Regulators are not writing new rulebooks for AI agents in banking. They are telling banks and insurers that existing duties apply, and that the more an AI system can act on its own, the stronger the controls, human oversight and accountability around it must be. The clearest example is the US OCC’s Bulletin 2026-54, issued on 4 November 2026: an examiner framework for generative and agentic AI whose illustrative tiering puts an autonomous payment agent in the top risk tier regardless of how well it tests. Singapore, Australia, Canada, the UK and India are moving in the same direction.
This analysis is for boards, chief risk officers, CISOs and model risk teams at banks and insurers piloting or running AI agents.
What the OCC published on 4 November 2026
OCC Bulletin 2026-54, “Super Intelligence: Supervisory Framework”, is addressed to national banks, federal savings associations and federal branches and agencies. Citing Executive Order 14434, the OCC uses “super intelligence” (SI) for the technologies previously called artificial intelligence. In practice, the framework is about generative AI and AI agents.
Four points matter most:
- It is an examiner resource, not a new rule. It does not establish supervisory expectations or enforceable standards, but it shows what examiners will look at.
- AI is not a separate risk category. Banks do not need a separate SI risk appetite; AI risk sits inside existing operational, compliance and credit risk frameworks.
- Consequence beats sophistication. Control intensity could increase with the potential consequences of an SI agent’s failure or misuse. A simple rules-based agent authorised to initiate transactions could present greater risk than a sophisticated model limited to internal drafting.
- It fills a gap. The April 2026 interagency model risk guidance (OCC Bulletin 2026-13) left generative and agentic AI out of scope; this framework covers them.
The framework groups 16 components under six pillars: Purpose, Risk, Control, Oversight, Resilience and Assurance. A bank may document what each agent is permitted to do, what it can access and which actions need human review; activity outside that purpose may be treated as a control failure “even when the resulting output or action appears reasonable”. The resilience section addresses an agent that “could remain available while taking inappropriate or harmful actions”, listing circuit breakers, rate limits, least-privilege permissions and kill switches as containment tools.
Human in the loop versus human on the loop
The OCC separates two kinds of oversight. Human in the loop means review and approval before an output or action is used. Human on the loop means active monitoring with the authority and practical ability to intervene. The framework warns that approval should be “more than a nominal click”, that bulk approvals cause alert fatigue, and that reviewers need the expertise and authority to reject, pause or halt an action.
The OCC’s illustrative risk tiers
Use cases are scored on decision consequence, reversibility, customer and financial impact, autonomy, data sensitivity, explainability and provenance, then combined into a composite tier:
| Tier | Profile | Illustrative use cases |
|---|---|---|
| Tier 1: Critical | Direct action, low or no reversibility, high impact, autonomous or partially autonomous | Autonomous payment agent, real-time credit decisioning, fraud-hold release, market access |
| Tier 2: High | Direct action or high-weight decision support, human on the loop | AI-assisted underwriting, agent-initiated customer communications, AML alert disposition |
| Tier 3: Moderate | Decision support with mandatory human in the loop, reversible | Summarising documents for credit memos, agent-assisted research |
| Tier 4: Low | No customer or financial data exposure, no direct action | Internal productivity tools, general employee Q&A |
How other regulators treat AI that takes action
Singapore: MAS Guidelines on AI Risk Management
MAS issued its final Guidelines on Artificial Intelligence Risk Management on 7 October 2026. They take effect on 7 October 2027, with Sections 3 and 4 (oversight and risk management systems) expected from then and Sections 5 and 6 (life cycle controls and capability) by 7 October 2028. They apply to all financial institutions, including insurers, and explicitly cover AI agents. MAS warns that an agent with tool access could “autonomously execute unauthorised or erroneous actions”, suggests inventories record agent identifiers, tool access and guardrails, and expects human oversight proportionate to risk materiality. It plans to consult on further agentic AI guidance in 2027.
