1. Open manivox.ai and start the interactive demo.
2. Choose a voice, then French, English or Spanish.
3. Enter your details: the agent will greet you by your first name.
4. Call the number shown, and ask for an appointment.

• It answered straight away, and greeted you by name.
• You could interrupt it, and it stopped to listen.
• It told you it was an AI, and it still felt natural.
• It did something: the appointment was confirmed on your screen.

• Busy hours: the phone rings, and nobody is free to pick up.
• Lunchtime and after hours: nobody answers at all.
• Voicemail: the caller has to wait for a callback. How many hang up instead? Nobody counts.
• The caller tries the next business on the list.

General practitioners in the Paris region spend about seven hours a week on the phone, and 45% of the calls they receive are appointment requests (URPS Île-de-France, survey of 3,734 calls, 2010).

• Practice: an empty slot in tomorrow’s diary, and a patient who has found another practice.
• Garage: a service booked with the competitor across the road.
• Hotel: a room booked through a booking site, commission included.
• Social landlord: a leak reported the next morning, when the flat below is already wet.

A Manivox agent answers every call, understands the request and handles it, in the caller’s language. It replies with low latency, inside a normal conversational pause.

In conversation, people reply to each other after a gap of about 200 ms, in every language studied (Stivers et al., PNAS, 2009). A silence of about one second is the most we tolerate before it feels like a problem (Jefferson, 1989). A voice agent has to reply inside that window, or the conversation breaks.

• They can say it in their own words: no menu to learn.
• They can interrupt, and the agent stops to listen.
• They get an answer or a booking, not a promise of a callback.
• They are told they are talking to an AI, and still reach a person when they ask.

• Ears: speech becomes text while the caller is still talking.
• Brain: understands what the caller wants and decides what to do next.
• Voice: a natural voice by Cartesia, the world leader in natural speech.
• Memory: the knowledge base, with the client’s documents, price lists and opening hours.
• Hands: connections to the calendar, the CRM or the ticketing system.

The agent does not improvise the job. It follows a workflow the business has approved: greet, understand the reason for the call, check the diary, book, confirm. Inside each step, it talks freely.

• Answer every call, at any hour.
• Understand free speech in 42 languages, and switch language when the caller does.
• Look up answers in the client’s knowledge base.
• Act: book, cancel, update a record or open a ticket in the client’s systems.
• Transfer to a person when the caller asks, or when the case needs one.

One agent can speak several languages. It follows the caller’s language during the call, with its own voice, speed and messages for each language.

• Full support, 8 languages: French, English, Spanish, German, Italian, Dutch, Polish and Portuguese. Relative dates (“next Tuesday”), numbers said in words and spelled-out email addresses are understood natively.
• Standard support, 34 languages: the full conversation and multilingual agents. Relative dates, numbers said in words and spelled-out addresses can be understood less precisely.

Arabic · Bengali · Bulgarian · Chinese · Croatian · Czech · Danish · Finnish · Greek · Gujarati · Hebrew · Hindi · Hungarian · Indonesian · Japanese · Kannada · Korean · Malay · Malayalam · Marathi · Norwegian · Punjabi · Romanian · Russian · Slovak · Swedish · Tagalog · Tamil · Telugu · Thai · Turkish · Ukrainian · Urdu · Vietnamese.

Hebrew and Urdu are understood, and their voices are being activated: they are not offered in the editor yet.

• Cold calling French consumers without their prior consent, required since 11 August 2026. Business-to-business calls are outside that law.
• An agent that hides that it is an AI, or pretends to be a real person.
• Medical advice or triage, and emergency or crisis lines. The agent can pass an emergency on, never handle it.
• Decisions about people, such as recruitment, creditworthiness or access to essential services, without Manivox’s written agreement.
• Manipulation, hard-sell debt collection, emotion recognition and voice cloning.

Today: the phone rings and nobody can pick up. Voicemail. The patient calls the next practice on the list.

With Manivox: the agent answers straight away, finds a free slot in the diary and books it. The assistant finds the appointment already in the calendar.

Today: the phone rings in an empty reception. A customer wants to book a service and asks whether their car is ready. Someone wipes their hands and walks over, or nobody does.

