Implementation Guide

    FINRA Rule 3110(b)(4) Risk-Based Review of AI-Drafted Correspondence

    Treat the risk-based, documented review system that Rule 3110(b)(4) calls for as the design specification. Assess risk, define tiers, assign a review action to each tier (allow, redact, block or escalate), and retain evidence for every decision. Runtime controls can apply principal-approved parameters to individual interactions. Supervisory judgment and the written procedures remain with registered principals.

    Evidence to retain for each decision

    • Policy and parameter version in force, with the principal approval of those parameters
    • Risk tier assigned and the criteria that triggered it
    • Decision outcome: allowed, redacted, blocked or escalated
    • Model version used and the time of the interaction
    • For escalations: the principal reviewer, date of review and steps taken
    • Prompt and output records handled according to your client-data and retention policies

    What the rule requires and what it leaves to the firm

    Rule 3110(b)(4) requires procedures for reviewing incoming and outgoing written, including electronic, correspondence relating to the member's investment banking or securities business. Under firm procedures, the review must identify and handle customer complaints, instructions, funds and securities, and communications whose subject matter requires review under FINRA rules and federal securities laws. A registered principal conducts the review, and it is evidenced in writing, electronically or on paper.

    The risk-based element sits in the rule's Supplementary Material. By employing risk-based principles, a member decides how far additional policies and procedures extend to correspondence outside the listed subject matters. FINRA has described "risk-based" as a methodology for identifying and prioritizing the areas that pose the greatest risk of potential securities law and SRO rule violations. Where procedures do not require all correspondence to be reviewed before use or distribution, they must provide for further elements.

    The rule does not mandate specific technology. Regulatory Notice 24-09 states that FINRA rules apply to generative AI as they apply to any other technology, and that Rule 2210 content standards apply whether a human or a tool generated the communication. An AI draft is therefore judged against the same standards as one an advisor typed. What changes is volume and variability, which is why a single lexicon or a post-send sample is a poor fit on its own.

    Start with a risk assessment and tiering criteria

    Begin with a written assessment of where AI-drafted correspondence could produce the outcomes your procedures exist to prevent: unbalanced or promissory statements, unapproved performance claims, and client data appearing in a prompt or an output. FINRA's 2026 oversight material also flags hallucinations, meaning inaccurate or misleading output presented as fact, and inaccurate client or market data. Record the rationale in the written supervisory procedures so the tiers can be traced to the assessment.

    The four subject-matter categories in the rule and the Rule 2210 content standards are a practical starting point for tier criteria. Add the attributes a firm can actually determine at the time of an interaction: content type, recipient, product discussed, retail versus institutional audience, presence of client data, and advisor history. None of these attributes is prescribed. They are design choices, and each should be justified in the assessment.

    Keep the criteria few enough to apply consistently. A tier definition that a principal cannot explain to an examiner in a sentence or two is usually too complicated to defend.

    Example tiers and review actions

    These tiers are illustrative design suggestions, not requirements. Each firm sets its own criteria and actions based on its assessment.

    • Routine, no client data, no product or performance content: Allow. Include in post-send sampling as defined in the written supervisory procedures.
    • Client data present in a prompt or output: Redact the data that should not appear, then apply the review action for the underlying content tier.
    • Product discussion to a retail audience: Apply Rule 2210 content checks. Escalate to a principal where the content is not clearly within approved language.
    • Performance claims or promissory language not tied to approved material: Block, or escalate for principal approval before anything is sent.
    • Complaint, instruction, or funds and securities subject matter: Escalate to a registered principal and record the review in writing.

    What can be enforced at runtime and what stays with principals

    Review after the fact works on what was already sent. Enforcing at runtime means the decision is applied to the individual interaction before the draft leaves the AI path: allow it, redact part of it, block it, or hold it for human approval. For a rule built on risk prioritization, that moves the highest-risk tiers from detection to prevention.

    The limit is who owns the judgment. FINRA has stated that a principal's use of an automated supervisory tool is a direct exercise of supervision by that principal, who remains responsible for compliance. Final responsibility rests with the member. Regulatory Notice 14-10 addresses technology-based review parameters in the transaction context, but its logic is a sound model here: a principal reviews the parameters that decide what merits further review and documents that review in writing. Runtime rules should therefore be principal-approved, versioned, and changed through a documented process.

    What automation can apply: pattern and content checks, redaction of client data, blocking of defined categories, and routing to an approver. What stays with principals: setting the tiers, approving thresholds, deciding escalated items, and maintaining the written supervisory procedures. A runtime log can support the principal review requirement, but it does not by itself establish that the principal review occurred.

    Trussed enforces policy in the execution path across models, agents and tools, with allow, redact, block and escalate-for-human-approval actions at the level of individual interactions. It generates audit-ready evidence from that enforcement. It applies the policies a firm defines and does not replace principal judgment. Its deepest regulated-sector experience is in insurance.

    Validate the components and connect to model risk management

    Regulatory Notice 24-09 says that where generative AI is used within a supervisory system for electronic correspondence review, policies should address model risk management, data privacy and integrity, and the model's reliability and accuracy. FINRA's AI report describes validation covering input data, algorithms, parameters such as risk thresholds, and outputs. Apply the same discipline to the drafting model and to the control layer that decides what happens to its output.

    In practice, test the tier thresholds against representative drafts before use, repeat that testing when the model, the policy or the thresholds change, and keep the results with the policy version they apply to. FINRA's report also refers to formal review and approval of generative AI use cases and to documentation under a governance or model risk framework. Per-interaction evidence supports that framework: it shows the control operated as designed, not just that it was configured.

    A useful test before examination: pick one AI-drafted message at random and reconstruct its policy version, decision, model version and any principal review. If that takes more than a query, the evidence design needs work.

    As of November 2026, the rule text and guidance described here are the current basis for this approach. Confirm the full text of the Supplementary Material when drafting procedures.

    Risk-based review program for AI-drafted correspondence

    Assess risk

    Identify where AI-drafted content creates the greatest potential for rule violations.

    Define tiers

    Classify by content type, recipient, product, client data and advisor history.

    Apply review action

    Allow, redact, block or escalate to a registered principal.

    Retain evidence

    Keep the policy version, decision, model version and principal review for each item.

    Show an examiner what your AI was allowed to do

    See how enforcement in the execution path produces per-interaction, audit-ready evidence that supports a principal-owned, risk-based review program.

    See How Evidence Is Produced