Implementation Guide

    Supervising AI Chatbots on Wealth Management Client Portals as Retail Communications

    FINRA says firms are responsible for chatbot communications and must supervise them under FINRA rules. Classify each interaction type, turn the Rule 2210 content standards into policies applied to every response, define escalation to a human in written supervisory procedures, and retain full chat sessions with the decisions made on them. Enforcing policy at runtime gives the supervisory system something testable.

    Pre-launch checklist

    • Interaction types classified, with a named approver for each determination
    • Pre-use approval exceptions assessed and documented per use case
    • Content standards expressed as runtime policies for prompts and responses
    • Escalation triggers, reviewer roles and response times written into procedures
    • Full chat sessions retained, with decisions and reviewer actions linked by session
    • Retrieval tested, and Notice 26-14 tracked for procedural impact

    Start with classification

    FINRA's guidance says chatbot communications used with investors may be correspondence, retail communications or institutional communications, depending on the nature and number of persons receiving them. The firm is responsible for the content whether a human or AI generates it. That responsibility covers supervision, SEC and FINRA recordkeeping, and the content standards in Rules 2210 and 2220. Using a vendor's model does not transfer it.

    The guidance does not sort portal interactions into those categories. The determination belongs to the firm, and it should be written down with a named approver. A workable approach is to classify by interaction type: general education, account-specific information, product-related responses and recommendations, and planning guidance. Each type is then assessed on content and likely recipient count.

    Classification matters because current Rule 2210 requires a qualified registered principal to approve retail communications before use, subject to exceptions. The requirement does not currently apply to retail communications on an online interactive electronic forum, or to communications that make no financial or investment recommendation and do not promote a member product or service. Whether a chatbot use case fits either exception should be resolved and documented for that use case. Responses that recommend or promote products are the least likely to qualify for flexible treatment.

    Do not plan around relief that has not arrived. As of November 2026, Regulatory Notice 26-14 is a proposal to replace the prescriptive pre-use approval requirement with risk-based standards. Current rule text governs.

    Translate content standards into runtime policy

    FINRA's content standards are the testable criteria. Communications must be fair and balanced, and they must not contain false, exaggerated, unwarranted, promissory or misleading statements. Each can be expressed as a policy evaluated on every response: a performance statement without balancing context, language that implies a guaranteed outcome, or a product claim that omits risk.

    Placement is a design choice. Sampled review after the fact finds a problem after the client has read it. A policy applied in the execution path acts before the response is delivered, and the decision can be recorded against the specific session. The standards apply whether content comes from a person or from technology, so the firm needs a way to show they were applied to dynamic output.

    Apply policy to inputs as well as outputs. Commenters have raised whether text customers submit to a chatbot is treated differently from text the chatbot generates, and the question is not settled. Capturing and evaluating both is the conservative design.

    The chatbot's descriptions of itself fall under the same standards. Retail communications that mention AI tools should accurately describe how AI is incorporated and balance benefits with risks.

    • Allow: Education and service responses that meet the content standards pass unchanged, and the decision is recorded.
    • Redact: Remove specific content, such as account details outside the interaction's scope, and deliver the rest.
    • Block: Stop promissory, exaggerated, unbalanced or misleading statements before they reach the client.
    • Escalate: Hold the response and route the conversation to a qualified person when risk conditions are met.

    Write the supervisory procedures

    FINRA's FAQ says firms must review chat content in the same manner as required for other correspondence under Rule 3110(b) and 3110.06 through .09. That is the stated baseline for written supervisory procedures. The procedures should say how chatbot content is classified, who reviews it, how review is evidenced, and how often outputs are tested.

    Escalation criteria to a human are not defined in the guidance, so the firm must set them and be ready to explain the rationale to examiners. Tying triggers to interaction type and risk is a defensible structure. Examples include a request for a product recommendation, a question about expected returns, a planning scenario that depends on a client's circumstances, or a repeated policy block within one conversation. Name the reviewer role, the time expected for response, and what the client sees while a response is held.

    Exceptions need their own procedure. Define what happens when a block fires, when a reviewer overrides an automated decision, and when a response is later found to have breached a standard. Each case should produce a record with a reason and an owner.

    Before deployment, evaluate the tool against applicable rules, as the 2026 Oversight Report advises, and monitor new chatbot features as they are released. Track the outcome of Regulatory Notice 26-14 and identify which current controls depend on principal pre-use approval.

    Define the evidence and retention

    The 2026 Oversight Report names supervision and retention of chatbot communications, and retention of the chat sessions, as an effective practice. The full session is therefore the core artifact. Retention rules for broker-dealers come from both the SEC and FINRA, so the retention schedule should be set against those rules and tested for retrieval, not only for storage.

    Beyond the session, an examiner will reasonably ask whether the firm can show that content standards were applied and what happened as a result. FINRA does not list additional required artifacts, so the following is design judgment: store with each session the policy evaluated, the decision taken, any escalation, and the reviewer's action and timing. Linking these by session identifier turns supervision from an assertion into something that can be inspected one conversation at a time.

    This also supports model risk management. A record of what the system was allowed to do, and what it did, lets the firm show that supervision operates on the deployed behavior and not only on the pre-deployment design.

    Where runtime enforcement fits

    Written supervisory procedures state what the firm requires. Runtime enforcement is one mechanism that carries those requirements out on each interaction. The procedures remain the firm's responsibility, including classification, escalation criteria and reviewer accountability.

    Trussed is one way to enforce policy in the execution path. It evaluates individual interactions and applies allow, redact, block, rate limit or escalate for human approval decisions at runtime, not after the fact, and it generates audit-ready evidence from that enforcement. It can be deployed self-managed in the customer's cloud, with configurable retention and data residency. Trussed's deepest experience is in insurance, a regulated financial sector, and it operates in four regulated verticals: insurance, pharma, healthcare and education. Rule content for retail communications is defined by the firm and its counsel.

    Supervision by design for a portal chatbot

    Classify

    Assign each interaction type to correspondence, retail or institutional communication.

    Enforce

    Apply content standards to prompts and responses: allow, redact, block or escalate.

    Escalate

    Route defined risk conditions to a human reviewer under written procedures.

    Evidence

    Retain sessions, policy decisions and reviewer actions for examiners.

    See supervision evidence produced at runtime

    Review how policy enforcement in the execution path produces session-level, audit-ready evidence for regulated enterprises.

    See How Evidence Is Produced