FINRA Rule 2210 and AI-Generated Communications: Enforcing Limits on Projections and Promissory Claims

What the rule says and what changed

Rule 2210(d)(1)(B) states that no member may make any false, exaggerated, unwarranted, promissory or misleading statement or claim in any communication. Rule 2210(d)(1)(F) prohibits predictions or projections of performance, implying that past performance will recur, and any exaggerated or unwarranted claim, opinion or forecast. Three exceptions have long applied: certain hypothetical illustrations of mathematical principles, certain investment analysis tools or the reports they produce, and certain price targets in research reports on debt or equity securities.

On 2026-10-20 the SEC published an order approving a FINRA change, as modified by Partial Amendment No. 1, that adds a fourth exception. It permits projected performance or targeted returns in limited circumstances. As originally filed in February 2026, the proposal would allow a member to project the performance of, or provide a targeted return for, a security, a securities portfolio, or an asset allocation or other investment strategy, subject to conditions meant to ensure projections are carefully derived from a sound basis.

For AI-drafted content the practical consequence is that a rule reading "never emit a projection" is too blunt. FINRA's Rule 2210 FAQ, which predates the approval, says retail communications may not project or predict returns, including yields, income, dividends, capital appreciation or percentages, and that the prohibition extends to target returns. Policy written from that FAQ alone will need revisiting. Regulatory Notice 24-09 is clear that the content standards apply whether a communication is generated by a human or a technology tool, so the firm owns the output either way.

Build a claim taxonomy from the rule text

Start from the language of paragraphs (B) and (F), not from a keyword list. The categories below are a design inference from the rule text, not a list prescribed by FINRA. Add examples from Regulatory Notice 20-21 and the FAQ as test cases: assertions about the likelihood of a future public offering, claims about the future success of a new or untried business model, and projections of investment performance.

The taxonomy must also name what is conditionally permitted. The FAQ separates returns to investors, which are prohibited, from reasonable forecasts of issuer operating metrics such as sales, revenues or customer acquisition, which are not prohibited if the communication provides a sound basis for evaluating the facts. A detector therefore has to determine what a number refers to, not only whether a number is present.

• Projected or targeted investor returns: Forward-looking yields, income, dividends, capital appreciation or percentage figures for a security, portfolio or strategy.
• Implied recurrence of past performance: Phrasing that suggests historical results will repeat.
• Promissory language: Guarantees, assurances of outcome and "can't lose" phrasing.
• Exaggerated or unwarranted claims: Opinions or forecasts stated with implied certainty and no basis.
• Conditionally permitted content: Hypothetical illustrations, analysis tool output, research price targets, issuer operating forecasts, and content that may fall under the fourth exception.

Map actions to violation types

The mapping below is a design proposal, not a FINRA requirement. The aim is to match the severity of the action to the confidence and the consequence of the finding.

Block clear promissory or guarantee language. There is no compliant version to preserve, and a block stops the draft before it enters a review queue as if it were a candidate for approval. Redact or rewrite an isolated unsupported return figure where the rest of the draft is compliant, and record what was removed. Allow content that matches a defined permitted pattern, such as a calculator-style illustration that computes the outcome of assumed variables without predicting their likelihood or the outcome.

Escalate anything that requires judgment: hypothetical illustrations with ambiguous framing, issuer metrics near a return claim, price-target text, and any projection or targeted-return content that could fall under the fourth exception. The calculator-versus-projection line turns on whether the content predicts the likelihood of variables or outcomes, which is a semantic judgment that keyword matching is unlikely to make reliably. Until the firm has defined in policy the conditions it applies to the fourth exception, with a named owner, route that content to human approval rather than allowing it automatically.

• Allow: Content matching a defined, owner-approved permitted pattern.
• Redact: Isolated unsupported figures or phrases where the remainder is compliant.
• Block: Guarantees, promissory language and clear return projections outside any permitted path.
• Escalate for human approval: Ambiguous illustrations, issuer metrics, price targets and possible fourth-exception content.

Enforce in the execution path and manage false positives

Rule 2210(d)(1)(B) and (F) apply to content however it is produced, so checks can sit at generation time on model outputs. Runtime enforcement is a design choice for the firm, not something the rules mandate. Apply checks to outputs and also to tool and retrieval results that feed drafts, because a projected figure can enter through source material and be repeated by the model as fact.

Set policy by communication type. FINRA's FAQ says chatbot communications may be correspondence, retail communications or institutional communications depending on the nature and number of recipients, and the firm must supervise them under applicable rules. Document which AI use cases produce which category, then attach different rule sets and escalation routes to each.

Build a labeled test set before switching enforcement on. Include hypothetical calculators, issuer operating forecasts, price-target text, and past-performance statements with and without recurrence implications. Measure false positives and false negatives per category, and start new rules in an escalate-only mode. Define each exception as a rule-level entry with a named owner and an approval record, not an ad hoc override. Where applying the rules to a specific AI use is ambiguous, FINRA says firms may seek interpretive guidance.

Where runtime enforcement fits with principal review

Regulatory Notice 26-14, published 2026-07-09, states that Rule 2210 currently requires a qualified registered principal to approve each retail communication before the earlier of its use or filing, with exceptions for retail communications posted on an online interactive electronic forum and those that make no financial or investment recommendation and do not promote a member product or service. FINRA acknowledges that pre-use approval presents significant challenges for some social media communications and some generated by AI. Nothing in that notice says automated controls can replace principal approval or supervisory procedures.

Runtime enforcement is best understood as a filter ahead of the principal. It stops clear violations before they reach a queue, narrows what reviewers see to the ambiguous cases, and gives them the matched rule and the evidence of what the model produced. Enforcement decisions should feed supervisory procedures, not bypass them.

Trussed is the enforcement and evidence layer for this pattern. It enforces policy in the execution path across models, agents, tools and MCP servers, applying allow, redact, block, rate limit and escalate for human approval to individual interactions and tool calls, and it produces audit-ready evidence from those decisions. It supports self-managed deployment in the customer cloud with configurable retention and data residency. It has run continuously for 18+ months in a regulated enterprise, with 100+ models under one policy plane. Its customers are in insurance, pharma, healthcare and education, and insurance is where it is deepest. The claim rules, exceptions and reviewer routes remain the firm's to define.