Australia: APRA’s letter to industry
On 30 April 2026, APRA wrote to all regulated entities, including banks, insurers and superannuation trustees, after finding that governance and assurance were not keeping pace with AI adoption. It created no new standard. Its expectations include an inventory of AI tools and use cases, human involvement in high-risk decisions, staff training, controls over “agentic and autonomous workflows” and identity management for “nonhuman actors such as AI agents”.
Canada: OSFI Guideline E-23
OSFI’s revised Guideline E-23 on model risk management takes effect on 1 May 2027 for all federally regulated financial institutions, including banks and insurers. It explicitly covers AI and machine learning models, including third-party models. A July 2026 OSFI technology risk bulletin on generative and agentic AI adds sound practices institutions can consider: limits on autonomy, approval checkpoints for high-risk actions, logging of agent activity, short-lived credentials and manual fallbacks.
United Kingdom: FCA and PRA
The FCA published “AI in financial services: good practice and areas for improvement” on 19 October 2026. It creates no new requirements, relying on the Consumer Duty, the Senior Managers and Certification Regime and systems and controls rules. It suggests a central inventory of agents showing owners, access and controls, and warns that human review can become “box-ticking”. In a speech the same day, the FCA said it would work with industry, regulators and frontier labs on a framework for agentic finance covering agent identity, delegated authority, accountability and execution. The PRA applies its model risk principles (SS1/23) to AI and, in April 2026, said it keeps its technology-agnostic approach under review.
India: RBI FREE-AI and draft model risk guidance
The RBI published its FREE-AI committee report on 13 August 2025, with seven principles (“sutras”) and 26 recommendations. It is a committee report, not a binding direction. On 24 June 2026, the RBI issued draft model risk management guidance covering AI and machine learning models, including third-party models, at banks, NBFCs and other regulated entities. The draft refers to human-in-the-loop and human-on-the-loop arrangements and “kill-switch arrangements”. Comments closed on 24 July 2026; we found no final version at the time of writing.
Comparison table: AI agents in banking regulation
| Regulator | Instrument and date | Status | Covers agents explicitly? | Human oversight emphasis |
|---|---|---|---|---|
| OCC (US) | Bulletin 2026-54, 4 November 2026 | Examiner resource, no new expectations | Yes, SI agents and agentic SI | Human in or on the loop, calibrated to consequence |
| MAS (Singapore) | AI Risk Management Guidelines, 7 October 2026 | Supervisory guidelines, effective 7 October 2027 | Yes, AI agents and multi-agent systems | Proportionate to risk materiality |
| APRA (Australia) | Letter to industry, 30 April 2026 | Supervisory expectations under existing standards | Yes, agentic and autonomous workflows | Human involvement in high-risk decisions |
| OSFI (Canada) | Guideline E-23, effective 1 May 2027; July 2026 bulletin | Guideline; bulletin is sound practice | Bulletin yes; E-23 covers AI models | Approval checkpoints for high-risk actions |
| FCA and PRA (UK) | FCA good practice, 19 October 2026; SS1/23 | No new rules; existing regimes apply | Yes, agent inventories and governance | Risk-based, wary of box-ticking review |
| RBI (India) | FREE-AI report, 13 August 2025; draft guidance, 24 June 2026 | Report and draft, not final | Not by name; covers autonomy of AI models | Human in or on the loop, kill switches |
What boards and risk teams should do now: a checklist
- Inventory every agentic system. Include vendor features and staff-built agents, with owner, purpose, access, permitted actions and provider.
- Assign autonomy tiers. Score each use case on consequence, reversibility, impact and autonomy. Anything that moves money, changes customer records or communicates with customers on its own belongs at the top.
- Define human approval points. Decide which actions need approval before execution and which only need monitoring. Check that reviewers have the time, skills and authority to say no.
- Extend model risk management. Adapt validation to non-deterministic outputs, frequent vendor and prompt changes, and agent tool use. Plan for E-23, MAS and RBI timelines where they apply.
