Reference · Securities

FINRA 2026 Oversight Report: GenAI and AI Agent Expectations

How FINRA's existing rules apply to generative AI and AI agents, and the practices the report describes

Last reviewed November 4, 2026

FINRA published the 2026 FINRA Annual Regulatory Oversight Report on 9 December 2025.FINRA release It adds a topic area on generative AI, "GenAI: Continuing and Emerging Trends," marked new for 2026.FINRA 2026 Report · Introduction The topic covers AI agents as well as GenAI tools. FINRA's survey found firms starting with efficiency gains in internal processes and information retrieval, and summarization and information extraction as the most common use case among member firms. The report's list of use case types also includes conversational AI and question answering, content generation and drafting, translation and coding.FINRA 2026 Report, GenAI · How Member Firms Use GenAI These are the large language model (LLM) uses that Notice 24-09 addressed by name, among them summarizing large and complex documents.Notice 24-09

The report "does not create any new legal or regulatory requirements or new interpretations of existing requirements." FINRA also tells readers not to treat its practices as required beyond what existing rules demand.FINRA 2026 Report · Introduction What it gives is FINRA's view of how those rules reach generative AI and LLMs, and the practices it has seen firms adopt.

FINRA's GenAI and LLM guidance spans two notices and two oversight reports

Regulatory Notice 24-09, issued 27 June 2024, is FINRA's reminder of member firms' obligations when they use generative AI and large language models. It creates no new requirements. FINRA's rules are technology-neutral and apply whether a firm builds a GenAI tool, uses a third party's, or uses GenAI features embedded in a product it already runs.Notice 24-09

The 2025 Annual Regulatory Oversight Report, published 28 January 2025, took up GenAI inside its third-party risk topic. FINRA observed firms proceeding cautiously, mostly with vendor-supported GenAI tools for internal efficiency, and listed practices such as asking vendors whether their products incorporate GenAI and barring firm or customer sensitive information from a vendor's open-source GenAI tool.FINRA 2025 Report · Third-Party Risk

The 2026 report, published 9 December 2025, gives GenAI a topic of its own and adds AI agents: six risks particular to agents and four supervisory considerations, covering system access and data handling, human-in-the-loop oversight, tracking of actions and decisions, and guardrails.FINRA 2026 Report, GenAI · Agents Its third-party risk topic names LLMs directly, listing a supervision, governance or model risk management framework for "AI/LLM development, implementation, use and monitoring" among the practices FINRA has observed.FINRA 2026 Report, Third-Party Risk · Technology Management

In Regulatory Notice 26-14, issued 9 July 2026, FINRA proposed modernizing its communications rule, citing advances in GenAI among its reasons, and stated that members are responsible for their communications whether a person or AI technology generates them.Notice 26-14 FINRA's Artificial Intelligence key topics page collects these statements and its other AI resources, and repeats the position they share: the rules continue to apply when member firms use GenAI, whether they develop it in house or source it from a third party.FINRA AI key topics

Existing rules already cover GenAI

FINRA repeats the position of Regulatory Notice 24-09: its rules are technology-neutral and "continue to apply when firms use GenAI," so supervision, communications, recordkeeping and fair dealing cover GenAI as they cover any other tool.FINRA 2026 Report, GenAI · Regulatory Obligations Notice 24-09

Supervision is the rule the report cites by number. Rule 3110 requires each member firm to keep a supervisory system "reasonably designed" to achieve compliance with the securities laws and FINRA rules.Rule 3110 · (a) The report adds that the system must be tailored to the firm's business, and that a firm relying on GenAI inside its supervisory system may consider the integrity, reliability and accuracy of the model in its procedures.FINRA 2026 Report, GenAI · Regulatory Obligations

Recordkeeping is one of those rules, so retention periods for AI-generated records, including LLM prompts and outputs that are required records, come from the books and records rules, as for any other record. The SEC recordkeeping entry covers which AI records Rule 17a-4 reaches.

Vendor tools are in scope

The GenAI section addresses FINRA member firms and draws no line between a tool a firm builds and one a vendor supplies. Its cybersecurity practice covers the risks of both the firm's and its third-party vendors' use of GenAI.FINRA 2026 Report, GenAI · Current Practices Notice 24-09 says the same of the rules themselves: they apply whether a firm develops a GenAI tool or uses a third party's, including features embedded in products it already runs.Notice 24-09

For supervising vendors, the section's resources list Regulatory Notice 21-29, FINRA's reminder of a firm's supervisory obligations when it outsources to third-party vendors.FINRA 2026 Report, GenAI · Additional Resources Notice 21-29 FINRA also notes that some of the report's content may not be relevant to a given firm's business model, size or practices.FINRA 2026 Report · Introduction

Approve, test and monitor each GenAI use

FINRA offers these as practices a firm contemplating GenAI may want to consider.FINRA 2026 Report, GenAI · Current Practices

Run GenAI as an enterprise program. Supervisory processes for developing and using GenAI across the firm, ways to find and reduce hallucination and bias, and a cybersecurity program that accounts for GenAI risk at the firm and its vendors.
Approve each use case formally. Review by business and technology experts of each opportunity and the controls it needs, inside a supervision, governance or model risk management framework that keeps "comprehensive documentation throughout."
Test the model. Testing that establishes the model's capabilities, limitations and performance, in areas such as privacy, integrity, reliability and accuracy.
Monitor it in use. Ongoing monitoring of prompts, responses and outputs to confirm the tool performs as expected and behaves compliantly. FINRA's examples are prompt and output logs kept for accountability and troubleshooting, a record of which model version was used and when, and human review of outputs with regular checks for errors or bias.FINRA 2026 Report, GenAI · Current Practices

