Banks, insurers, payment firms and broker-dealers are moving AI agents from drafting text to taking actions, and an agent that can move money, change a customer record or place an order is subject to model risk, conduct and authorisation rules that were written for people and spreadsheets. This guide covers the security and control questions specific to AI agents in regulated financial services: which supervisory expectations apply (model risk management, senior manager accountability, the EU AI Act, DORA, FINRA and MAS), how to authorise and limit agent actions on payments and customer data, what to log for supervisors, and where agents fail. It is written for CISOs, IAM and fraud teams, model risk managers and second-line risk functions. Identity controls for the wider financial estate are in the Financial Services Identity Security Guide; this page goes deeper on agents. Regulatory status was checked on 27 October 2026.
Key takeaways
- An agent is a model and an actor at once. Model risk frameworks cover its reasoning; identity, authorisation and conduct rules cover what it does. Most firms govern one and forget the other.
- Supervisors are applying existing rules. The FCA says it does not plan extra AI regulations and relies on frameworks such as the Consumer Duty and the Senior Managers and Certification Regime; the PRA's SS1/23 applies to all models regardless of technology.
- FINRA has named agent-specific risks. Its 2026 Annual Regulatory Oversight Report flags agents acting without human validation, exceeding the user's intended authority, and multi-step reasoning that is hard to audit.
- Payment and trading actions need deterministic limits. Amount ceilings, beneficiary allowlists, dual control and approval steps must sit outside the model, where a prompt injection cannot talk them away.
- Credit and insurance uses are high-risk under the EU AI Act. Creditworthiness assessment and life and health insurance pricing are in Annex III, with deployer duties and a fundamental rights impact assessment; the high-risk rules for Annex III apply from 2 December 2027 as amended.
Where agents are being used and what is at stake
| Use case | Typical agent actions | Primary risk | Regulatory hook |
|---|---|---|---|
| Customer service and complaints | Read account data, answer queries, raise refunds, update contact details | Account takeover via social engineering of the agent; unfair outcomes | Consumer Duty (UK); EU AI Act Article 50(1) interaction disclosure |
| Payments operations | Prepare and release payments, match invoices, change beneficiary details | Fraudulent or mistaken payments; beneficiary change abuse | PSD2 strong customer authentication; internal dual-control policy |
| Credit decisioning support | Gather documents, summarise, recommend or score | Discrimination, unexplainable decisions | EU AI Act Annex III 5(b); model risk (SS1/23, SR 11-7) |
| Fraud and AML investigation | Triage alerts, query transaction history, draft suspicious activity reports | Missed typologies; data over-exposure to the agent | AML obligations; model validation |
| Trading and wealth | Research, draft client communications, pre-trade checks, order preparation | Unsuitable advice, unauthorised orders, record-keeping gaps | FINRA supervision and communications rules; MiFID II record keeping |
| Internal IT and security operations | Ticket handling, access requests, log queries | Privilege escalation through the agent's service identity | DORA ICT risk management |
Attackers are using agents too. In the AI agent retail card theft campaign, Gambit Security reported that an operator used open-source agents to exploit web flaws, dump cloud secrets and steal card data at scale with little human involvement (interim figures). Defences built around human attacker speed need revisiting.
Supervisory expectations that apply to agents
| Jurisdiction or body | Instrument | Status as checked | What it means for agents |
|---|---|---|---|
| UK PRA | SS1/23 Model risk management principles for banks | Effective 17 May 2024; current version published 23 April 2026 | Five principles (model identification and risk classification; governance; development, implementation and use; independent validation; risk mitigants) apply to all models regardless of technology, with a sub-principle on AI used in modelling. Formally scoped to banks with internal model approval, widely used as a benchmark |
| UK FCA | AI approach page; Consumer Duty; SM&CR | No AI-specific rules planned (page updated 2 October 2026) | A named senior manager is accountable for the agent's outcomes; customer outcomes must meet the Consumer Duty |
| US Federal Reserve and OCC | SR 11-7 Guidance on Model Risk Management (2011) | Check current supervisory status before relying on it | Model inventory, independent validation, ongoing monitoring and outcomes analysis, effective challenge |
| US FINRA | 2026 Annual Regulatory Oversight Report, GenAI section | Published December 2025 | Consider whether agent autonomy needs tailored supervisory processes; monitor agent access and data handling; decide where human-in-the-loop applies; track agent actions; set guardrails |
| EU | AI Act (Regulation (EU) 2024/1689) | In force; Annex III high-risk rules from 2 December 2027 as amended by the Digital Omnibus on AI | Credit scoring and insurance pricing are high-risk; financial institutions meet the deployer monitoring and log-keeping duties through their financial services governance rules (Article 26(5) and (6)); FRIA under Article 27 |
| EU | DORA | Applies to EU financial entities | AI providers and model APIs are ICT third-party services: register, contract, test and report incidents. See the DORA ICT Third-Party Risk and Register of Information Guide |
| Singapore MAS | Consultation Paper P017-2025 on Guidelines on AI Risk Management | Consultation closed 31 January 2026; MAS response published 7 October 2026 | Expectations on oversight of AI risk, risk management systems and policies, lifecycle controls and capabilities, for all financial institutions |
For the EU role question (are you the provider or deployer of an agent built on a vendor model?), see the EU AI Act Provider vs Deployer Obligations Guide. For model governance in depth, see the AI Model Risk Management Guide.
