Comparison

    Lexicon Surveillance vs Runtime Policy Enforcement for Advisor AI Assistants: What Each Catches

    Lexicon surveillance is a detective control. It flags listed words and phrases in communications after they are produced or delivered. Runtime policy enforcement acts in the execution path, deciding on each interaction and tool call before it completes. Lexicon review can surface known phrasing for human follow-up, but it acts late and sees only the text it is given. Runtime enforcement can allow, redact, block, rate limit, or escalate individual actions, including ones that never appear as keywords. The two address different failure modes, and many programs will run both.

    Criteria for evaluating runtime enforcement

    • Actions: does it support allow, redact, block, rate limit, and escalate for human approval, and at what granularity (per interaction or per tool call)?
    • Coverage: which models, agents, tools, and MCP servers sit in the path, and how is bypass handled?
    • Detection quality: how does it handle paraphrased language, context-dependent statements, sensitive data in prompts, and injection inputs that humans cannot see?
    • Evidence: what record is produced for each decision, how durable is it, and how is it retained and exported to auditors?
    • Deployment and residency: can it be self-managed in your cloud, and are retention and data residency configurable?
    • Reconciliation: how are its findings matched with existing surveillance alerts, and who owns escalation and suspension decisions?

    How each control works

    Lexicon-based compliance monitoring compares communications against a maintained list of keywords and phrases. When text matches, the system raises an alert and a reviewer decides whether a violation occurred. Compliance teams have used this model for years on email and chat, and it remains a legitimate detective control. It is simple to explain, and reviewers can tune the lists as new concerns appear.

    Runtime policy enforcement for AI assistants works differently. Policy is evaluated in the execution path, at the moment an assistant receives a prompt, produces a response, or calls a tool. The policy outcome can be to allow, redact, block, rate limit, or escalate for human approval. The unit of control is the individual interaction or tool call, not a batch of messages reviewed later.

    The first control asks whether known language appeared in a record. The second asks whether this specific action is permitted right now. That difference drives most of the contrasts below.

    Side-by-side comparison

    The table compares the two approaches on the dimensions most relevant to advisor AI assistant programs.

    • Detection mechanism: Lexicon: matches listed words and phrases. Runtime: evaluates policy against the interaction or tool call, including its context and the action requested.
    • Timing: Lexicon: acts after content is produced or delivered. Runtime: acts before the interaction completes.
    • Scope: Lexicon: the text that is fed into the review system. Runtime: individual interactions and tool calls that pass through the execution path.
    • Available responses: Lexicon: alert, review, escalate, remediate afterward. Runtime: allow, redact, block, rate limit, escalate for human approval.
    • Typical failure mode: Lexicon: paraphrase, context-dependent statements, and content never ingested. Runtime: policy gaps or misconfigured rules, and traffic that bypasses the path.
    • Evidence output: Lexicon: flagged items and reviewer dispositions. Runtime: a record of each policy decision and the action taken.

    What each control catches and misses by risk category

    Map risks one category at a time. Treat any mapping as a hypothesis to test in your own environment, because results depend on configuration and on what each tool can see.

    Prohibited statements are where lexicon review is strongest, when the language is predictable. A phrase that is on the list will be flagged. The weakness follows from the mechanism. A list matches what it contains, so a paraphrase that carries the same meaning in different words can pass unflagged. A statement that is only problematic given the client's situation also has no keyword to match. AI assistants generate fluent variations of the same idea, which makes fixed lists a weaker fit than they are for human-written messages.

    Sensitive data in prompts is a different problem. The exposure occurs when an advisor submits the prompt, and review after the fact cannot undo it. Enforcement in the path can redact or block the data before it reaches a model. Lexicon review may help only if prompts are in the monitored feed and the data follows a pattern someone listed.

    Unauthorized tool or data access is largely outside what keywords describe. A tool call is an action with parameters, not a communication. Policy in the execution path can allow, block, rate limit, or require human approval for that call. A keyword list has little to say about whether an action was permitted.

    Prompt injection adds a further case. It is defined as input that alters a model's behavior in unintended ways, and injected content need not be visible or readable to humans as long as the model parses it. Review that depends on human-readable text may not surface it. Neither control should be assumed to stop it, so use injection scenarios as a test set for both.

    Timing, regulation, and what the evidence shows

    Timing is the clearest operational difference. After-the-fact review can identify and document a problem, but the content has already been produced or delivered. Enforcement before an interaction completes can prevent the outcome, and it records the decision it made.

    Evidence follows from that. A lexicon system produces alerts and the notes of reviewers who handled them. That shows what was flagged and what people concluded. It does not show what was prevented, and it does not show interactions that never matched a term. Runtime enforcement produces a record tied to each decision, so a reviewer can see what policy applied and what happened. Trussed generates continuous audit-ready evidence from enforcement, mapped to HIPAA, NAIC, NYDFS, EU AI Act, and NIST, with configurable retention and data residency. Whether any evidence set meets a specific recordkeeping obligation is a question for your counsel and regulators.

    On the EU side, as of November 2026, the AI Act has been amended and its high-risk rules are expected to apply from 2 December 2027 for stand-alone high-risk systems and 2 August 2028 for those embedded in products. Confirm first whether a given advisor use case is classified as high-risk. If it is, deployers must monitor operation according to the instructions for use. If they have reason to consider that use may present a risk, they must inform the provider or distributor and the market surveillance authority without undue delay, and suspend use.

    That suspension duty is a system-level response. It is not per-interaction enforcement, so treat them as separate layers. Name the owners of monitoring, incident reporting, and suspension decisions in advance.

    Using both controls together

    Whether the controls are complementary is a design decision, not a regulatory finding. A common arrangement keeps lexicon surveillance for what it does well and places runtime enforcement where prevention matters.

    Under that arrangement, lexicon review continues to cover the communication archive, including human-written messages and known phrasing in AI outputs. Runtime enforcement covers the assistant's prompts, responses, and tool calls, where actions can be allowed, redacted, blocked, rate limited, or escalated for approval. Escalations give human reviewers a decision to make before something is sent, rather than a record to read afterward.

    The practical work is reconciliation. A lexicon hit on content that enforcement already redacted should connect to the enforcement record. A pattern in enforcement decisions can also suggest new terms for the lexicon. Decide in advance which system is the record of truth for each interaction, so reviewers do not work the same event twice.

    The limits of the runtime approach also deserve a plain statement. It protects only traffic that passes through the execution path, and it is only as good as the policy written for it. It does not replace review of communications outside that path.

    Two controls, two points in time

    Lexicon surveillance

    Matches listed terms in communications after the fact and routes hits to reviewers.

    Runtime policy enforcement

    Applies policy in the execution path before an interaction or tool call completes.

    Evidence

    Alerts and review outcomes versus a decision record for each enforced action.

    Relationship

    A design choice: detective review alongside preventive enforcement.

    Map your advisor AI risks to the right control

    Walk through how policy is enforced in the execution path for individual interactions and tool calls, and what evidence that enforcement produces.

    Request a Demo