How to Prevent an Advisor AI Assistant from Making Performance Guarantee Statements
Define promissory and guarantee language as a versioned policy, enforce it on model outputs in the execution path before anything reaches a client, map each violation type to allow, redact, block or escalate for human approval, and retain per-interaction evidence of the policy version, the evaluated output and the action taken. Treat prompt instructions as one input that lowers violation frequency, not as the control.
Where enforcement sits in the request and response path
- Request context: attach communication type, audience and the policy version in force before the model is called.
- Prompt layer: include policy instructions to reduce violation frequency, treated as an input only.
- Model call: generate the draft or response.
- Output evaluation: assess the full output against each policy category before it is shown to a client or released to an advisor for sending.
- Enforcement action: allow, redact, block or escalate for human approval.
- Evidence write: record the interaction and the action as the enforcement occurs.
Evidence to retain for each interaction
- Interaction identifier, timestamp, communication type and audience
- Policy name and version in force at the time of evaluation
- The model output that was evaluated, and the output delivered if it was changed
- Policy categories matched and the enforcement action taken: allow, redact, block or escalate
- For escalations: the reviewer, the decision, the timestamp and any rationale
- Detection thresholds or configuration in effect, so results can be reproduced and tuned
Start with a policy definition that maps to rule concepts
Enforcement can only be as precise as the definition behind it. "Promissory" is not a single pattern. FINRA Rule 2210 generally prohibits member communications from predicting or projecting performance, implying that past performance will recur, or making exaggerated or unwarranted claims, opinions or forecasts. Its content standards also require communications to be fair and balanced and bar false, exaggerated, unwarranted, promissory or misleading statements. For investment advisers, the SEC marketing rule adds performance-presentation requirements, including that gross performance not be shown without net performance of at least equal prominence, over the same period and using the same return type and methodology.
A detector has to separate these overlapping concepts, so write the policy as distinct categories: explicit guarantees or promises of outcome, predictions or projections, targeted returns, implied recurrence of past performance, exaggerated or unwarranted claims, and unbalanced performance presentation. For each category, add firm-specific examples, including paraphrased and implied versions such as "you won't lose money" or "this strategy will keep delivering." Label these examples as firm interpretation, not regulator text.
Version the policy by communication type and audience. FINRA has approved exceptions to the projection prohibition, including for institutional communications and certain communications to qualified purchasers or knowledgeable employees about specified private offerings, and a February 2026 FINRA notice proposed allowing projections or targeted returns subject to conditions. As of November 2026, compliance should own the decision on whether any projection content is permitted, for which audience, and under which policy version. Until that decision is recorded, the conservative default is to treat projection content as not permitted.
Choosing an outcome for each violation type
No regulator prescribes which outcome fits which violation, so this mapping is a firm design decision to validate with compliance and legal. The table below is a starting point grounded in the rule concepts above.
Redaction deserves care. Removing a sentence can leave surrounding text that is still misleading, so re-evaluate the output after any redaction. Blocking is the safest outcome for clear violations but needs an approved fallback message so the advisor is not left with nothing. Escalation handles ambiguity but adds reviewer load and latency, so reserve it for cases where a human judgment is genuinely needed.
- Explicit guarantee or promise of outcome: Block. There is no compliant version to preserve, and an approved fallback response can be returned instead.
- Isolated implied-recurrence or exaggerated phrase: Redact, then re-evaluate the remaining text. If the meaning is still affected, block.
- Projection or targeted return: Depends on audience and policy version. Block for audiences where projections are not permitted. Escalate for approval where they may be permitted under conditions.
- Performance shown without required elements: Escalate or block. Redaction cannot supply missing net performance or a matching time period.
- Borderline or low-confidence detection: Escalate for human approval, with the evaluated output and the matched policy category shown to the reviewer.
- No policy category matched: Allow, and record the evaluation so the control's operation is still evidenced.
Why prompts and after-the-fact review are not enough
NIST reports that many mitigations for adversarial machine learning are empirical and lack provable guarantees, and states that so long as a model has any probability of exhibiting an undesired behavior, prompts exist that can trigger it. Alignment that attenuates a behavior without removing it remains vulnerable. OWASP similarly says that training a model toward a behavior can be effective but does not guarantee adherence. For a compliance leader, the implication is direct: a system prompt saying "never promise returns" lowers the rate of violations but cannot be cited as evidence that violations are prevented.
Review after the fact fails differently. It can find a prohibited statement, but only once the statement may already have reached a client. Sampling-based review also leaves unreviewed interactions with no demonstrated control. Post-hoc review remains useful as a supervisory and tuning feedback loop, but it does not stop delivery.
Output-side detection has limits of its own. NIST cites research on the impossibility of model-based detection preventing all impermissible outputs, so any classifier for promissory language will have residual misses. Test against paraphrased and implied guarantees, not only keywords. Measure both false negatives and false positives. Aggressive thresholds block legitimate, balanced commentary and push advisors toward workarounds. Permissive thresholds let more through. Document a residual-risk position, set thresholds per category, and use escalation to absorb uncertainty near the boundary.
Where runtime policy enforcement fits
Trussed enforces policy in the execution path, so outputs are evaluated at runtime, not reviewed after the fact. It supports the outcomes described above: allow, redact, block and escalate for human approval, applied to individual interactions. Each enforcement decision generates audit-ready evidence, which gives compliance teams a per-interaction record of what the control did. It is self-managed in the customer cloud, with configurable retention and data residency.
The policy definition, the audience rules and the choice of outcome per violation type remain compliance decisions. Whatever tooling a firm uses, the same sequence applies: define, enforce before delivery, handle borderline cases through approval, and keep evidence that the control operated.
Layered control for client-facing AI output
1. Define
Versioned policy by communication type and audience
2. Enforce
Evaluate the model output before delivery
3. Decide
Allow, redact, block or escalate for approval
4. Evidence
Record policy version, output, action and approver
Review your enforcement path for client-facing AI output
Walk through where policy is enforced on model outputs and what evidence each interaction produces.
Request a Demo