Writing a Supervisory Lexicon for AI-Generated Advisor Communications
A supervisory lexicon for AI-generated communications is a versioned policy artifact, not a word list. Define the interactions in scope, group terms and patterns into risk categories and severity tiers, map each tier to an action (allow, redact, block, or escalate for human approval, or leave to downstream review), test against realistic AI output, assign owners, and retain evidence of the lexicon version applied to each communication.
What changes when the author is a model
Most existing lexicons were written for human-authored emails and messages. They assume a bounded volume, a recognizable set of phrasings and a clear point of origin: an identified advisor typing a message. AI-generated drafts and responses alter all three. Volume can rise sharply, because drafts are cheap to produce. Phrasing varies more, because a model can express the same claim in many ways. Origin shifts, because the text may be produced inside an application before any person reads it.
Legacy keyword lists respond poorly to this. Fixed terms miss paraphrases, which creates gaps. Broad terms fire on harmless text at scale, which creates false positives and review backlogs. The answer is not a longer list. It is a lexicon with structure: categories, severity tiers, different match types, context rules, and a defined action for each tier.
NIST's generative AI profile can serve as a general reference for governing generative AI risk. It does not replace the firm's own supervisory procedures, which the lexicon should be written into.
Mapping tiers to enforcement actions
The central design decision is which matches are acted on before content leaves the system and which are reviewed afterward. A practical rule: enforce at runtime when the cost of one bad communication reaching a client is high and the rule is precise enough to trust. Leave to downstream review when judgment is needed, precision is low, or the pattern is better understood in aggregate.
The mapping below is a starting point. Calibrate it against your own risk appetite and supervisory procedures.
- Critical, precise match: Block. Examples: explicit guarantees of returns, or prohibited disclosures of sensitive client data.
- Critical, ambiguous match: Escalate for human approval. A reviewer decides before the communication is released.
- High, data-bearing match: Redact. Remove the sensitive element and allow the remainder, recording that redaction occurred.
- Moderate: Allow and flag for downstream supervisory review, sampled or full depending on volume.
- Informational: Allow and retain the match for trend analysis and later tuning.
Test, tune and respect the limits of matching
Lexical matching sees strings, not intent. A model can convey a prohibited idea without using any listed term, and it can use a listed term in an acceptable sentence. Pattern and context rules narrow both problems but do not remove them. Plan for the lexicon to be one control among several, not a complete test of whether content is compliant.
Build a test set before release. Include real examples of past communications that should trigger each category, paraphrased variants of those examples, and clean text that contains near-miss wording. Run the lexicon against the set and measure two things per rule: what it misses and what it flags wrongly. Rerun after every change and compare against the prior version. Add AI-generated samples from the actual prompts and workflows in scope, since model output differs from human writing.
Tuning follows from reviewer outcomes. If a rule produces mostly dismissed alerts, tighten its context or lower its tier. If reviewers find problems no rule caught, add a pattern or category. Keep the reasoning for each change.
Do not treat AI-text detection as a substitute for content review. NIST's GenAI pilot work notes that the detection approaches it describes may not fully capture the broader landscape of AI-generated content. Detecting that text is machine-written does not tell you whether it is compliant.
Versioning, ownership and evidence
Give every lexicon release an identifier, an effective date, an approver and a change summary. Store it in version control, with test results attached to each release. Assign one named owner per category, with compliance holding approval authority and technology owning deployment. Changes should follow a defined path: proposal, test, approval, release, and a post-release check of alert volumes.
Evidence should let you answer a specific question for any communication: which lexicon version was applied, what matched, and what happened as a result. Retain the items below.
- Lexicon version identifier and effective date at the time of the interaction
- Interaction scope: which AI-assisted workflow produced the content
- Rule identifiers that matched, with tier and category
- Action taken: allowed, redacted, blocked or escalated
- For escalations: reviewer identity, decision and timestamp
- Approval records and test results for each lexicon release
Where runtime enforcement complements supervisory review
Downstream review looks at communications after they exist. It is well suited to judgment, sampling and trend detection. It cannot stop a single non-compliant message from reaching a client. Runtime enforcement acts before release, on the individual interaction, and is well suited to precise, high-severity rules. The two are complementary: enforcement narrows what reaches review, and review feeds the tuning of enforced rules.
Trussed enforces policy in the execution path, at runtime, not after the fact. It can allow, redact, block, rate limit, or escalate for human approval on individual interactions and tool calls, and it generates continuous audit-ready evidence from that enforcement, mapped to frameworks including NAIC, NYDFS and NIST. It can be self-managed in the customer cloud, with configurable retention and data residency. Tiers you assign to runtime action are the kind of rule this layer applies, and the evidence it produces supports the records listed above.
Whether you use Trussed or another approach, the test is the same: for any AI-generated communication, can you show what policy applied, what it did, and which version was in force?
Anatomy of an AI-ready lexicon
Scope
Which AI-assisted interactions the lexicon covers
Categories and tiers
Risk themes with severity levels
Match rules
Exact terms, patterns and context conditions
Enforcement map
Runtime action or downstream review per tier
Evidence
Lexicon version and outcome per communication
Turn the lexicon into enforced policy and evidence
Discuss how tiered lexicon rules can be enforced in the execution path and how the resulting audit-ready evidence supports supervisory review.
See How Evidence Is ProducedHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.