Authorising agent actions on money and customer data
The control principle is that the model proposes and deterministic systems decide. An agent's identity and tools should make unsafe actions impossible, not merely discouraged in a system prompt.
- Dedicated identity per agent: each agent has its own workload identity and credentials, registered with an owner, never a shared operations account (NIST SP 800-53 AC-2, IA-9). See the Agent Identity and Access Guide.
- On-behalf-of authority: when an agent acts for a customer or employee, carry that person's identity and entitlements through delegated tokens, so the agent can never exceed the authority of the person it serves. FINRA's concern about agents exceeding the user's intended authority is this control. See the Delegation Chains and Attenuated Authority for AI Agents.
- Hard transaction limits: per-transaction and daily value ceilings, currency and corridor restrictions and beneficiary allowlists enforced by the payment system, not the agent. See the AI Agent Action Approvals and Spending Limits Guide.
- Separation of duties: an agent that prepares a payment must not release it, and an agent must never both change beneficiary details and pay the new beneficiary (AC-5).
- Customer authentication stays with the customer: agent-initiated payments must not bypass strong customer authentication; the customer's authentication and dynamic linking requirements still apply. See the PSD2/PSD3 Strong Customer Authentication Guide and the Agent Payments Protocols Guide.
- Minimum data: agents query customer data through scoped APIs returning only the fields the task needs, not direct database access.
| Action | Autonomy allowed | Control |
|---|---|---|
| Answer a balance or transaction query for an authenticated customer | Autonomous | On-behalf-of token, read-only scope |
| Issue a refund under [threshold] | Autonomous within limit | Value ceiling, daily cap, reason code, sampling review |
| Change contact details, phone number or email | Human approval or step-up authentication of the customer | High account takeover risk; out-of-band confirmation |
| Add or change a payment beneficiary | Never autonomous | Customer SCA plus cooling-off or confirmation of payee |
| Release an outbound payment | Human release, or dual control | Maker-checker with the agent as maker only |
| Make or recommend a credit decision | Recommendation only | Named credit officer decides; reasons recorded |
| Place a client order | Prepare only | Licensed person approves; pre-trade checks in the order system |
Inbound agents acting for customers
The reverse case is growing: customers' own AI agents and browser agents logging in to online banking or calling payment APIs on their behalf. Treat these as a distinct channel rather than either a human or a bot to be blocked.
- Prefer delegated API access over screen scraping: an agent that holds a scoped, revocable OAuth grant through an open banking or partner API is far safer than one that stores the customer's password and drives the web interface.
- Distinguish agent traffic: classify declared agents, undeclared automation and humans separately, and apply step-up authentication when an agent session attempts a high-risk action. See the Bot vs AI Agent Traffic Management Guide.
- Keep liability clear: terms and customer communications should state what an authorised agent may do and how the customer revokes it, because authorised push payment fraud claims will test this. See the Authorised Push Payment (APP) Fraud Guide.
Model risk management for agents
- Inventory the agent as a model: record the base model and version, prompts, tools, retrieval sources and the decision it supports, with a risk tier (SS1/23 principle 1).
- Validate the system, not just the model: independent validation should test the full agent with its tools and data, including adversarial tests for prompt injection and goal hijacking. See the AI Agent Security Testing Checklist.
- Treat vendor model updates as model changes: pin model versions where the provider allows it, and re-validate before a new version reaches production.
- Ongoing monitoring and outcomes analysis: SR 11-7's monitoring and outcomes analysis map to tracking agent decision quality, override rates and complaint outcomes.
- Effective challenge: second-line reviewers need access to traces and the authority to block deployment.
Records, audit and evidence for supervisors
- Action log tied to identity: for every action record the agent identity, the person on whose behalf it acted, the tool, parameters, result and approval reference (AU-12 Audit Record Generation; AU-10 Non-repudiation).
- Reasoning trace retention: keep the prompts, retrieved data references and outputs behind customer-affecting decisions for the period your record-keeping rules require, and at least six months for EU high-risk deployments under Article 26(6), which financial institutions fulfil through their financial services documentation.
- Communications records: client communications drafted or sent by agents are business records for broker-dealer and investment firm rules.
- Explainability for complaints: be able to reconstruct why the agent took an action for an individual customer.
See the AI Agent Audit Trails for Regulators Guide.
