Overview

Guardrails define boundaries for your agent's behavior. Control what agents can say and do to reduce risk and keep interactions predictable.

The Guardrails page controls what an agent can say and do during live calls, and how compliance obligations are handled. It has two areas. Guardrails sets behavioral boundaries on the agent. Compliance holds per-agent settings for opt-out handling, call evidence capture, and the call disclosures the platform adds to each call.

Contents

• Guardrails
• Compliance
  • Opt-out
• Call Evidence
• Disclosures

Guardrails

Each guardrail row shows its name, a tag for the part of the conversation it acts on, and a status badge on the right. Click a row to configure it.

Content Guardrails Toggles
Configure content moderation for this agent. Enabled categories are evaluated on the agent's responses by a lightweight classifier, honoring the sensitivity threshold, violation action, and execution mode below.

• Sexual: Blocks erotic or sexually explicit conversations.
• Violence: Blocks content that depicts acts of violence or physical harm.
• Harassment: Stops abusive or threatening behavior.
• Self Harm: Prevents engagement with self-harm or suicidal content.
• Profanity: Prevents use of vulgar or explicit language.
• Politics and religion: Avoids political and religious discussions.
• Medical and Legal Information: Prevents discussion about medical or legal topics.

Compliance

Opt-out handling covers opt-out requests and consent revocation, with reference to FCC 24-24. It is controlled by the Enable opt-out detection switch.

When the switch is on, the agent reacts as soon as a caller says an opt-out phrase or presses the opt-out key. It speaks a compliant closing line, ends the call, and writes an audit record of the revocation.

When the switch is off, the agent does not speak a closing line or end the call. A clear removal request is still recorded as an opt-out. On dialer calls, the call is marked Do Not Call and the number is added to your Do Not Call list.

Opt-out

Deterministic opt-out and consent-revocation handling (FCC 24-24). When a caller says an opt-out phrase — or presses the opt-out key — the agent immediately speaks a compliant close and ends the call, with an audit record of the revocation.

Enable opt-out detection
With this switch off, the agent does not speak a closing line or end the call — but when a caller clearly asks to be removed, the request is still recorded as an opt-out. On dialer calls the call is marked Do Not Call and the number is added to your Do Not Call list. Each opt-out check is billed like your other guardrail checks.

Extra opt-out phrases
These phrases still record an opt-out silently. The agent says nothing.

Semantic = the phrase anywhere in the caller's sentence; Exact = the whole utterance is the phrase. 0/24 used.

When opt-out detection is on

The agent acts as soon as it detects an opt-out. It speaks a compliant closing line, ends the call, and writes an audit record of the revocation. Turning the switch on reveals the following settings:

Call Evidence

Call Evidence captures detailed evidence on every call the agent places. This includes the exact system prompt, a lossless record of tool calls, UTC timestamps for compliance-relevant moments, and a log of guardrail triggers. Changes apply from the next call without a restart. Editing this section requires an admin role.

• Evidence capture
  Master switch. When on, this agent's calls are treated as compliance evidence: call-trace persistence is always on and the toggles below take effect. UTC capture-point timestamps (connect, first AI response vs. called-party greeting completion, opt-out detection and DNC recording) are stamped on every call regardless.
   
• Per-call prompt snapshot
  Store the exact rendered system prompt each call ran with — a SHA-256 fingerprint plus the retrievable text, captured after per-call variable substitution. Covers unpinned seats too, so every call is defensible even without a pinned agent version.
   
• Lossless tool-call & trace log
  Never drop an evidence write: the per-call trace writer (tool calls — DNC checks, CRM lookups, webhooks — turns, and stage events) switches to a lossless queue instead of dropping under burst load.
   
• Safety trigger log
  Record every guardrail hit (content-filter matches and script deviations, on caller input and agent output) into an exportable per-call log with the rule, category, action taken, and UTC time.

Disclosures

Disclosures are composed by the platform and spoken at the start of every call the agent places, with reference to 47 CFR § 64.1200(b). They play in this order:

1. Who is calling (the legal entity)
2. The opt-out announcement, for telemarketing calls only
3. The recording disclosure
4. A callback number

The platform injects this opening, so it cannot be removed or reworded through the agent’s prompt or greeting text. The only way to turn it off is the Call disclosures switch.

If required values are missing, a warning banner says disclosures are not configured. They start working once valid values are saved.

The Preview button at the bottom of the section plays the disclosure as it will open the call, in the agent’s voice. It uses the current settings, including changes you haven’t published yet.