Signal Guard
Signal Guard—built-in guardrails for AI-generated replies. Catches prohibited phrases, flags compliance risks, and routes expert-knowledge requests to humans.
Overview
Signal Guard protects your business from AI replies that could create legal, medical, or compliance risk. Every AI-generated reply is checked before it reaches a customer:
- Quick safety check (instant): scans for prohibited phrases and escalation keywords
- AI review (takes about half a second): a separate AI evaluates the reply against your compliance rules
The combined result produces a recommendation: send, review, escalate, or block.
Dashboard tabs
Overview tab
Shows at a glance:
- Stats cards: total evaluations, passed, reviewed, escalated, blocked
- Active preset: which industry Signal Guard preset is loaded (for example, Medical Aesthetics, Insurance)
- AI Judge toggle: enable or disable the LLM-as-judge second pass
- Rule counts: number of prohibited phrases, escalation keywords, and compliance rules active
Rules tab
View and customize Signal Guard rules:
- Custom Rules: additional rules written in plain English that the AI checks every reply against. Add requirements specific to your business (for example, "Never mention competitor clinics by name," "Always recommend in-person consultation for pricing questions").
- Prohibited Phrases: patterns that trigger immediate violations. Shown as badges from the industry preset plus any custom additions.
- Escalation Keywords: patterns that flag messages for expert review (for example, "lawsuit," "allergic reaction," "malpractice").
- Compliance Rules: industry-specific requirements the AI judge checks against (for example, "Never guarantee specific medical outcomes").
Custom rules merge with (not replace) your industry preset. You can't accidentally disable industry-standard protections.
Audit log tab
A compliance audit trail of every Signal Guard evaluation:
- Event ID: links to the signal that triggered the evaluation
- Recommendation: color-coded badge (send/review/escalate/block)
- Violations: which prohibited phrases were matched
- Judge: whether the LLM judge ran and its verdict
- Date: when the evaluation occurred
Export the audit log as CSV or JSON for compliance reporting.
How the two-stage pipeline works
- Quick safety check (always runs): scans for prohibited phrases and escalation keywords. If a prohibited phrase is found, the reply is immediately flagged.
- AI review (if enabled): evaluates the reply against your compliance rules and checks if the topic needs a licensed professional.
- Final decision: the more cautious result wins. The four possible outcomes are:
- Send—safe to send automatically
- Review—held for your review before sending
- Escalate—routed to a team member or specialist
- Block—not sent under any circumstances
How decisions are ranked
When the two checks disagree, Signal Guard always chooses the safer option:
| Priority | Recommendation | Meaning |
|---|---|---|
| 1 (lowest) | send | Safe to auto-send |
| 2 | review | Flag for human review |
| 3 | escalate | Route to expert |
| 4 (highest) | block | Never send |
Signal Guard-to-routing integration
When Signal Guard flags a reply as needing escalation or blocking, your routing rules get a chance to respond. This lets you create rules like:
IF _requires_expert = true
THEN escalate: reason "guardrail_expert_required"
create_task: priority high, reason "expert_knowledge_needed"Compliance-sensitive industry presets (medical, insurance, legal, financial, dental) include a default Signal Guard escalation rule automatically.
Industry presets
Five industry Signal Guard presets are available:
| Industry | Prohibited Phrases | Escalation Keywords | Compliance Rules |
|---|---|---|---|
| Medical Aesthetics | diagnosing, prescribing, guaranteed results, cure, permanent | allergic reaction, infection, malpractice | No medical advice, recommend consultation |
| Insurance | guaranteed coverage, binding quote | lawsuit, fraud, bad faith | No binding commitments, refer to agent |
| Legal | legal advice, guaranteed outcome | malpractice, negligence | No legal counsel, refer to attorney |
| Financial | guaranteed return, risk-free | SEC, fiduciary breach | No investment advice, include disclaimers |
| General | (minimal) | (minimal) | Professional communication standards |
AI judge details
The AI review uses a fast, cost-efficient AI model (about half a second per check). It:
- Checks the draft reply against all your compliance rules (preset + custom)
- Identifies when the topic needs a licensed professional
- Explains exactly what was flagged and why
- If the AI review is unavailable, the quick safety check still runs on its own
The AI review uses the same AI provider as reply generation. If you use your own AI key (BYOK), review checks count toward your AI usage. On platform-managed plans, the review is included in your AI signal cost.
Tips
- Start with your industry preset. It covers the most common compliance risks.
- Add custom principles for business-specific rules the preset doesn't cover.
- Keep the AI judge enabled for compliance-sensitive industries. Pattern matching catches literal phrases but misses semantic violations like "based on your symptoms, you probably have..."
- Review the audit log regularly. It shows patterns in what's being flagged.
- Export audit data for compliance documentation or regulatory review.
- Test with the Test Runner. Paste a message that should trigger Signal Guard and verify the evaluation result.