Short version. The receptionist demo on this site uses AI-generated voices and AI-generated translations. The receptionist product we sell is also AI — it answers calls and books appointments on a business's behalf using a large language model and a synthetic voice. Both the demo and the product are clearly identified as AI. Customers who deploy our receptionist must tell their callers they are speaking to AI; we provide guidance below.
1. AI in the demo on this site
The "receptionist demo" embedded on the homepage and feature pages is a marketing illustration, not a live phone call. It plays back pre-rendered audio clips of three scripted scenarios: a real-estate brokerage (Ava), a roofing contractor (Cole), and a luxury flooring boutique (Nina).
- Voices are generated using ElevenLabs's text-to-speech models from a fixed script. They are not recordings of real people, and they are not generated in real time — they are baked into static MP3 files at build time.
- Translations into Spanish, Mandarin (Simplified), and Arabic were drafted with the help of a large language model, reviewed by a human i18n auditor, then rendered to audio. The full audit trail is published alongside this site in
audits/translation-review-{es,zh,ar}.md.
- Subtitles and on-screen copy on the demo are written by us. They are not generated at runtime.
Nothing about the demo collects, processes, or learns from your voice. The site does not have microphone access and never asks for it.
2. AI in the receptionist product
The AI receptionist that we sell (separate from this site) is a real, live AI system. When a business deploys it:
- Inbound calls to a configured phone number are routed through a telephony provider to our system.
- The caller's speech is transcribed in real time using a speech-to-text model.
- A large language model reads the transcript, the business's playbook (hours, services, FAQ, calendar), and the conversation history, then generates a response.
- That response is rendered to speech with a synthetic voice (typically ElevenLabs or an equivalent), and played back to the caller.
- Booked appointments are written to the business's connected calendar (Google Calendar, Calendly, etc.).
There is no human in the loop on a per-call basis. Logs and transcripts are available to the business for review, retraining, and quality control.
3. How we identify AI to your callers
Customer obligation. If you deploy our receptionist for your business, you are responsible for disclosing to your callers that they are speaking to an automated AI system, in line with the laws of the jurisdictions you do business in. We design the product to make compliance easy, but the legal duty sits with you as the operator of the call.
3.1 Default disclosure script
By default, every receptionist persona begins the call with a disclosure equivalent to "Hi, this is an AI receptionist for [Business Name] — how can I help?" The exact phrasing is configurable per business but the AI-identification phrase cannot be disabled in our product UI.
3.2 Specific regulatory frameworks
| Jurisdiction | Rule | What our product does by default |
|---|
| California (USA) |
SB 1001 — bots interacting with Californians for commercial purposes must disclose they are not human. |
Self-identifies as an AI receptionist on every call. |
| Utah (USA) |
SB 149 (AI Policy Act) — regulated occupations must affirmatively disclose AI on consumer interaction. |
Self-identifies. Customers in regulated occupations should consult counsel for their specific disclosure language. |
| Colorado (USA) |
SB 205 (effective 2026) — duty to inform consumers when "high-risk" AI is used in a consequential decision. |
Self-identifies. If the receptionist will gate access to scheduling on protected classes (e.g., screening criteria), the business is responsible for the high-risk impact assessment. |
| European Union |
EU AI Act Art. 50 — providers and deployers of conversational AI must inform exposed natural persons they are interacting with AI, unless this is obvious. |
Self-identifies in the language the caller is dialing in. Multilingual disclosure phrases are configurable. |
| Telephone Consumer Protection Act (USA, federal) |
TCPA — restrictions on automated outbound calls. The FCC's 2024 declaratory ruling treats AI-generated voices as "artificial or prerecorded" voices. |
Our product is configured for inbound calls by default. If you enable outbound calling, you take on TCPA consent obligations and our admin UI walks you through them. |
| Federal Trade Commission (USA) |
Section 5 — prohibition on unfair and deceptive practices. Impersonating a human when a customer asks "are you a person?" is deceptive. |
If a caller asks any variant of "am I talking to a human?" the receptionist confirms it is AI. This response cannot be disabled. |
4. What our receptionist does not do
- It does not impersonate a specific real person. Personas (Ava, Cole, Nina, etc.) are clearly named characters, not clones of identifiable individuals.
- It does not record or store voiceprints for biometric identification.
- It does not train the foundation model on your callers' voices or transcripts. Logs are scoped to your business's tenant and used only to operate and improve the service for you. We do not feed customer-call data back into ElevenLabs model training; ElevenLabs's standard terms confirm this for our subscription tier.
- It does not make automated decisions with legal or similarly significant effect on the caller. It schedules appointments and answers FAQs; it does not approve or deny applications.
5. Limitations and known failure modes
You should know what an AI receptionist is bad at:
- Heavy accents and high background noise degrade transcription accuracy. We provide a confidence-threshold setting; calls below the threshold are forwarded to a human (you, or a configured forwarding number).
- Edge-case requests outside the playbook get a graceful "let me have someone follow up with you" response and a notification to you. The receptionist will not improvise legal, medical, or financial advice.
- Emergencies. If the caller indicates a life-safety emergency, the receptionist immediately tells them to hang up and dial the appropriate emergency number.
6. Bias, fairness, and content controls
The foundation models we use ship with general-purpose safety tuning. On top of that, we enforce:
- A refusal policy for discriminatory triage (e.g., declining to book based on a protected class).
- A refusal policy for impersonation of authority figures (police, judges, regulators).
- Profanity and abuse handling that de-escalates rather than mirrors.
7. Models and providers we use
Subject to change as the market evolves; we update this list when the underlying stack changes materially.
- Voice synthesis: ElevenLabs (primary).
- Language model: Anthropic Claude family (primary) and OpenAI GPT family (fallback).
- Speech-to-text: Deepgram and OpenAI Whisper (depending on latency profile).
- Translation: Anthropic Claude with human review.
8. Your rights as a caller
- Right to know. You will always be told you are talking to AI at the start of the call, and any time you ask.
- Right to a human. Ask to be transferred and the receptionist will route you to the business's configured human contact or take a message.
- Right to access and delete the transcript of your call. Email the business directly; if you cannot reach them, email us at lester@blankpages.app and we will forward your request.
9. Updates
We update this disclosure when the AI stack, our product behavior, or applicable law changes. Material changes are dated and flagged at the top of the page.
10. Contact
Questions about how we use AI: lester@blankpages.app.