Risks of AI in banking: the 10 risks regulators name, with the banks' own controls.

What are the risks of AI in banking?

Ten, by the count of what the regulators' documents name: model risk, bias and fair lending, data privacy, third-party and vendor risk, cyber and adversarial attack, hallucination, concentration and systemic risk, operational resilience, explainability, and consumer harm or deception. The table maps each to the documents that name it in this site's tracker of 238 documents and to the banks that disclose a control for it; 137 of the 140 banks profiled here disclose at least one.

The list is assembled from what the supervisors wrote rather than from a vendor's taxonomy. Model risk and third-party risk are the two with their own guidance and examination handbooks; bias, explainability and consumer harm are the three that consumer law reaches; privacy is a data-protection matter that AI sharpened; cyber, hallucination and resilience are the three the generative wave added to the supervisors' letters in 2024 to 2026; concentration is the one the financial-stability bodies watch. They overlap (a hallucination that reaches a customer is consumer harm; a vendor model that fails is a resilience event), which is why banks govern them through one framework rather than ten programmes. For each risk the table gives the documents that name it, the number of banks on this site that disclose a control, and the control in one line.

Which documents name each risk, and which banks disclose a control?

Model risk: SR 26-2 and PRA SS1/23, with 96 banks disclosing validation or inventory controls. Bias: ECOA and the 2026 Regulation B rule, 5 banks. Privacy: GDPR Article 22 and the ICO guidance, 29 banks. Third parties: SR 23-4 and the Basel principles, 112 banks. Cyber: Treasury's 2024 report and the ECB's 2026 letter, 82 banks. Hallucination: NIST AI 600-1 and the CFPB chatbot spotlight, 22 banks. Concentration: the FSB's 2024 and 2025 reports, 8 banks. Resilience: DORA, 23 banks. Explainability: FCRA and the ECB's internal-models guide, 35 banks. Consumer harm: the Consumer Duty and Colorado's ADMT Act, 9 banks.

The bank column is computed: it counts the banks whose profiles record an activity in production or rolling out that describes a control for the risk, and names the five largest. A bank absent from a row has not disclosed a control, which is not the same as not having one; the largest banks disclose most, in their 10-Ks and 20-Fs, and the smallest disclose least. Click a bank for the record and its source; click a document for its plain-language answer and official link.

Three rows deserve a note. Hallucination is the risk with the widest gap between the regulators' attention and the banks' disclosure: the documents are explicit, but most banks describe the control only as “human in the loop”; the ones that say more (Deutsche Bank's inline citations and audit logging, Goldman's Marquee AI attributing statements to source lines, SoFi's planners evaluating Coach's answers) are the ones to copy. Concentration is a system-level risk that a single bank cannot remove, and the control it can disclose is optionality: Bank of America's Orchestra routes simple tasks to open-weight models on its own GPUs, Goldman's platform lets users choose among GPT, Gemini and Claude, BNP Paribas runs Mistral on premises for its most sensitive KYC steps. Privacy is where the newest controls are: JPMorgan does not let third parties log in with customers' credentials, so outside agents can read through open-banking connectors but cannot move money, and Goldman's prompt controls check what staff enter before it reaches a model.

What do the banks themselves say the risks are?

103 banks describe AI in the risk factors of their annual filings. Goldman Sachs names generative-AI output, bias, opacity and third-party models and a Firmwide AI Risk Committee co-chaired by the CIO and the chief risk officer; Bank of America's 10-K covers autonomous agents, third-party models with limited visibility into training data, inaccurate outputs and bias; Citi's covers generative AI, autonomous agents, explainability, bias and third-party models with board Technology Committee review; JPMorgan's puts machine-learning and AI models inside model risk management and names generative AI; Morgan Stanley's says risk models may not keep pace with new technology.

The most candid statements of AI risk come from the banks' own annual filings, where the risk-factor section is written by lawyers for shareholders rather than by marketing for customers. The table lists every bank whose profile records what its 10-K, 20-F or annual report says about AI risk and governance; the largest file the fullest accounts, and the pattern across them is consistent: generative output that may be inaccurate or biased, autonomous agents, third-party models the bank cannot see inside, regulatory uncertainty, and the possibility that risk models do not keep pace with the technology.

What is the AI risk management framework regulators expect?

An AI committee or council that approves use cases, an inventory with risk tiers, validation and monitoring for models in scope, data and prompt controls, vendor due diligence, human oversight with a route to a person, and incident reporting to the board; mapped to SR 26-2, SR 23-4, the CFPB's adverse-action rules, the EU AI Act, NIST's AI RMF and generative-AI profile, Treasury's financial-services adaptation of it, DORA and the FSB's sound practices. The AI governance page sets out the framework element by element; this page is the risk view of the same material.

Every supervisor that has written about it describes the same architecture, because it is the architecture banks already run for models, technology and operations. The board owns an inventory of what is running; a committee approves use cases against a risk assessment; models in scope are validated and monitored; data is owned and traceable; vendors are diligenced and monitored; a person can explain and override any decision that touches a customer; incidents are reported upward. The AI governance in banking page sets that out element by element with the documents that require each and the banks that disclose each; model risk management covers the validation discipline in detail, and the agentic-banking control plane covers what changes when the AI acts rather than advises: identity for agents, limits, kill switches, audit trails and dispute processes.

The guides on this site explain the framework documents one at a time: SR 11-7 and what SR 26-2 changed, PRA SS1/23, the NIST AI RMF, DORA and BCBS 239.

Which documents matter most for AI risk in banking?

12 documents cover the ten risks between them: SR 26-2 and SR 23-4 for models and vendors in the US, the 2026 Regulation B rule and the EU AI Act for bias, NIST AI 600-1 for generative-AI risks, Treasury's 2024 report, the NY DFS letter and the ECB's 2026 letter for cyber, the FSB's 2024 report and 2026 consultation for systemic risk, DORA for resilience and the Basel third-party principles for the supply chain.

The documents below cover the ten risks between them. Each links to the tracker page with its plain-language answer, its status and its official text; the full set by use case is on the by-use-case page, and the deadlines page lists what is due when, including the ECB's 31 October 2026 action plans and the EU AI Act's deferred high-risk date of 2 December 2027.

Which vendors sell AI risk management tools?

5 published vendors in this site's directory sell AI governance or model-risk software (inventory, tiering, validation workflow, drift and bias monitoring, policy packs for the EU AI Act, NIST AI RMF and ISO 42001): Credo AI, Monitaur, SAS Model Risk, ValidMind, Yields. None of the 120 US banks here names one on the record; the controls banks describe are built on their existing model-risk and GRC platforms.

A small category with no bank customer yet named on the record here, which says more about disclosure than about the market: a bank that buys a model-inventory tool rarely announces it. The model risk and AI governance category guide explains what the tools do and which documents they map to.