With Manivox: the agent books the service in the workshop calendar and reads the repair status from the workshop system. The mechanics keep working.

Today: one night porter, a queue at the desk, and a guest who does not speak the local language.

With Manivox: the agent answers in German, explains late check-in, and leaves a message in Reception for the morning team. If the guest switches to English, so does the agent.

1. A telecom operator. Manifone is a declared telecom operator: it provides the phone numbers and runs the telephony itself, not a layer on top of someone else’s.
2. Hosted in France. The core of the platform runs in France on infrastructure Manifone owns: the language models, the main voice engine, the knowledge base and the telephony. Speech recognition runs in the EU, with a US-based provider.
3. Replies with low latency, and the response time of every reply is visible in the call detail.
4. Voices by Cartesia, the world leader in natural speech, with automatic language switching during the call.

• Developer platforms (Vapi, Retell): powerful building blocks for teams that write code and assemble their own stack.
• Phone-system add-ons (Aircall, Ringover, Allo): simple to switch on, focused on answering and taking messages.
• Enterprise contact-centre suites (Cognigy, Parloa, Zaion): built for large contact centres, with long projects and large budgets.
• Manivox: a complete agent a partner can build, test and run for a small or mid-sized business, on a phone network we operate.

64% of customers say they would rather companies did not use AI in customer service, and their top fear is not being able to reach a person (Gartner, 2024). So every agent offers a way to reach a person, and says so.

It answers from the client’s own knowledge base. KB Doctor checks the documents for contradictions, placeholders and missing basics such as opening hours. Agent Lab tests the agent with difficult callers before it goes live. Every real call is transcribed, so the client can read exactly what was said.

• The core of the platform is hosted in France. Speech recognition runs in the EU with a US-based provider, and the contract says so plainly.
• The business is the data controller, its partner the processor, Manivox the sub-processor. Every further sub-processor is listed in the client portal, or in the partner’s contract.
• The greeting says it is an AI, as the EU AI Act has required since 2 August 2026.
• Call data is never used to train AI models. There is no voice biometrics and no emotion recognition.
• Recordings are deleted after 7 days by default. They can be switched off for each agent, or sent straight to the business’s own storage.

Telling them is mandatory, so the question is how. For service calls (booking, status, information), a short and useful greeting works: “I’m the practice’s AI assistant. I can book you in right now.”

It takes the repetitive calls, the peaks, lunchtime and the nights. Staff keep the calls that need them, and start each day with the overnight messages already waiting for them.

The agent is built in a visual editor: the steps of the call, the routes between them and the tools each step may use. Calendars and CRMs connect through ready-made app connectors (Google Calendar, HubSpot, Salesforce and many more) or through the client’s own web services. The partner sets it up, and the business reviews and approves it before it goes live.

• Your client is the data controller: it decides why calls are handled.
• You are its processor: you design and run the agent for it.
• Manivox is your sub-processor: it provides the infrastructure.
• Your own agents, such as your reception or sales line, make you the controller and Manivox your direct processor.
• The account an agent sits in, yours or a client’s, decides which case applies.

• Accepting the Terms of Use also accepts the data processing agreement. A signed confirmation copy can be generated from the Console at any time.
• Manivox brand: you sell under the Manivox brand, with a licence to use it.
• White label: you sell under your own brand. In the Console and the client portal, only the logo changes; the portal stays on manivox.ai.
• For an interface fully under your own name and domain, build your own on the Manivox API.
• Switching between the two takes 60 days’ notice.

• Manivox brand: Manivox is the provider of the AI system; you and your client are deployers.
• White label: you become the provider, because the system is sold under your name. Manivox gives you the information you need about its capabilities and limits.
• In both cases, the greeting of every agent must say it is an AI. Setting it up is your job.
• Machine-readable marking of the synthetic voice is Manivox’s job, within what telephony allows. Never switch it off.

• A contract with data protection terms equivalent to the ones you accepted.
• The list of further sub-processors. Clients who use the client portal see it there; for the others, put it in your contract.
• Notice that transcripts and recordings exist, and that you can access their portal with “view as account”.
• The rules on recording calls in their country, and what to tell callers.
• Any notice you receive from Manivox about a change of further sub-processor.