- Tighten third-party risk. Require notice when vendors add AI features or change models, clarify whether your data trains their models and test exit options. Accountability stays with you.
- Log everything material. Keep inputs, outputs, tool calls, approvals and overrides so a decision can be reconstructed. Give each agent its own identity and least-privilege credentials.
- Prepare AI incident response. Define what counts as an AI incident (misfire, hallucination, drift, manipulation), set escalation thresholds, test kill switches and fallbacks, and plan customer remediation.
- Train people by role. Every regulator above stresses people as much as technology. Boards need enough AI literacy to challenge management (see AI Risk & Strategy for Leaders); risk, compliance and audit teams need to assess AI use cases (AI Governance & Compliance training); engineers need agent security skills (Securing AI Agents training).
For tooling that discovers agents and enforces guardrails at runtime, our partner AccuroAI offers AI agent security.
Frequently asked questions
Does OCC Bulletin 2026-54 create new rules for banks?
No. The OCC says it is a resource for examiners and does not establish supervisory expectations or enforceable standards. It does show how examiners will assess generative and agentic AI, so banks can use it to test their controls.
What does “super intelligence” mean in the OCC bulletin?
It is the OCC’s term, following Executive Order 14434, for technologies previously called artificial intelligence. It is not a new class of technology.
Which AI uses do regulators see as highest risk?
Systems that take direct, hard-to-reverse actions with customer or financial impact and little real-time human control, such as the OCC’s examples of autonomous payment agents and fraud-hold release.
Do these frameworks apply to insurers?
The OCC bulletin covers banks only. MAS’s guidelines apply to all financial institutions, APRA’s letter covers insurers, OSFI’s E-23 covers insurers, and the FCA’s paper applies to all regulated firms.
Is a human approving every action enough?
Not on its own. The OCC, FCA and MAS all warn about automation bias and nominal approvals. Reviewers need competence, authority and realistic workloads.
This article is general information, not legal or regulatory advice.
Get your teams ready
Product Innovation Academy runs corporate AI governance and AI security programmes for financial institutions, with AI Security & Governance programmes delivered in partnership with AccuroAI. Request a proposal and we will tailor training to your agents, regulators and risk profile.
Sources
- OCC, Bulletin 2026-54, “Super Intelligence: Supervisory Framework”, 4 November 2026. occ.gov
- OCC, “OCC Supervisory Framework on Bank Use of Super Intelligence” (framework document), 4 November 2026. occ.gov
- OCC, Bulletin 2026-13, “Model Risk Management: Revised Guidance”, 17 April 2026. occ.gov
- MAS, “MAS Sets Out Supervisory Expectations on Responsible AI Adoption by Financial Institutions”, 7 October 2026. mas.gov.sg
- MAS, “Guidelines on Artificial Intelligence Risk Management”, 7 October 2026. mas.gov.sg
- APRA, “APRA Letter to Industry on Artificial Intelligence (AI)”, 30 April 2026. apra.gov.au
- OSFI, “Guideline E-23: Model Risk Management (2027)” letter, 11 September 2025. osfi-bsif.gc.ca
- OSFI, “Generative and Agentic Artificial Intelligence: Implications for Technology, Cyber Security, and Operational Resilience”, July 2026. osfi-bsif.gc.ca
- FCA, “AI in financial services: good practice and areas for improvement”, 19 October 2026. fca.org.uk
- FCA, “Building the future of finance in the age of AI” (speech), 19 October 2026. fca.org.uk
- Bank of England and PRA, “The Bank and the PRA’s response to HMT, DSIT and DBT on AI in financial services”, 1 April 2026. bankofengland.co.uk
- RBI, “Report of the Committee to develop a Framework for Responsible and Ethical Enablement of Artificial Intelligence (FREE-AI) in the Financial Sector”, press release, 13 August 2025. rbi.org.in
- RBI, “RBI issues draft Guidance on Regulatory Principles for Model Risk Management”, 24 June 2026. rbi.org.in