> ## Documentation Index
> Fetch the complete documentation index at: https://docs.orbit.devotel.io/llms.txt
> Use this file to discover all available pages before exploring further.

# FCC 24-17 — written consent for AI voices

> What the FCC's 24-17 ruling changed for AI-generated and cloned voices, how Orbit classifies the voice on your outbound calls, why the consent gate fails closed, and how to keep synthetic-voice campaigns inside the rule.

# FCC 24-17 — written consent for AI voices

If you run outbound voice with an AI agent, text-to-speech playback, or a
cloned voice, US federal law demands **prior express written consent** from
the recipient. Orbit enforces that bar on every outbound leg before the call
is placed, and blocks with `422 FCC_AI_VOICE_WRITTEN_CONSENT_REQUIRED` when
you haven't recorded it. This page explains what the rule changed, which
call-metadata signals trigger the gate, how it orders with the other
build-time guards, and how a legitimate synthetic-voice campaign complies.

## What FCC 24-17 changed

The FCC's February 2024 declaratory ruling (DA 24-17) put AI-generated
voices under the TCPA's "artificial or prerecorded voice" rule —
[47 CFR § 64.1200(a)(1)(iii)](https://www.ecfr.gov/current/title-47/chapter-I/subchapter-B/part-64/subpart-B/section-64.1200).
Two consequences follow for outbound callers:

* **Synthetic means artificial.** Text-to-speech of any kind — including a
  Devotel TTS voice, and including voices from third-party providers — is an
  "artificial voice" under the rule.
* **Cloned means artificial too.** A voice model trained to mimic a specific
  person still qualifies; the ruling attaches the same bar to both kinds.

The exposure is statutory: **$500–$1,500 per non-compliant call**
(47 U.S.C. § 227(b)(3)). For marketing calls under the TCPA, the prior
express **written** consent bar applies — not merely verbal or implied.

Orbit's role, consistent with the wider compliance model ([compliance
posture](/compliance/posture-overview)): the platform classifies the voice
on each outbound leg and enforces a consent check against **your** consent
records. The obligation to obtain valid written consent is the tenant's.

## Classification: how Orbit decides which voice a call uses

At send time the gate reads the call's request `metadata` and resolves one
of three content types — `human`, `synthetic`, or `cloned` — using this
precedence, first match wins:

| Priority | Metadata signal                                      | Verdict it yields    |
| -------- | ---------------------------------------------------- | -------------------- |
| 1        | `voice_content_type` in `{human, synthetic, cloned}` | that value, verbatim |
| 2        | `voice_cloned_model_id` (string)                     | `cloned`             |
| 3        | `voice_agent_id` (string)                            | `synthetic`          |
| 4        | `agent_id` (string)                                  | `synthetic`          |
| 5        | `tts_text` / `tts_body` (string)                     | `synthetic`          |
| —        | none of the above                                    | `human`              |

Voice campaigns stamp these keys for you — the outbound campaign path
writes `voice_agent_id` on AI-voice legs, and text-to-speech sends stamp
`tts_text` / `tts_body`. The `agent_id` fallback keeps direct-API callers
inside the gate. Only when none of those signals is present does the call
count as human (click-to-call, bridged agent, conference).

Worked examples (input metadata → verdict):

| Input metadata                           | Verdict     | Why                                                           |
| ---------------------------------------- | ----------- | ------------------------------------------------------------- |
| `{}`                                     | `human`     | no signal keys                                                |
| `{ "voice_content_type": "synthetic" }`  | `synthetic` | explicit override                                             |
| `{ "voice_content_type": "cloned" }`     | `cloned`    | explicit override                                             |
| `{ "voice_cloned_model_id": "vcm_123" }` | `cloned`    | cloned model                                                  |
| `{ "voice_agent_id": "ag_456" }`         | `synthetic` | AI voice agent                                                |
| `{ "agent_id": "ag_legacy" }`            | `synthetic` | fallback key                                                  |
| `{ "tts_text": "Hello…" }`               | `synthetic` | TTS playback                                                  |
| `{ "tts_body": "Hello…" }`               | `synthetic` | TTS playback                                                  |
| `{ "voice_agent_id": 12345 }`            | `human`     | non-string signal is ignored                                  |
| `{ "voice_content_type": "robot" }`      | `human`     | unknown value — blocking on free text would be false blocking |

A `human` verdict short-circuits the consent check entirely — no database
read, no block.

