Implementation Guide

    Supervising AI-Drafted Client Emails Before an Advisor Sends Them

    Supervise AI-drafted client emails as an enforced control placed between draft generation and delivery. Classify each draft by risk, run policy checks for sensitive data and prohibited content, redact or block what fails, route higher-risk drafts to a qualified human approver, and record the policy applied, the action taken and the approver for every draft. A send should not be possible until that decision exists.

    The pre-send workflow, step by step

    1. 1

      1. Classify the draft

      Assign a tier based on communication type, presence of sensitive data and the kinds of claims made.

    2. 2

      2. Run policy checks

      Check for sensitive client data, prohibited phrases and content that conflicts with communication policy.

    3. 3

      3. Redact or block

      Remove disallowed data where safe to do so. Block the draft when redaction would leave it misleading or non-compliant.

    4. 4

      4. Escalate where required

      Send drafts that meet escalation criteria to the named approver for the tier, with the reason for escalation attached.

    5. 5

      5. Release or reject

      Allow the send only after a policy decision, and an approval where required, is recorded for that draft.

    6. 6

      6. Retain the evidence

      Store the draft, policy version, action, reviewer and decision as a single reviewable record.

    Evidence records that show supervision happened

    • Draft content as generated, and final content as sent
    • Assigned review tier and the reason for it
    • Policy version applied and the result of each check
    • Enforcement action taken: allow, redact, block or escalate
    • Reviewer identity and role, with the approval or rejection decision and any edits
    • Timestamps for generation, each decision and delivery

    Supervision is a checkpoint, not a review habit

    Many firms start with advisors reading AI drafts before they hit send. That is a practice, not a control. Nothing enforces it, reviewers apply different criteria, and afterward nobody can show what was checked, who approved it or which policy applied.

    The regulatory frame points toward something more structured. For FINRA member firms, Regulatory Notice 26-14 (dated July 9, 2026) states that the rule requires an appropriately qualified registered principal to approve retail communications before use. It also states that members may adopt alternative supervisory procedures for reviewing correspondence and institutional communications, and that Rule 2210's principles-based content standards apply to all member communications. The notice's snippet does not establish how it treats AI-generated content, so confirm with counsel how each type of client follow-up email is classified and who the qualified principal is for each.

    NIST guidance supplies the risk language. It defines confabulation as the production of confidently stated but erroneous or false content, which is the risk behind unsupported statements in a draft. Its AI RMF GOVERN 1 outcome calls for policies, processes and procedures that are in place, transparent and implemented effectively. A supervision policy that exists only in a manual meets the first part of that and fails the last.

    Define risk-based review tiers

    Tiers are a design choice, not a sourced requirement, but they follow the FINRA split between principal pre-use approval and alternative supervisory procedures. The goal is to put human attention where the risk is and keep it off drafts that a policy check can handle.

    A workable model has three tiers. In the first, drafts contain no client-specific sensitive data and make no performance or product statements. They pass through policy checks only, with periodic sampling by a reviewer. In the second, drafts include account or personal details, or statements a reviewer would want to see. They require approval by a designated supervisory reviewer. In the third, drafts that fall under retail communication treatment require approval by a qualified registered principal. Content standards apply to all three tiers, because the Rule 2210 standards apply to all member communications.

    Because the evidence does not show that a policy check can reliably detect confabulation, treat performance claims, product claims and forward-looking language as triggers for human review rather than assuming automation will catch an unsupported statement.

    Specify policy checks and escalation criteria

    Two categories are well suited to pre-send checks: sensitive-data exposure and prohibited content. Both are pattern-detectable, which makes allow, redact or block outcomes practical. This is a design inference, and no detection accuracy data is assumed, so measure false positives and misses against your own sample drafts before relying on any check.

    Write each check as a testable rule tied to a policy version. Examples include account numbers or identifiers appearing in the body, references to another client, specific language your communication policy prohibits, and promissory or guarantee-style phrasing. A rule that cannot be tested cannot be shown to an examiner.

    Escalation criteria work as thresholds. NIST describes approval thresholds as part of go/no-go policies for deploying generative AI. The same logic applies to individual drafts: define what moves a draft from policy-only handling to human approval, and review those thresholds on a schedule. To avoid a bottleneck, keep triggers narrow, redact before escalating so reviewers see a cleaner draft, state why each draft was escalated, and assign queue ownership with a defined response expectation. Track approval turnaround and how often reviewers change or reject escalated drafts. A high rejection rate suggests thresholds or checks need adjustment.

    Treat the model as an untrusted component. OWASP describes a scenario in which an injected prompt leads an assistant to ignore guidelines, query private data and send emails, and notes that prompt-level instructions alone are not sufficient prevention. The checkpoint therefore has to sit outside the model, and it should be tested with penetration testing and breach simulations that treat the model as an untrusted user.

    Where runtime policy enforcement fits alongside human review

    Human review and runtime enforcement do different jobs. The reviewer applies judgment to a draft. Enforcement ensures the reviewer step cannot be skipped and that the same rules apply to every draft. Without enforcement in the path, approval depends on advisors following procedure.

    Trussed enforces policy in the execution path at the level of individual interactions and tool calls, including a send action. Outcomes are allow, redact, block or escalate for human approval, and enforcement generates audit-ready evidence of each decision. Trussed does not replace the principal's judgment. It makes the checkpoint mandatory and records it. Its scope is the individual interaction, so tiering, reviewer assignment and communication classification remain decisions your compliance function owns.

    Whatever tooling you use, assign named senior management accountability for the control framework. A proposed NYDFS circular letter on AI in underwriting and pricing, as of November 2026, frames senior management responsibility for AI systems this way. It does not address client email, but the accountability model is a reasonable design analogy. Also run an after-action assessment whenever an erroneous email is sent, to confirm that incident response and recovery steps were followed and worked.

    Pre-send supervision path

    Classify

    Assign a review tier to each draft before anything else happens.

    Check

    Apply policy checks for sensitive data and prohibited content.

    Enforce

    Allow, redact, block or escalate based on the result.

    Approve

    A qualified reviewer decides on drafts that require human approval.

    Record

    Retain the draft, policy version, action and decision as evidence.

    Review your pre-send checkpoint

    Talk through review tiers, policy checks and evidence records for AI-drafted client communications with a Trussed expert.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.