Building your AI voice agent
An AI phone receptionist wired into your tools
We build the agent that answers the phone, qualifies the call, writes into your calendar or your CRM, and hands over to a human when it should. Telephony, voice and model assembled by us, from scoping to monitoring.
- Senior product team
- voice agent in production
- latency and cost measured per call
What is a voice AI agent and when should you have one built?
A voice AI agent answers calls on behalf of your team, identifies the caller, understands their request and acts directly in your tools: booking an appointment, creating a ticket, updating a file. It does not replace a human advisor for complex cases, but it handles first-level calls with no waiting time. You have a custom voice agent built when inbound call volume is too high to handle manually, or when the phone journey needs to connect to a specific business application that an off-the-shelf solution cannot reach.
What our clients have a voice agent handle
Night and weekend front desk
Out-of-hours calls are answered, qualified and summarised instead of lost. The next morning your teams find a ticket with the transcript attached.
Booking and moving appointments
The agent reads your availability, books the slot in the calendar and sends the confirmation by SMS or email. Nothing to re-enter afterwards.
Inbound breakdown calls
Contract identified, severity assessed, the job created in your business tool, and escalation to the on-call engineer once the threshold is crossed.
Calling back quote requests
Incomplete forms get a call back, missing details are collected, and your sales team receives a qualified file instead of an empty line.
What it changes in day-to-day operations
Engineering in service of a measurable outcome: no more missed calls, less data entry, call reasons you can finally measure.
No more calls falling into the void
Nights, weekends and traffic peaks are covered. For a business that lives on appointments, that is revenue captured, not a headcount saving.
Human time goes to the calls that matter
The first tier goes to the agent. What stays human arrives with the context already collected, the contract identified and the reason written down.
Data entry disappears during the call
The agent writes straight into the calendar, the CRM or the ERP. The gap between what the customer asked and what is recorded drops to zero.
Call reasons become actual data
Every call produces a structured transcript. You finally know why people call you, in what proportion, and what deserves fixing upstream.
How we ship your voice agent
Scoping
Inbound or outbound, type of number, volumes, cases covered and cases excluded. The handover rule and the legal frame are settled before coding.
Call prototype
A test number you can dial in French from the second week, with the AI disclosure, one real intent and the handover to a human already wired in.
Measure and tune
Replay of a corpus of French calls, including noise, accents and spelling out. Tuning of interruption handling, latency and cost per call.
Monitoring
Latency per turn, human handover rate, cost per call and timestamped transcripts on a dashboard, with a daily spending cap in place.
What a voice agent can do today
- Real-time conversation
- The OpenAI realtime API connects over WebRTC for browser and mobile, over WebSocket when your server already receives the audio, over SIP for telephony.
- Telephony and call transfer
- An inbound call fires a webhook. The agent accepts, rejects with a SIP code, transfers to a number or hangs up. Transfer is a primitive, not a workaround.
- Handling interruptions
- Twilio ConversationRelay tunes interruptibility, sensitivity and backchannel filtering of “yes, right”. That is what separates a pleasant agent from an IVR.
- End-of-turn detection
- The LiveKit turn detector encodes the audio itself to catch intonation, covers French, and lowers the default endpointing delay to 0.3 second.
The vocabulary of a voice agent
- Endpointing
- Detecting when the caller has finished speaking. It is the first link in perceived latency: too short and the agent talks over people, too long and it feels slow. LiveKit lowers its default delay to 0.3 second, with a 2.5 second ceiling.
- Barge-in
- The caller's ability to cut in while the agent is speaking. Twilio does not expose this as a boolean but as degrees, and filters out backchannel “mhm yes” so the agent does not keep cutting itself off on acknowledgements.
- refer
- The call transfer operation towards a target, a phone number for instance. It is the primitive that makes the handover to a human clean, with the context already collected, instead of a “please call back later”.
- Context replay
- On a realtime API, every turn sends the conversation history back to the model. A badly designed ten-minute call costs several times a well-designed ten-minute call, within a 32,000 token window.
- Article 50 of the AI Act
- The obligation to inform a person that they are interacting with an AI system, unless it is contextually obvious. It covers callbots and voicebots, with penalties announced up to 15 million euros or 3 % of worldwide turnover.
- NPV number
- A number category required in France for automated outbound calling and sales. Twilio explicitly forbids those uses from a French geographic or mobile number, and requires a K-bis, a SIREN and a French address.
The real constraints of a voice agent
Latency is the product, not an option
It adds up: end of speech, transcription, model, synthesis, network, gateway. A single slow link is enough to make the caller hang up. That is why the OpenAI documentation recommends a low reasoning effort for voice in production.
