PrymeIntelligence

Regulation and standards·Regulation

FINRA and AI agents in broker-dealers

FINRA and AI agents: how supervision, communications, books and records, and vendor oversight rules apply when broker-dealers deploy agents that act.

FINRA's rules apply to AI agents the same way they apply to people and other technology at a broker-dealer. The firm must supervise what agents do, retain their communications, review what they say to customers, keep accurate books and records, and oversee vendors. An agent that acts for the firm is treated as the firm acting.

What does FINRA require, in plain words?

FINRA is the self-regulatory organisation for broker-dealers in the US, overseen by the SEC. Its rules are technology-neutral. In substance, the parts that matter most for agents are:

  • Supervision. Each firm must establish and maintain a supervisory system reasonably designed to achieve compliance with securities laws and FINRA rules, with written procedures, designated supervisors and review of business activities.
  • Communications with the public. Communications must be fair and balanced, must not be misleading, and some categories need principal approval before use. Correspondence and retail communications are reviewed under the firm's procedures.
  • Books and records. Firms must make and keep records of their business, including communications relating to it, under FINRA and SEC record-keeping rules.
  • Recommendations. Recommendations to customers must meet suitability or, for retail customers, the SEC's best-interest standard.
  • Outsourcing and vendors. Firms remain responsible for activities they outsource and must supervise them.
  • Cybersecurity and customer information protection.

The SEC's own expectations for advisers and broker-dealers are covered in SEC and AI agents in investment advisers and broker-dealers.

Who does it apply to?

FINRA member broker-dealers and their associated persons. Firms that are dually registered as investment advisers face SEC adviser rules as well. Vendors who supply agents are not FINRA members, but the firm's supervisory duties extend to what they provide.

What does supervision mean when the worker is an agent?

FINRA supervision was built around people: registered representatives whose activity is reviewed by principals. An agent is not a registered person, cannot hold a licence and cannot be disciplined. Its actions are the firm's actions, and the firm's supervisory system must cover them.

That raises four specific questions.

What counts as a communication? An agent that answers customer questions, drafts emails or summarises research for clients is producing communications. They must be retained and reviewed like any other. A chat with an agent is correspondence, not a casual conversation.

When is an agent making a recommendation? An agent that suggests a security, a strategy or an account type to a particular customer may be making a recommendation. The suitability and best-interest obligations then apply, and the firm must show the recommendation was appropriate for that customer. Most firms keep agents to education and factual responses, and route anything personalised to a registered representative.

Who reviews what the agent does? Principal review must be designed for volume. An agent can produce more communications in a day than a team of representatives. Sampling must be risk-based and documented, and pre-approval must apply where the rules require it.

What does the agent's record look like? Books and records rules require that business communications be preserved. For an agent, the record must include what it was asked, what it said and what it did, in a form that can be produced on request.

How do you meet FINRA expectations for an agent?

Requirement Control Evidence it leaves
Supervisory system Written procedures covering each agent: scope, limits, reviewer, escalation Updated written supervisory procedures
Designated supervisor Principal named for each agent's activity Supervisory assignments
Communications review Agent output retained and reviewed on a risk-based schedule; pre-approval where required Review logs; approvals
Fair and balanced communications Agent restricted to approved content for retail communications Approved content library; tests
Recommendations Agent barred from personalised recommendations; routed to registered persons Routing records; refusal records
Books and records Complete, unalterable record of agent interactions and actions Retained records; retrieval tests
Vendor oversight Due diligence and ongoing monitoring of the agent provider Due diligence file; review records
Cybersecurity Agent access limited per capability; outside content cannot trigger trades or transfers Access settings; refusal records
Change control Model, brief and source changes reviewed before release Change records

A deeper question for firms that build or heavily customise agents is whether they have taken on the developer's responsibilities. The EU concept of substantial modification is a useful parallel even for US firms, because the same line decides who must validate what.

What will an examiner ask?

  • Which agents interact with customers, and what can each do?
  • Show us the written supervisory procedures for agent activity.
  • How are agent communications retained and reviewed? Show a sample.
  • Can the agent make recommendations? How do you stop it?
  • What vendor due diligence did you do?
  • How would you produce every interaction an agent had with a particular customer?

The general approach regulators take to auditing agents is described in how regulators expect AI agents to be audited.

What do FINRA rules not cover?

They do not classify AI by risk tier. They do not set model validation standards; bank-style model risk guidance applies to banks, though many broker-dealers borrow its methods. They do not govern consumer privacy beyond customer information protection rules; GLBA and state laws apply. And they do not tell you which actions an agent may take; the firm decides and must supervise. For a wider view across regimes, see AI agent regulation, mapped to controls.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

Questions people ask

Does an AI agent need to be registered with FINRA?

No. Registration applies to people. The firm's supervisory system must cover the agent's activity, and any recommendation or communication it makes is the firm's.

Are chat logs with an AI agent business records?

If they relate to the firm's business, treat them as communications to be retained under the record-keeping rules. That includes internal agents whose output feeds customer business.

Can an agent place trades for customers?

Only within the firm's supervisory controls and the customer's instructions. Most firms keep order entry with registered persons or with tightly bounded systems, and hold any agent-proposed trade for approval.

Decide what your agents may do. Prove it. Enforce it.

Pryme Intelligence certifies each agent on your own cases, checks every action against five bounds at the moment it happens, and files every verdict.