• Health or other sensitive data: flag the agent to Manivox, shorten the retention, keep the self-hosted language model, and never paste such data into Copilot.
• Your own tools (your own language model, a web service, an MCP server, an external knowledge base): declare each new third party in the Console, and have your own data processing agreement with it.
• Outbound prospecting: tick the consent attestation before each campaign, and keep the proof of every consent yourself.
• The greeting says it is an AI, and the recording notice fits the country.

• In France, on infrastructure Manifone owns: language models, main voice engine, knowledge base, databases, telephony, client portal.
• Speech recognition: a US company, processing in the EU, storing nothing.
• Backup voice engine: used only during an incident on the main one, and it receives only the text of the agent’s replies, never the caller’s voice.
• App connectors: a US integration platform with no EU hosting. No data goes there unless you set up a connector.
• Copilot: an EU endpoint, and only what your team submits while building. Never live call data.

• Call data is never used to train AI models. No voice biometrics, no emotion recognition, no voice cloning.
• A data breach is notified to you within 48 hours, in phases, so your client can meet its own 72-hour deadline.
• 99.5% monthly availability. Support picks up a ticket within 1, 4 or 8 hours depending on severity, counted in French working hours (Monday to Friday, 9:00 to 18:00).
• Recordings are deleted after 7 days and transcripts after 365 days by default, and everything is deleted or returned within 30 days after the contract ends.

1. How many calls a day, and how many go unanswered or to voicemail?
2. What are the top three reasons people call?
3. What is one missed call worth: a booking, a patient, a repair?

4. Who answers after hours and at lunchtime?
5. Which calendar, CRM or practice software do you use?
6. Who decides, and what would success look like in 30 days?

• Build a demo agent from the prospect’s website (see “Build your first demo”).
• In the second meeting, call it in front of them, on speaker.
• Leave with one agreed measure for the first 30 days.

Starter, Connected and Advanced packages, at fixed monthly prices. Manivox bills you per minute of conversation, all-in (speech recognition, language model, voice, knowledge base and recording), with no per-seat fees. Phone numbers, outbound calls and transfers are billed separately. The package prices go here once they are decided.

The canvas shows the call as a chain of nodes: the Start node greets, Agent nodes handle each topic, a Router sends the caller to the right one, Transfer and Hangup end the call. This is the brain’s job description.

Each node lists its exit routes: “the caller wants to book”, “the caller asks for a person”. A classifier picks the route from what the caller says; rules pick it from a variable or a tool’s result. Build course: lesson 7.

Documents, price lists and web pages, in one place, attached to the whole agent or to one node. KB Doctor checks them after each upload; “Run quality check” runs it again on demand.

Web Services and App connectors: this is where the agent gets access to the calendar, the CRM or the ticketing system, one tool at a time, on the node that needs it.

An in-browser test call runs the version you are editing, or the live version. It is the fastest way to hear a change, and it works before the agent has a phone number.

Test callers, including difficult ones, call the agent and check the result. When a test fails, Agent Lab suggests a fix. Tests never touch the client’s live systems.

Order a number for the client’s account (once the account’s identity check is approved) and assign it to the workflow. “Make live” makes the version you see the one callers hear.

The call history holds every call, with its transcript, the steps it went through, the variables it collected, a summary, the recording and the response time of each reply.

Calls, duration and completion, when calls come in, response times, and the quality score against the client’s goals.

One account per client, with its own agents, numbers and statistics. The client logs in to the portal, with your logo, and sees their calls and messages in Reception. The portal stays on manivox.ai, even in white label.

1. In your demo account, create a new agent workflow. In the Initialization tab, choose the language, the model and a voice.
2. On the Start node, write a welcome message that says it is an AI, and a prompt with the prospect’s services and opening hours, taken from their website.
3. Add a Hangup node and an exit route “the caller is done”, and connect them.
4. Save, and try it with an in-browser test call.
5. In the meeting, run the in-browser test call from your laptop, on speaker.

You also get a demo agent that presents your offer on your own number. Start from the template and adapt it.

One demo agent you built, reviewed by Manivox. Integrator partners go further with the capstone of the Build course: a complete agent built from a brief in their training account.