Financial Regulators and AI: what the SEC, FINRA, FCA, PRA and EU supervisors expect

Financial regulators have mostly chosen not to write new AI rules, which means banks, brokers and asset managers must show that existing supervision, model risk, conduct and record-keeping rules already cover their AI models and agents. This guide sets out what the US Securities and Exchange Commission (SEC), the Financial Industry Regulatory Authority (FINRA), the UK Financial Conduct Authority (FCA) and Prudential Regulation Authority (PRA), and EU financial supervisors under the AI Act expect, with dates and status checked against each regulator's own pages on 27 October 2026. It focuses on supervisory expectations that touch security and identity: who an AI agent acts as, what it may do, and how its actions are recorded. It is written for compliance, model risk, security and technology leaders at regulated firms. For the model risk discipline itself, see the AI Model Risk Management Guide; this page covers what the regulators say.

Key takeaways

• Existing rules apply. FINRA calls its rules technology neutral, the FCA states it does not plan to introduce extra regulations for AI, and the PRA folds AI into its model risk principles.
• The SEC pulled its AI-specific proposal. The 2023 predictive data analytics proposal was withdrawn with effect from 17 June 2025; examiners now test AI claims and AI supervision under existing rules.
• Agents are on the supervisory agenda. FINRA's 2026 oversight report names agents acting without human validation and beyond their intended scope and authority as specific risks.
• In the EU, financial supervisors enforce the AI Act for financial services. Article 74(6) makes the national financial supervisor the market surveillance authority for high-risk AI used in direct connection with financial services.
• Identity is the control regulators can test. Every expectation (supervision, accountability, logging, scope of authority) depends on agents having distinct, owned, least-privilege identities with auditable actions.

Scope: who and what is covered

For the wider cyber regulation picture by jurisdiction, see the Financial Services Cyber Regulation Map.

Key obligations

United States: SEC

• Accurate AI claims: on 18 March 2024 the SEC settled charges against two advisers, Delphia (USA) Inc. and Global Predictions Inc., for false and misleading statements about their use of AI, with civil penalties of US$225,000 and US$175,000. Both consented without admitting or denying the findings. The charges were brought under the Marketing Rule and the Advisers Act: no AI-specific rule was needed.
• Examination focus: the fiscal year 2026 priorities say the Division of Examinations will review the accuracy of registrants' representations about AI capabilities and whether firms have policies to monitor and supervise AI used in fraud prevention, back-office operations, anti-money laundering and trading. Automated investment tools are reviewed for whether advice fits investor profiles and matches disclosures.
• Customer data: the same priorities cover Regulation S-P incident response programmes and vendor oversight, and Regulation S-ID identity theft programmes, both of which apply when AI tools process customer information. See the SEC Regulation S-P Amendments Guide.

United States: FINRA

• Supervision (Rule 3110): where generative AI is used in a supervisory system, for example reviewing electronic correspondence, policies should address model risk management, data privacy and integrity, and reliability and accuracy.
• Communications (Rule 2210): content standards apply whether a person or a tool generated the communication.
• Agents: the 2026 oversight report lists agent risks (acting without human validation, acting beyond intended scope and authority, hard-to-trace multi-step reasoning, mishandling sensitive data) and points firms to monitoring agent system access and data handling, deciding where human in the loop oversight is required, tracking agent actions and decisions, and guardrails that limit agent behaviours.
• Logging: the report's effective practices include logging prompts and outputs and tracking model versions.

United Kingdom: FCA and PRA

• FCA, outcomes not new rules: the FCA relies on existing frameworks and a principles-based approach. The Consumer Duty applies to AI-driven products, communications and support, and SM&CR means a named senior manager is accountable for the business area where AI is used.
• PRA SS1/23: five principles: model identification and model risk classification; governance; model development, implementation and use; independent model validation; and model risk mitigants. A sub-principle covers the risks of AI in modelling techniques such as machine learning. Vendor models are in scope. SS1/23 was published on 17 May 2023, took effect on 17 May 2024 and was re-issued on 23 April 2026 with low-impact amendments.

European Union: AI Act in financial services

• Who supervises: under Article 74(6), for high-risk AI placed on the market, put into service or used by regulated financial institutions in direct connection with financial services, the national financial supervisor is the market surveillance authority. Under Article 74(7), authorities in the Single Supervisory Mechanism report relevant findings to the European Central Bank without delay.
• Deployer duties adapted for finance: Article 26(5) treats a financial institution's monitoring duty as met when it complies with its internal governance rules under financial services law, and Article 26(6) lets it keep AI system logs within the records it already maintains under that law (the general minimum is six months).
• High-risk uses in finance: Annex III includes creditworthiness assessment and credit scoring of natural persons and risk assessment and pricing for life and health insurance. See the EU AI Act High-Risk AI Systems Guide.
• ICT risk: AI services are ICT services, so DORA's ICT risk, incident reporting and third-party register obligations apply to them. See the DORA Guide and the EBA ICT and Security Risk Guidelines Guide.

Timeline and current status

Dates can move: the EU high-risk dates were already pushed back once. Recheck the AI Act consolidated text and each regulator's page before relying on a date. See the EU AI Act Implementation Timeline Tracker.