## Where the gate runs

The gate sits **before any paid or media work**:

1. It runs **after** the federal dialing-window guard — the one
   platform-global control that stops calls outside 8 AM–9 PM recipient-
   local time ([the federal voice guard](/concepts/tcpa-federal-voice-guard)).
   That guard always runs first; this one layers on top of it and never
   relaxes its outputs.
2. It runs **before** the send claim takes a billing hold, **before** any
   media room is created, and **before** the leg is dispatched to the voice
   fabric.

That ordering is the point: a blocked consent call never burns a billing
hold, never opens a media path, and never interacts with the disposition
lock that would otherwise record a half-routed call. Human-voice calls are
unaffected because the classifier returns `human` without a consent check.

## Failure posture: fail closed

The consent lookup blocks on two failure shapes, not just on "no consent":

* **Consent schema missing.** A workspace whose consent records table was
  never provisioned (older tenants) blocks — migration instructions included
  in the returned error.
* **Lookup error.** Any non-`does not exist` database error on the
  existence probe is captured and blocks, same shape.

Fail-open would silently treat an AI voice as compliant whenever the
database hiccups. At $500–$1,500 per call of statutory exposure, that risk
is the platform's to take on the block side, not yours to wave on the
allow side. The tradeoff is a blocked call (fixable) versus a class-action
discovery request you have to answer.

## Consent evidence queue

When a block fires, the event lands in your audit log as a
`voice.fcc_ai_consent_blocked` entry — see below. From there the entry shows
up as a consent-evidence gap you can chase and close in your own audit
exports: the recipient, the voice type the classifier saw, and the reason
(`no_written_consent_on_record`, `tenant_consent_schema_missing`, or
`consent_lookup_error`). Process those entries the way you process any other
consent-evidence gap, through [the consent register](/compliance/consent-management).

## Audit trail

Every block writes a `voice.fcc_ai_consent_blocked` audit row to the
append-only, tamper-evident chain before the error is thrown, so the
regulatory evidence survives whatever happens to the request upstream.
The row carries the classification verdict, a redacted recipient number,
the reason, and the campaign/agent id when those were stamped. Allow-path
calls do not write — the usual voice-call audit downstream covers granted-
consent calls. Filter on `voice.fcc_ai_consent_blocked` in your [audit
log](/guides/audit-log) to work the queue.

## Complying with a legitimate synthetic-voice campaign

1. **Obtain prior express written consent** for every recipient under
   [the consent register](/compliance/consent-management) — either under
   consent type `fcc_24_17_written_consent`, or with a free-text type plus
   a `lawful_basis` entry of `fcc_24_17_written_consent` in the record's
   metadata.
2. **Stamp it in the right shape.** The gate accepts either marker on a
   granted, non-revoked record; both unblock a `synthetic` or `cloned`
   call for that recipient.
3. **If you're blocked, the error tells you which.** `422
   FCC_AI_VOICE_WRITTEN_CONSENT_REQUIRED` carries the reason and a hint
   describing the exact consent row shape to add. Human-voice paths are
   outside the gate entirely.

## Tenant ownership

Recording and holding valid written consent is **your** obligation — Orbit
classifies, enforces, and audits; it does not obtain consent for you. This
page is not legal advice; run the posture by qualified counsel, same as any
US voice-consent question.

## See also

* [Compliance posture overview](/compliance/posture-overview) — where this gate sits in the tenant-versus-platform split
* [EU AI Act transparency](/compliance/eu-ai-act) — the sibling page: that one covers **AI disclosure**, this one covers **US voice consent** — different obligations, different pages
* [TCPA federal voice guard](/concepts/tcpa-federal-voice-guard) — the 8 AM–9 PM recipient-local window this gate layers on top of
* [Consent management](/compliance/consent-management) — where to record the written consent row the gate reads