How to implement
- Register agents: list every agent, its owner, accountable senior manager, model, tools, data and action rights.
- Classify by action risk: read-only, reversible write, irreversible or financial action, regulated decision.
- Engineer controls outside the model: limits, allowlists, maker-checker and step-up authentication in the core systems.
- Pilot in shadow mode: the agent proposes actions that humans execute for [4 to 8 weeks]; compare its proposals with human decisions and measure error and override rates.
- Graduate autonomy by action class: enable autonomous actions only for the lowest-risk class first, with value caps.
- Prepare rollback and kill switch: one control to revoke the agent's tokens and return the workflow to manual handling, tested. See the AI Agent Kill Switch and Containment Guide.
- Bring in third-party risk: treat the model and agent platform providers as critical ICT suppliers where they support important functions.
Decision guidance
- Build or buy: a vendor agent inside a core banking or CRM platform shifts development work to the vendor but not accountability; you still need validation evidence, logs and limits.
- Smaller firms: keep agents read-only or proposal-only for customer money; the control cost of safe autonomous payments is high.
- Large groups: standardise a single agent gateway that enforces identity, tool allowlists, limits and logging for every business line.
Detection and measurement
- Monitor: agent actions against baseline (volume, value, time, customer segments) (DE.CM-09); tool calls outside the agent's registered set; refused or failed actions, which can indicate manipulation attempts; data volume returned per session.
- Correlate with fraud signals: beneficiary changes, contact detail changes and refunds initiated through agent channels feed the fraud engine like any other channel. See the Fraud and Cyber Convergence Guide.
| Metric | Why it matters |
|---|---|
| Agents with a named owner, accountable senior manager and model risk tier | Accountability and inventory completeness |
| Agent actions blocked by deterministic limits per 1,000 actions | Shows limits are working and where agents push against them |
| Human override rate on agent proposals | Decision quality and drift |
| Customer complaints attributed to agent handling | Conduct outcome measure |
| Agents validated after the most recent model version change | Change control discipline |
| Time to revoke an agent's access in an exercise | Containment readiness |
Framework mapping
| Framework | Reference | Agent control |
|---|---|---|
| NIST SP 800-53 Rev. 5 | AC-2, AC-5, AC-6, IA-9, AU-10, AU-12, CA-7 | Agent accounts, separation of duties, least privilege, service authentication, non-repudiable logs, monitoring |
| NIST CSF 2.0 | GV.RR-02, PR.AA-05, DE.CM-03, DE.CM-09 | Accountability, scoped permissions, activity monitoring |
| PRA SS1/23 | Principles 1 to 5 | Inventory, governance, development and use, validation, mitigants |
| EU AI Act | Article 26, Article 27, Annex III point 5 | Deployer duties, FRIA, high-risk credit and insurance uses |
| ISO/IEC 42001:2023 | AI management system | Governance of agent development and use |
Common failure patterns
- Limits in the prompt: "never pay more than [amount]" written into instructions instead of enforced in the payment system.
- Shared service account: several agents run under one privileged operations identity, so no action can be attributed and one compromise reaches everything.
- Agent as account takeover helper: a support agent changes phone numbers or email addresses on request, giving attackers the recovery channel. The Meta AI support assistant account takeover shows the pattern outside banking.
- Model risk without identity: the agent passes validation but holds broad API rights nobody reviews.
- Silent vendor model changes: the provider updates the model and agent behaviour shifts with no re-validation.
- No trace for complaints: the firm cannot explain to a customer or ombudsman why the agent acted.
Practitioner checklist
- Every agent is in the model inventory and the identity registry, with an owner and accountable senior manager.
- Agents use dedicated identities and on-behalf-of tokens bounded by the user's own entitlements.
- Payment, beneficiary, contact-detail and credit actions have deterministic limits and human or dual control outside the model.
- Strong customer authentication is never bypassed for agent-initiated payments.
- Independent validation covers the full agent, including adversarial testing.
- Model version changes trigger re-validation before production.
- Action logs tie each action to agent, principal, tool and approval, with retention matching record-keeping rules.
- Agent actions feed fraud monitoring and anomaly detection.
- AI providers are assessed and registered as ICT third parties where they support important functions.
- A tested kill switch returns each agent workflow to manual handling.
Standards and references
- Bank of England PRA: SS1/23 Model risk management principles for banks
- Federal Reserve: SR 11-7 Guidance on Model Risk Management
- FINRA: 2026 Annual Regulatory Oversight Report, GenAI
- FCA: Our approach to AI
- MAS: Consultation Paper on Guidelines on Artificial Intelligence Risk Management
- European Commission AI Act Service Desk: Article 26 Obligations of deployers
- NIST: SP 800-53 Rev. 5
Related NHI Mgmt Group resources: Financial Services Identity Security Guide · Financial Services Cyber Regulation Map · AI Agent Authorisation Guide · DORA Guide