Enforcement and penalties

• SEC: AI misrepresentation is enforced under existing anti-fraud and Marketing Rule provisions: censure, cease and desist orders and civil penalties, as in the March 2024 cases.
• FINRA: failures in AI-assisted supervision or communications are failures under Rule 3110 or Rule 2210 and are disciplined as such.
• FCA and PRA: breaches of the Consumer Duty, Principles or SM&CR duties carry the usual enforcement tools, and a senior manager can be held personally accountable.
• EU: AI Act penalties are set by Member States within the Act's maximums, and financial supervisors act as market surveillance authorities for financial services uses. See the EU AI Act Compliance Guide for the penalty tiers.

What it means for AI agents and non-human identities

None of these regulators has written an identity rule for agents, but each expectation becomes testable only through identity and access controls. Map them like this:

See the AI Agents in Financial Services Security Guide, the AI Agent Action Approvals and Spending Limits Guide and the AI Agent Audit Trails for Regulators Guide.

Compliance steps

1. Inventory: list every AI model, generative AI tool and agent, in-house and embedded in vendor products (FINRA's notice covers both), with business owner, accountable senior manager and the identities each uses.
2. Classify: apply your model risk tiering (PRA SS1/23 principle 1 if in scope); flag EU uses that fall in Annex III, such as credit scoring.
3. Map rules to uses: for each use, record which existing rules apply: supervision, communications, Consumer Duty, Marketing Rule, Reg S-P, DORA, AI Act.
4. Control agent authority: give each agent its own identity, scope its permissions to the task, require human approval above set thresholds, and block actions outside its mandate.
5. Log and retain: log prompts, outputs, tool calls and model versions under your record-keeping controls, attributable to the agent and the human principal.
6. Validate and test: independent validation for models in scope, adversarial and accuracy testing for generative AI and agents.
7. Check claims: compare marketing and disclosures about AI against the inventory before publication.
8. Report and review: report AI risk to the board and accountable senior managers, and review after each regulatory update.

Small firms versus large groups

• Small firms: often use AI through vendors. Focus on vendor due diligence, contract terms on data use and logging, and supervising outputs under existing procedures.
• Large groups: need one model and agent inventory across jurisdictions, mapped to each regulator's expectations, with the strictest requirement set as the group baseline.

Measuring it

• Inventory coverage: percentage of AI uses (including vendor-embedded) in the inventory with an accountable owner.
• Agent identity hygiene: percentage of agents with a unique identity and task-scoped permissions.
• Traceability: percentage of agent actions in logs attributable to agent, human principal and model version.
• Validation currency: share of in-scope models validated within the cycle set by your model risk policy.
• Claims review: number of AI-related marketing statements checked against the inventory before release.

Common failure patterns

• Waiting for AI rules. Firms delay controls because no AI-specific rule exists, while examiners already test AI under supervision, conduct and marketing rules.
• Missing vendor-embedded AI. The inventory covers in-house models but not generative AI features switched on inside CRM, email or surveillance products, which FINRA's notice explicitly includes.
• Agents on shared or human credentials. An agent that runs on an employee's account or a shared service account cannot show it stayed within its intended scope and authority, and its actions cannot be supervised separately.
• Logs outside regulated records. Prompt and tool-call logs sit in a developer platform with short retention, outside the records the firm must keep.
• Marketing ahead of capability. Claims about AI-driven advice or forecasting that the systems cannot support, the exact pattern in the SEC's March 2024 cases.
• No named owner. Under SM&CR and similar regimes, an AI use without an accountable senior manager is a governance gap in itself.

Practitioner checklist

• Build one inventory of AI models, tools and agents, including vendor-embedded AI.
• Assign an accountable owner and senior manager to each.
• Map each use to the existing rules that apply in each jurisdiction.
• Give every agent its own least-privilege identity; ban shared service accounts.
• Define which agent actions need human approval and record approvers.
• Log prompts, outputs, tool calls and model versions within regulated records.
• Validate models in scope of SS1/23 or your model risk policy.
• Check AI claims in marketing against what systems actually do.
• Track EU Annex III uses against the 2 December 2027 date.
• Recheck regulator pages each quarter for new guidance.

Standards and references

• FINRA: Regulatory Notice 24-09, Generative AI and Large Language Models
• FINRA: 2026 Annual Regulatory Oversight Report, GenAI
• SEC: Division of Examinations Fiscal Year 2026 Priorities
• SEC: Withdrawal of Proposed Regulatory Actions (S7-12-23)
• SEC: Charges against two investment advisers for misleading statements about AI
• Bank of England PRA: SS1/23 Model risk management principles for banks
• FCA: Our approach to AI
• European Commission AI Act Service Desk: Article 74

Related NHI Mgmt Group resources: AI Model Risk Management Guide · AI Agents in Financial Services Security Guide · Financial Services Identity Security Guide · Financial Services Cyber Regulation Map. This guide summarises supervisory expectations of the SEC, FINRA, FCA, PRA and EU supervisors on AI as checked on 27 October 2026 and is not legal advice.