment, data governance, explainability, performance monitoring, human control, cyber resilience and third party management. The October step is significant because the work has moved from consultation toward finalised guidance while supervisors are asking more concrete questions about implementation. The direction is clear even if national rules remain different.

Efficiency can create a common dependency

A bank has strong reasons to use an external model or cloud service. Building frontier capability internally is expensive. It requires specialised talent, large computing capacity, continuous testing and access to data. A shared provider can spread those fixed costs across many customers and update the service faster than each bank could on its own.

The economic benefit is similar to other forms of outsourcing. The bank converts a large fixed investment into a service expense and gains access to capabilities that would otherwise be unavailable. Smaller institutions can compete with larger ones. Large institutions can deploy tools across more functions without rebuilding the same infrastructure repeatedly.

Concentration emerges because the best economics often point in the same direction. Banks choose providers with scale, recognised security teams, established integrations and the ability to satisfy procurement requirements. The more firms that use a provider, the more attractive its ecosystem becomes. Developers learn its tools, consultants build around it, internal systems become compatible with it and switching costs rise.

This creates a paradox. Every bank can make a rational decision and the system can still become less diverse. Each institution may believe it has chosen a robust vendor. Yet if many institutions choose the same vendor, one software defect, compromised update, model behaviour change or cloud disruption can affect several firms simultaneously. The risk is not simply that a supplier fails. It is that the supplier becomes a transmission channel.

Block2Learn previously examined how an AI safety pact can turn governance into a valuation test. The banking version is more demanding because a control failure can interrupt payments, credit decisions, fraud screening or market operations. It is no longer enough to ask whether a company has an AI policy. Investors must ask which critical services depend on the same models, compute providers and data pipelines as everyone else.

Agentic systems expand the operational blast radius

Agentic AI changes the control problem because actions can be connected. A cyber defence agent might scan alerts, rank threats, isolate devices, recommend patches and initiate containment. A compliance agent might review transactions, gather documents and escalate suspicious activity. A trading support agent might collect market data, compare hedges and prepare instructions.

None of these functions needs to be fully autonomous to matter. Risk increases when a system can move across tools, inherit permissions or create a rapid sequence of actions that operators struggle to reconstruct. The relevant questions are practical: what can the agent access, which actions require approval, how are decisions logged, how can the system be halted and what happens if the underlying model changes?

The Bank of England AI Consortium discussed the need for outcome based validation, stronger oversight of outside providers and clearer governance. Its June minutes also identified concentration at the model and computing layers, where alternatives may be limited. Members linked this directly to important business services whose disruption could threaten a firm or financial stability.

The FSB framework approaches the same problem from the life cycle of the system. It asks institutions to assess materiality, understand the model, maintain reliable data, test performance and preserve effective human oversight. Those controls become more important as autonomy increases. A person who merely approves a recommendation after a complex chain of machine actions may provide the appearance of control without the substance of it.

This is why the market should not treat an AI deployment count as a simple innovation metric. Ten low materiality assistants are not equivalent to one agent embedded in cyber response, liquidity management or credit approval. The value and the risk depend on where the system sits, what it can do and whether the bank can continue operating without it.

Cyber defence and cyber attack are using the same technology

AI can strengthen financial security. It can process large volumes of signals, identify unusual behaviour, accelerate investigation and help teams prioritise vulnerabilities. The FSB consultation described applications in fraud detection, anti money laundering, cyber security and market surveillance. These benefits explain why refusing to adopt AI can itself become a risk.

The same capabilities can help attackers discover vulnerabilities, generate convincing identities, automate social engineering and scale intrusion attempts. The European Central Bank said in June that more than 85% of banks under its supervision use AI, while warning that frontier models are changing the asymmetry between defenders and adversaries. It cited an example in which criminals used synthetic identities to create thousands of fake customers and cause millions in losses.

That dual use feature makes the risk dynamic. A bank can improve its defences and still face a faster moving threat environment. It can deploy an AI tool to close vulnerabilities while depending on a provider that is also a target. It can automate detection while creating new permissions and interfaces that must be secured.

The FSB chair made the systemic connection explicit in August. He described cyber risk as the most immediate concern from frontier models and called for robust response and recovery capabilities at financial firms and their critical technology providers. The emphasis on providers is crucial. A bank cannot claim resilience solely because its own internal systems are well managed if a common service remains a single point of failure.