Cost is the sum of the layers
At Retell, the public grid runs from 0.07 to 0.31 dollar per minute: voice infrastructure, synthesis, telephony, then per-minute options such as the knowledge base or the removal of personal data.
Disclosing the machine is mandatory
Article 50 of the AI Act requires telling the caller they are speaking to an AI. The French legal sources we consulted place its application on 2 August 2026 for callbots. We build it into the greeting itself.
Outbound cold calling changed regime
Since 11 August 2026, article L. 223-1 of the French consumer code requires prior consent for any telephone canvassing. An outbound consumer prospecting agent gets legally scoped before the first line of code.
Turnkey platform or custom assembly?
Two routes to a voice agent in production. The right one depends on your deadline and on what becomes structural next: latency, cost at scale or data sovereignty.
| Criterion | PlatformVapi, Retell | AssemblyLiveKit, realtime |
|---|---|---|
| Time to go live | A few days | Three to four weeks |
| Cost structure | Platform margin per minute | Cost of the blocks alone |
| Control over latency | The settings exposed | Every link tunable |
| Business logic | Provided tools and workflows | Your code, no ceiling |
| Hosting | At the provider | Cloud or self-hosted |
| Swapping a block | Whatever the catalogue offers | A configuration change |
| The right case | Validate the use case fast | Hold the volume and the cost |
The wrong choice is not one or the other, it is being unable to move from the first to the second. We put every block behind an internal interface so the switch stays a configuration change and a regression test.
What we measure on a voice agent
The other AI building blocks
A voice agent combines with these blocks more often than it replaces them.
GeminiThe model that decides, when text is enough and voice is not the point.
MistralThe sovereign option, hosted in France, when the data requires it.
HeyGenThe video avatar, when voice alone does not carry the message.
ElevenLabsWe build your ElevenLabs connectorTranscriptionWe build your transcription connector
Anthropic (Claude)A model that calls your tools, not one more chat
OpenAIThe model writes into your tools, it does not chat
CursorThe agent reaches your internal tools, not just your codeWe combine AI voice agents with
The stack that surrounds a voice agent on our projects.
AI voice agent: your questions
Four steps. First scope it: inbound or outbound, which call reasons are covered, which are excluded, and on what signal the agent hands over. Then choose the assembly: a platform such as Vapi or Retell to validate the use case in a few days, or a direct assembly on a realtime API and a SIP trunk when latency and cost become structural. Then wire in the business tools, calendar, CRM or ERP, so the agent writes during the call instead of leaving a message. Finally measure on your own calls: latency per turn, human handover rate, average cost. The hard part is not making a model talk, it is holding a conversation in French over the phone when there is background noise and the caller spells out their name.
The real cost is the sum of the layers, not the model. Retell's public grid runs from 0.07 to 0.31 dollar per minute, broken down across voice infrastructure, speech synthesis, telephony, and per-minute options such as the knowledge base or the removal of personal data. At OpenAI the realtime model is billed per audio token, which moves the question to design: every turn sends the history back, so a badly designed call costs several times a well-designed one. We set a spending cap per call and per day, with controlled degradation as the cap approaches, and we hand you the breakdown before we build.
The first tier, yes: identify the caller, understand the reason, book an appointment, open a ticket, route. Beyond that, no, and that is the only honest deployment. Conversations with several turns of ambiguity, emotional or contentious requests, long spelling out, very noisy environments, callers talking over the agent and language switching mid-call all stay out of reach. An agent that claims to handle everything damages the brand faster than an IVR that owns what it is. So we design the human exit first: intent unrecognised twice, a trigger word, detected irritation, a tool failure, an excluded topic.
Yes. Article 50 of the AI Act requires informing a person that they are interacting with an AI system, unless it is contextually obvious, and the French legal sources we consulted confirm it applies to callbots and voicebots from 2 August 2026, with penalties announced up to 15 million euros or 3 % of worldwide turnover. On top of that comes the CNIL guidance on listening to and recording calls: information given at the time of the call, purpose, legal basis and recipients. In practice that means a greeting disclosure, a defined retention period, a separation between operational and debugging data, and an up-to-date processing register. We set that frame during scoping, not after go-live.
That is the most constrained case. The law of 30 June 2025 rewrote article L. 223-1 of the French consumer code, in force in that wording since 11 August 2026: prior consent from the consumer is required for any telephone canvassing, Bloctel disappears, and a contract concluded without consent is void. On top of that sits a numbering constraint, since Twilio explicitly forbids automated outbound calling and sales from a French geographic or mobile number. An outbound agent stays feasible on a consented base, calling back an inbound request for instance, or running a satisfaction survey after a job. Legal scoping comes before development, not after.
An AI voice agent project?
Let's talk. 30 minutes to scope the call reasons you want covered, the applicable regulatory frame and tell you honestly what is feasible today.
Discuss my voice project