Voice Moderation and AI Disclosure

Overview

Two per-agent settings, both keys in the agent's config, set with PATCH /api/v1/bots/{bot_id}/config like any other agent setting:

• ai_disclosure speaks a short sentence before the greeting so callers know they are talking to an AI.
• moderation screens the call: what the caller says, what the agent says, and a list of words the agent must never speak.

Every finding is stored, listed by GET /api/v1/voice/moderation-events and sent to your webhooks as call.moderation_flagged.

Authentication

Authorization: Bearer cm_your_api_key

Changing an agent's config needs the bots:write scope. Listing findings needs bots:read.

AI disclosure

curl -X PATCH https://api.callmissed.com/api/v1/bots/{bot_id}/config \
  -H "Authorization: Bearer cm_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{ "values": { "ai_disclosure": { "enabled": true, "text": "Heads up, you are speaking with an AI assistant." } } }'

• The disclosure is spoken once per call, ahead of the recording notice when the call is recorded, whether the call opens on the greeting, a call menu or a queue hold.
• It is spoken as written, before anything else the agent says.
• New agents created from the starter template have it turned on. Existing agents are not changed; turn it off with { "unset": ["ai_disclosure"] } or "enabled": false.

Moderation

curl -X PATCH https://api.callmissed.com/api/v1/bots/{bot_id}/config \
  -H "Authorization: Bearer cm_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{ "values": { "moderation": {
        "enabled": true,
        "categories": ["hate", "violence"],
        "action": "warn",
        "severity_threshold": "medium",
        "blocklist": ["darn"]
      } } }'

A malformed setting (an unknown category or action, an oversized blocklist) is rejected with 422 and nothing is saved.

What the caller says

Each finished caller sentence is classified as the call goes on. When a category in categories is found at or above severity_threshold, the finding is stored and action runs:

On a call handed between agents (squads, call menus, call queues), each sentence is screened with the settings of the agent answering at that moment; the blocklist stays the one of the agent that answered the call.

Screening never slows the call down. If the classifier is slow, unavailable or not enabled for your workspace, the sentence is treated as not flagged and the call carries on. Moderation is a safeguard, not a guarantee that every harmful sentence is caught.

What the agent says

• Blocklist. A blocked word is replaced in the speech with blocklist_replacement before the caller hears it. Matching is whole-word and case-insensitive. The saved transcript keeps the agent's original words, and each rewrite is stored as a finding with category blocklist.
• Classification. Each finished agent message is also classified afterwards. Findings about the agent are log only: the call is never acted on for what the agent said.

List findings

curl "https://api.callmissed.com/api/v1/voice/moderation-events?direction=caller&limit=20" \
  -H "Authorization: Bearer cm_your_api_key"

[
  {
    "id": "7f0b3c1e-0d5a-4f29-9a5e-5b1c2f0d9e11",
    "session_id": "c3a1c2a0-61f4-4c8e-8a39-0f6b1f6c8a52",
    "bot_id": "2b7e6a53-27f0-4d2c-8d3e-1c0a4f9a7b10",
    "direction": "caller",
    "category": "hate",
    "severity": "high",
    "action": "warn",
    "excerpt": "...",
    "created_at": "2026-10-08T09:41:07.512Z"
  }
]

Newest first. severity is low, medium or high.

Excerpts and data retention

excerpt holds at most the first 200 characters of the flagged speech. When the agent, or your whole workspace, runs with zero data retention, the finding is kept but excerpt is null: you learn that a category was hit, never the words.

Webhook

Subscribe to call.moderation_flagged (webhooks) to be told as it happens. The payload carries the same fields as a listed finding, plus event_id:

{
  "event_id": "7f0b3c1e-0d5a-4f29-9a5e-5b1c2f0d9e11",
  "session_id": "c3a1c2a0-61f4-4c8e-8a39-0f6b1f6c8a52",
  "bot_id": "2b7e6a53-27f0-4d2c-8d3e-1c0a4f9a7b10",
  "direction": "caller",
  "category": "hate",
  "severity": "high",
  "action": "warn",
  "excerpt": null,
  "created_at": "2026-10-08T09:41:07.512Z"
}

Limits

• Moderation and AI disclosure apply to calls your agents take by phone, on the web and on WhatsApp. They do not apply to a managed voice agent WebSocket connection, where you set the opening line yourself.
• Content screening is off for your workspace until it is switched on for you; until then blocklist and ai_disclosure work and the other settings are stored without effect.