​

EU AI Act — Article 50 Transparency

​

Overview

Not automatic geo-detection. Orbit does not infer a contact’s jurisdiction from a phone number, address, or IP — enabling a rule applies it workspace-wide to every AI interaction, it does not switch on or off per contact based on location. If your organization’s traffic touches the EU, enable the eu_ai_act_enabled rule for the whole workspace. The one exception is the California SB 243 minor reminder, which fires only for contacts you have explicitly flagged as a minor — Orbit never infers minor status from a birth date or other signal.

​

What ships today

1. Chat notice — a configurable notice (“You are chatting with an AI assistant…”) is prepended to an AI agent’s first reply in a conversation. It is sent once per conversation — it does not repeat on every turn.
2. Voice intro — a configurable spoken introduction (“This call is being handled by an AI agent…”) is played before an AI voice agent begins a call. You can supply a pre-recorded audio URL or let Orbit synthesize the text.
3. AI-generated content marking — messages sent by an AI agent are stamped with metadata.ai_generated: true. Recorded calls with any AI involvement are stamped with EU AI Act Article 50 provenance metadata on the archived recording. Where the configured text-to-speech provider embeds an inaudible content-provenance watermark, that support is recorded per call — watermarking coverage varies by TTS provider today, so treat it as an additional signal rather than a universal guarantee.
4. California SB 243 minor reminder — when enabled, contacts explicitly flagged as a minor get a periodic “you’re talking to an AI” reminder during long-running sessions, on a configurable interval.
5. Conformity dossier export — a one-click, per-agent evidence pack that maps the agent’s transparency configuration, evaluation/red-team results, human-oversight approvals, and decision audit log to EU AI Act Articles 12, 14, 15, and 50.
6. AI-disclosure ledger export — a privacy-preserving, tamper-evident, cryptographically signed ledger proving when and where an AI (not a human) handled an interaction, independently verifiable without trusting Orbit’s UI.

​

Configuring AI disclosure

​

Dashboard

​

API

GET /api/v1/compliance/ai-disclosure

{
  "data": {
    "settings": {
      "id": "aidisc_default",
      "defaultEnabled": true,
      "chatNoticeText": "You are chatting with an AI assistant…",
      "voiceIntroAudioUrl": null,
      "voiceIntroText": "This call is being handled by an AI agent…",
      "minorBreakReminderMinutes": 60,
      "marketingDisclosureRequired": true,
      "krDisclosureEnabled": false,
      "caMinorReminderEnabled": true,
      "euAiActEnabled": true
    }
  }
}

PUT /api/v1/compliance/ai-disclosure

{
  "default_enabled": true,
  "chat_notice_text": "You are chatting with an AI assistant.",
  "voice_intro_text": "This call is being handled by an AI voice agent.",
  "eu_ai_act_enabled": true
}

​

Verifying it’s working

1. Live check — with default_enabled and eu_ai_act_enabled on, start a new chat session with an AI agent and confirm the configured chat_notice_text appears on the first reply; place a test call to an AI voice agent and confirm the voice intro plays before the agent speaks.
2. Per-turn metadata — inspect a chat message sent by the agent and confirm metadata.ai_generated is true.
3. Conformity dossier — export the evidence pack for an agent:
   GET /api/v1/agents/{id}/conformity-dossier?format=json
   
   Add ?format=html for a downloadable, human-readable evidence pack suitable for an auditor or enterprise security questionnaire. Each section reports present, partial, absent, or unavailable — a legacy tenant missing a backing table degrades that one section rather than failing the whole export.
4. AI-disclosure ledger — export the signed, tamper-evident ledger for an agent:
   GET /api/v1/agents/{id}/ai-disclosure-ledger?format=json
   
   The ledger carries provenance metadata only (agent id, model, outcome, timestamps, aggregate counts) — never verbatim prompt/response text — and is signed with a detached content credential so a copy handed to a regulator or partner can be verified without trusting Orbit’s UI.

​

Remediating a degraded dossier section

​

Verifying the signed ledger

1. Export the JSON ledger (this is the payload to verify):
   GET /api/v1/agents/{id}/ai-disclosure-ledger?format=json
   
   The response wraps the ledger in data.ledger. Its shape: ledger_version, generated_at, tenant_id, agent, disclosure_posture, jurisdiction_resolution, summary, entries, note, disclaimer, and a detached signature object:
   {
     "signature": {
       "v": "ovl1",
       "alg": "HMAC-SHA256",
       "content_sha256": "…",
       "mac": "…"
     }
   }
   
2. Post the ledger to the verifier endpoint. The detached credential lives inside the JSON itself — nothing else to download:
   curl -X POST https://orbit.devotel.io/api/v1/agents/ai-disclosure-ledger/verify \
     -H "Authorization: Bearer <token>" \
     -H "Content-Type: application/json" \
     -d '{"ledger": <the ledger object>}'
   
   # → { "data": { "valid": true, "reason": "ok" } }
   
   reason is one of ok, content_mismatch (body edited after signing), signature_invalid (MAC does not match the content), unsupported_version (unknown credential version), or malformed_ledger.
3. Verify offline if you want the check independent of Orbit’s API: strip the signature field, serialize the remaining body as canonical JSON (object keys sorted recursively), and SHA-256 the result — it must equal signature.content_sha256. The MAC additionally requires the workspace’s signing key, so a tampered ledger can be detected by anyone, but producing the MAC is Orbit-side only.
   body   = ledger minus the "signature" field
   digest = SHA-256(canonical_json(body))   # keys sorted recursively
   assert digest == ledger.signature.content_sha256
   

​

Confirming the notice fired on a conversation

1. Read the conversation — on a text channel the chat notice is prepended to the agent’s first reply, so the notice text is visible verbatim in the conversation’s message list in the dashboard and via the conversations API. The same property holds for voice: the intro plays before the agent speaks, so a recorded call carries it at the start.
2. Check the per-conversation timestamp — once the chat notice is delivered, the platform stamps first_disclosure_sent_at on that conversation and the notice never repeats there. If you (or your compliance tooling) retain exports or warehouse snapshots of conversation records, a non-null value is durable proof the notice fired for that conversation.
3. Check the audit log for configuration changes — every change to the disclosure settings is recorded under the audit action compliance.ai_disclosure_updated, and each dossier or ledger export is logged as agent.conformity_dossier.exported or agent.ai_disclosure_ledger.exported. If a notice did not fire on a conversation you expected it to, the audit trail for those actions plus the conversation timestamp above isolates whether disclosure was off at the time, or the notice genuinely failed to send.

​

Posture matrix per regulation

​

Known limitations

• Jurisdiction rules are workspace-wide opt-ins, not automatic per-contact geolocation — see the note above.
• TTS-embedded audio watermarking depends on the configured voice provider; not every provider supports it as of mid-2026.
• This page describes the transparency (Article 50) surface. It does not cover the EU AI Act’s separate high-risk-system (Annex III) obligations — see the pointer below.

​

Annex III high-risk obligations