---
title: What does FINRA expect from broker-dealers using AI agents?
url: https://opencontroller.lyzr.com/blog/finra-expectations-for-ai-agents
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# What does FINRA expect from broker-dealers using AI agents?

**FINRA has not written new rules for AI agents. Its 2026 Annual Regulatory Oversight Report, released in December 2025, added a section on AI agents that names their risks (autonomy, scope creep, auditability, data handling) and lists what firms should consider: agent-specific supervision, monitoring of system and data access, human-in-the-loop points, tracking of agent actions, and guardrails. Existing rules on supervision, communications, recordkeeping and fair dealing still apply to whatever an agent does.**

That combination is the practical message for broker-dealers. The obligations are old. The evidence needed to show you meet them, when an agent rather than a person did the work, is new.

## How does FINRA define an AI agent?

The 2026 report describes AI agents as systems or programs capable of autonomously performing and completing tasks on behalf of a user. FINRA also published a blog post, "Observations on AI Agents", alongside its GenAI material. In 2024, Regulatory Notice 24-09 reminded firms that FINRA's rules are technology neutral and continue to apply when firms use generative AI.

## What risks did FINRA flag for agents?

The report lists agent-specific risks on top of familiar GenAI risks such as hallucination and bias:

- **Autonomy:** agents acting without human validation and approval.
- **Scope and authority:** agents acting beyond the user's actual or intended authority.
- **Auditability and transparency:** multi-step reasoning and actions that are hard to trace or explain.
- **Sensitive data:** agents storing, exploring, disclosing or misusing sensitive or proprietary information.
- **Domain knowledge:** general-purpose agents lacking the expertise for complex, industry-specific tasks.
- **Misaligned incentives:** reward functions that push agents toward outcomes that harm investors, firms or markets.

## Which existing rules does agent activity touch?

FINRA's report says GenAI use may implicate rules on supervision, communications, recordkeeping and fair dealing. The table maps the main ones to what an agent changes.

| Rule area | Example rule | What an agent changes | Evidence to keep |
|---|---|---|---|
| Supervision | FINRA Rule 3110 | Work is done by software, so supervisory procedures must cover agent actions | Written supervisory procedures naming agents; review records; escalation logs |
| Communications | FINRA Rule 2210 | Agents may draft or send customer communications | Content review and approval records; retained outputs |
| Books and records | FINRA Rule 4511, SEA Rule 17a-4 | Agent conversations and actions may be business records | Retained prompts, outputs and tool actions under the firm's retention policy |
| Fair dealing and best interest | Regulation Best Interest, FINRA Rule 2111 | Agent output may influence recommendations | Testing for accuracy and suitability; human review before recommendations |
| Cybersecurity and data | Regulation S-P | Agents access customer data and third-party tools | Access scopes, data-loss controls, vendor records |

## What should a supervisory programme for agents include?

FINRA's considerations translate into a concrete programme:

1. **Inventory.** Know every agent in use, including agents inside vendor products and agents built by business teams.
2. **Agent-specific written supervisory procedures.** Name who supervises each agent, what they review, and how often.
3. **Scoped access.** Give each agent its own identity and only the permissions its task needs. Avoid shared service accounts.
4. **Human-in-the-loop points.** Define which actions require a person to approve before they happen, such as sending customer communications or touching accounts.
5. **Action tracking.** Log every model call and tool call with the agent, version, user and outcome, so you can reconstruct a multi-step task.
6. **Guardrails.** Constrain what the agent can say and do at runtime, and test the guardrails.
7. **Ongoing testing.** Evaluate accuracy and adherence before release and on sampled production traffic.
8. **Stop controls.** Be able to block or quarantine an agent quickly when it misbehaves, and record that you did.

Our posts on [evidence for human review of agent outputs](/blog/evidence-for-human-review-of-ai-agent-outputs) and [AI agent controls for banks' model risk](/blog/ai-agent-controls-for-banks-model-risk) go deeper on two of these steps. For mapping across frameworks, start with [how to map AI agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks).

## What do examiners likely ask?

Expect questions that test whether the programme is real: show me the list of agents; show me the supervisory procedure for this one; show me who approved this customer communication; reconstruct what this agent did on this date; show me what happens when it breaks a rule. If any answer depends on searching several platforms by hand, that is the gap to fix first.

## How Opencontroller does this

Opencontroller helps produce evidence for these supervisory expectations; it does not make a firm compliant. It discovers agents, models and API keys across AWS, Azure and Google Cloud and connects to third-party platforms and enterprise AI workspaces, building a governed registry. Delegated on-behalf-of identity and agent workload identity federation tie actions to users and agents. The LLM Gateway and Agent Gateway sit in the traffic path for observability and cost attribution; live guardrails, LLM judges on production traffic, and per-agent Block and Quarantine controls cover runtime. Policy as code maps controls to OWASP, MITRE ATLAS, NIST and ISO 42001. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Has FINRA issued a rule specific to AI agents?**
No. FINRA treats its rules as technology neutral. The 2026 oversight report sets out observations and considerations, not new rules.

**Do agent conversations need to be retained?**
If they are business communications or records under FINRA Rule 4511 and SEA Rule 17a-4, yes. Firms should decide which agent outputs are records and retain them accordingly.

**What is the first step for a broker-dealer?**
Build an inventory of every agent in use, including those inside vendor tools, and assign each one a supervisor.

## Key facts

FINRA's 2026 Annual Regulatory Oversight Report (released December 2025) includes a section on AI agents, described as systems or programs capable of autonomously performing and completing tasks on behalf of a user. Risks named: autonomy, scope and authority, auditability and transparency, sensitive data handling, domain knowledge, misaligned reward functions. Considerations: agent-specific supervisory processes, monitoring of system and data access, human-in-the-loop points, tracking agent actions, guardrails. Regulatory Notice 24-09 (2024) says FINRA rules are technology neutral. Relevant rules include FINRA 3110, 2210, 4511 and SEA 17a-4. Opencontroller provides evidence for supervision via registry, identity, gateways, guardrails and Block/Quarantine. Licensing starts from $250K per deployment per year.

## Related

- [What evidence should you keep when a human approves an AI agent's output?](https://opencontroller.lyzr.com/blog/evidence-for-human-review-of-ai-agent-outputs.md)
- [What AI agent controls do banks need now that model risk guidance excludes agentic AI?](https://opencontroller.lyzr.com/blog/ai-agent-controls-for-banks-model-risk.md)
- [How does DORA apply to AI agents in financial services?](https://opencontroller.lyzr.com/blog/dora-and-ai-agents-in-financial-services.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