Recovery capacity is the hidden capital expense

Markets often assume that better software lowers the cost base. That can be true in normal operation, but resilience creates a second cost curve. Critical systems need inventories, access controls, testing, independent backups, alternative processes, trained staff and recovery environments that do not share the same vulnerabilities as the primary system.

The ECB tested cyber resilience at 109 banks in 2024, with 28 receiving a deeper assessment of their ability to respond to and recover from a severe incident. The exercise found that banks had frameworks in place but also identified areas for improvement. By June 2026, almost three quarters of those findings had been addressed. This is evidence of progress, not a reason to assume the work is finished.

Recovery becomes harder when the disruption affects a widely used provider. A bank may have a technically independent backup, but restoration can still depend on data formats, identity services, model interfaces or external specialists. A backup that cannot be operated at scale is not a recovery plan. A switch that requires the failed vendor to cooperate is not independent.

This connects AI adoption to the broader question of bank resolvability. In an earlier analysis of the missing liquidity bridge in bank resolution, the central issue was whether authorities can preserve critical functions after a financial failure. AI concentration raises a related operational question before insolvency occurs: can critical functions continue when the institution remains solvent but a shared technology layer is impaired?

The answer requires money. Banks may need parallel infrastructure, more frequent recovery exercises, additional cyber insurance, independent validation and contracts that provide access to data and models during disruption. Some of those costs will appear as technology spending. Others will emerge as higher operational risk charges, slower deployment or duplicated capacity. The efficiency case for AI remains intact, but the net saving may be smaller than the headline automation benefit suggests.

Model risk is becoming infrastructure risk

Traditional model risk focuses on the possibility that a model is wrong, poorly designed, misused or applied outside its intended conditions. Those concerns remain. In April, the Federal Reserve, OCC and FDIC revised their model risk management guidance, replacing frameworks dating from 2011 and emphasising a risk based approach tailored to each bank’s profile, size and complexity.

AI concentration adds a different layer. A model can perform acceptably at each institution and still create system risk if many firms respond similarly. Common data may produce common signals. Common models may identify the same threats, reject the same customers or recommend the same defensive actions. A provider update can change behaviour across several firms at once.

The distinction can be framed simply:

Risk layer	Institution level question	System level consequence
Model performance	Does the system produce reliable outputs?	Common errors can align decisions across firms.
Data dependence	Are inputs accurate, secure and representative?	Shared data failures can spread across products and markets.
Provider concentration	Can the bank replace or isolate the service?	A single disruption can affect many institutions together.
Agent authority	Can humans monitor, limit and stop actions?	Automated responses can amplify a shock before coordination begins.
Recovery	Can critical operations restart independently?	Common recovery dependencies can delay the whole system.

The most dangerous case is not necessarily a dramatic model failure. It may be a small defect combined with speed, scale and common use. A flawed security update, a false fraud signal or an incorrect classification can be manageable in one institution. It becomes more serious if several institutions take similar actions at the same time and customers lose access to services across providers.

The cross market transmission runs through trust

A severe AI related disruption would first appear as an operational event. The financial effects could then move through several markets.

Bank equities would price remediation costs, service interruption and possible customer attrition. Credit spreads could widen if investors believed operating losses or litigation would weaken capital generation. Cyber insurance pricing could rise, while policy exclusions become more restrictive. Technology vendors might benefit from greater demand for resilience services even as the largest providers face more scrutiny and contractual liability.

Payment and market infrastructure could transmit the shock further. If banks cannot process transactions, confirm positions or manage collateral normally, clients may conserve cash and reduce activity. Liquidity can deteriorate even when balance sheets remain solvent. A technical incident can therefore create a temporary demand for central bank liquidity or supervisory coordination without beginning as a credit problem.

The confidence channel may be more important than the direct loss. Depositors and corporate treasurers do not need to understand the model failure to move funds. They only need to doubt access. Markets then reprice the institution that appears most exposed, the providers associated with the disruption and any firms believed to share the same architecture.

This is where Europe’s debate about supervision becomes relevant. Block2Learn’s analysis of Europe’s two speed market union showed how fragmented oversight can preserve national control while complicating common market infrastructure. AI providers operate across borders, while bank recovery remains rooted in legal entities and national authorities. A cross border technology failure could therefore move faster than the supervisory response.

What appears priced and what still looks underpriced

The market has priced the revenue opportunity for large model and cloud providers. It has also begun to price the productivity gains available to banks. Investors routinely reward cost guidance that promises more automation and scalable service.