Agents carry six risks of their own

FINRA describes an AI agent as a system that can perform and complete tasks on a user's behalf, planning, deciding and acting toward a goal without predefined rules. Bias, hallucination and privacy risk carry over to agents and their outputs, and FINRA names six more.FINRA 2026 Report, GenAI · Agents

Autonomy: the agent acts without a person validating or approving the action.
Scope and authority: it acts beyond what the user actually intended or authorized.
Auditability and transparency: multi-step reasoning makes outcomes hard to trace or explain.
Data sensitivity: working on sensitive data, it may store, explore, disclose or misuse sensitive or proprietary information without meaning to.
Domain knowledge: a general-purpose agent may lack the knowledge to carry out complex, industry-specific tasks consistently.
Rewards and reinforcement: a misaligned or poorly designed reward function could steer it toward decisions that hurt investors, firms or markets.

Agents need supervision of their own

FINRA suggests agents may call for supervisory processes "specific to the type and scope of the AI agent being implemented." It names four things to consider: how to monitor the agent's system access and data handling, where to place human-in-the-loop oversight, how to track its actions and decisions, and what guardrails limit what it may do.FINRA 2026 Report, GenAI · Agents

Flow

Where FINRA's agent considerations sit on a single agent action

StartAgent picks an actionA tool call, a record update, a message
ConsiderationAccessWhich systems and data can this agent reach?
ConsiderationGuardrailsIs the action within the limits set for this agent?
ConsiderationHuman in the loopDoes a person approve this before it runs?
OutcomeAction runs or is stoppedEither way, the firm has something to supervise
ConsiderationTrackingThe action and the decision behind it are recorded

The four considerations are from the AI agents subsection of the 2026 FINRA Annual Regulatory Oversight Report. The report lists them without an order; placing them along one action is ours.

What an examiner can ask a firm to show

The report lists no examination requests and does not claim to be a complete inventory of what FINRA will examine.FINRA 2026 Report · Introduction Its practices still translate into records a firm using GenAI should be able to produce:

Which GenAI tools and use cases are approved, who approved them, and where the approval is recorded.
The written supervisory procedures that cover each approved use, including any GenAI the supervisory system itself relies on.
Test results for each model in use, and the version each result applies to.
Retained prompts and outputs for a given interaction, with the model version and timestamp.
Records of human review: who reviewed which outputs, when, and with what result.
For each agent: the systems and data it can reach, the actions it took, the points where a person approved an action, and the limits it runs under.

Control mapping

What a reviewer expects to be able to see.

Obligation	What the system must do	Evidence a reviewer expects
Supervision (Rule 3110)	Bring GenAI use, including GenAI inside the supervisory system, under reasonably designed procedures	WSPs naming approved GenAI tools, their permitted uses, and review procedures
Enterprise governance	Review and approve GenAI use cases formally, with business and technology experts	Approval records per use case: reviewers, controls required, and approval date
Testing	Test the capabilities, limitations and performance of each model in use	Test plans and results tied to a specific model and version
Ongoing monitoring	Monitor prompts, responses and outputs, and track which model version was used and when	Retained prompt and output logs carrying model version and timestamp
Human review	Validate outputs with human-in-the-loop review and regular checks for errors or bias	Review records attributable to a named reviewer, with outcome and date
Agent access and data handling	Monitor which systems and data each agent can reach	Per-agent inventory of tools, systems and data, with access records
Agent actions and guardrails	Track agent actions and decisions, and limit what agents may do	A record of each action (tool, parameters, outcome, time) and of each blocked or approved action
Communications and recordkeeping	Retain AI-generated communications and records under existing rules	Retained records, retrievable within required timeframes

Key dates

Timeline, drawn to scale

FINRA guidance on GenAI and LLMs since Notice 24-09

PublishedAdds AI agents

Publication dates of Regulatory Notice 24-09, the 2025 and 2026 Annual Regulatory Oversight Reports and Regulatory Notice 26-14, from finra.org. Status as of November 4, 2026.

27 June 2024Regulatory Notice 24-09 on generative AI and large language models issued.
28 January 20252025 FINRA Annual Regulatory Oversight Report published, addressing GenAI within third-party risk.
9 December 20252026 FINRA Annual Regulatory Oversight Report published, with a new GenAI topic and a subsection on AI agents.
9 July 2026Regulatory Notice 26-14 proposes modernizing Rule 2210, citing generative AI.

Sources cited

Common gaps

Where a firm's GenAI program usually falls short of the practices the report describes.

No enterprise-level view. Teams that adopt GenAI one at a time, each with its own tools and procedures, leave the firm without the enterprise-level supervisory process the report lists first.
Approvals with no record. A use case approved in a meeting, with nothing written down about who assessed it or which controls it needs, cannot be shown to an examiner.
Logs that cannot name the model. Prompt and output logs without the model version and timestamp cannot show which model produced a given output.
Testing once, at launch. The report pairs testing with ongoing monitoring. A model whose version changes after approval needs both.
Human review on paper. A procedure that puts a person in the loop needs a record of each review, attributable to that person. Without one, the review cannot be evidenced.
Agents supervised as chat assistants. An agent that calls tools and changes records needs a record of its actions and the limits it ran under. A transcript of its text covers only part of what it did.

In practice

How Meilynx maps to FINRA 24-09 →

Last reviewed November 4, 2026. This reference summarises publicly available regulatory guidance and is provided for general information. It is not legal advice. Obligations depend on an institution's charter, registration status, size, and activities. Verify against the primary sources cited above and consult counsel before relying on any summary here.