Some regulatory cost is also visible. Banks disclose cyber investment, technology modernisation and operational resilience spending. Large institutions already run extensive vendor risk programmes and recovery tests. The largest providers know they face greater scrutiny as critical infrastructure.

What still looks underpriced is the correlation between those exposures. Equity analysts often evaluate a bank’s technology plan against its own cost base. Credit investors assess operational risk within the issuer. Technology investors model customer growth at the provider. These perspectives can miss the common dependency that connects them.

Switching costs may also be underappreciated. A bank can negotiate a contract with an alternative supplier, but migration requires compatible data, validated models, trained staff and tested workflows. Redundancy on paper may not produce substitutability in practice. True resilience is measured by whether the alternative can carry the service under stress.

The third underpriced element is behavioural convergence. If institutions use similar tools to monitor markets, detect fraud or prioritise threats, they may react to the same signals in similar ways. This does not require autonomous trading. It can happen through risk limits, client restrictions, liquidity decisions or cyber containment actions. Correlated caution can tighten financial conditions without any institution intending to do so.

Three paths for the next phase

Managed adoption

Banks expand AI use while mapping critical dependencies, preserving human authority and testing independent recovery. Providers improve transparency, portability and audit access. Supervisors coordinate expectations without forcing one rigid technology standard. Efficiency gains exceed resilience costs and AI becomes another managed layer of financial infrastructure.

Concentration with higher resilience costs

Adoption continues, but the same small group of providers dominates models, cloud and data. Banks maintain more duplicated capacity, regulators demand stronger recovery plans and insurers charge more for cyber exposure. AI still improves productivity, but the net benefit is shared with vendors, consultants, security firms and backup providers. Bank valuations reflect a lower saving than optimistic cost plans imply.

A common provider shock

A model update, cyber incident or infrastructure failure affects several institutions at once. Banks restrict services while attempting to isolate systems. Customers move cash, market makers reduce risk and supervisors coordinate recovery. The financial loss may be contained, but the event changes procurement and capital allocation for years. Provider concentration becomes a formal financial stability issue rather than a supervisory concern.

What would invalidate the concentration thesis?

The thesis would weaken if banks demonstrate genuine portability across models and cloud services, including successful live recovery exercises. It would also weaken if critical AI use remains limited, with most systems confined to low materiality productivity tools and strong human control.

A broad, competitive provider market would reduce concentration. Open standards, interoperable data and credible smaller suppliers could lower switching costs. Institutions might also develop more important systems internally, preserving diversity in models and controls.

Finally, AI could improve defence faster than it expands the attack surface. If detection, patching and response become materially better, operational losses may fall even as adoption rises. The correct conclusion is not that AI necessarily makes banking less safe. It is that the distribution of providers and controls determines whether benefits remain local while failures become common.

What to monitor next

The first signal is implementation of the FSB sound practices. The document is not a binding global rule, so national supervisors will decide how its principles become examinations, guidance and recovery expectations.

The second is provider mapping. Investors should look for evidence that banks know which critical services depend on the same models, data and cloud infrastructure, including indirect dependencies through contractors.

The third is recovery testing. A useful exercise does not merely restore a server. It proves that the bank can continue a critical service when its primary model or provider is unavailable or untrusted.

The fourth is agent authority. Disclosures about human oversight matter only when they explain approval thresholds, logs, permissions and shutdown mechanisms. The Bank of England consortium’s discussion of intervention and kill switches points toward the practical standard supervisors may expect.

The fifth is the economics. Technology spending, cyber insurance, vendor pricing and duplicated infrastructure will reveal how much of the efficiency gain remains with the bank. A rising cost base does not mean adoption has failed. It may mean resilience is finally being priced honestly.

AI is becoming part of the financial system’s plumbing

The latest regulatory scrutiny is not a rejection of innovation. It is recognition that a tool can be valuable at the firm level and consequential at the system level. Banks are adopting AI because it improves analysis, security and service. The same adoption can create common providers, common models and common recovery dependencies.

That makes bank AI concentration risk a question of market structure. The winners will not simply be the institutions that deploy the most systems or promise the largest cost reductions. They will be the institutions that understand where autonomy sits, preserve operational choice and can recover without waiting for the same provider as everyone else.

The next phase of financial AI will be measured less by the number of experiments and more by the resilience of critical services. To build the concepts behind operational risk, financial institutions and market transmission in a structured sequence, continue with the Block2Learn Learning Path.