An AI voice agent is a system that picks up the phone, holds a real time conversation, understands a spoken request, triggers actions inside your tools and hands over to a human when the situation goes beyond it. In 2026, the technology is no longer the question. It works in French, with a credible voice, for a few cents a minute. The question is the scope you give it.
Tandem runs these deployments for small and mid-sized companies, from Santiane to KparK and Nomination, and sees very clearly where the line sits. Smart switchboards, appointment booking and form callbacks work. Cold outbound campaigns and long conversations still break. This article covers both sides of that line, with real costs checked on 9 August 2026 and the legal obligations that came into force this month. For context, 18% of French companies with 10 or more staff used at least one artificial intelligence technology in 2025, against 6% two years earlier, according to Insee Première no. 2120 published on 21 July 2026.
How an AI voice agent works
An AI voice agent chains four technical blocks on every conversational turn. Telephony opens the audio stream, through a carrier such as Twilio or a browser connection. Speech to text turns speech into text, with engines like Deepgram or Whisper. A language model reads that text and produces an answer. Text to speech makes the answer audible, with ElevenLabs or Cartesia voices. It all goes back into the call, and the loop starts again.
The block that makes the difference is none of those four. It is the orchestration layer sitting on top, the one that carries the instructions, keeps the memory of the call and hands tools to the agent. Without it, you have a talkative answering machine. With it, the agent checks a slot in a calendar, creates a record in the CRM, sends a confirmation text message and switches over to a human adviser. That layer often runs on a dedicated platform, sometimes straight on an orchestrator such as n8n.
The setup Louis Graffeuil settled on in May 2025 for French paired Deepgram Nova 2 for transcription, GPT-4o mini for reasoning and Cartesia for the voice, with a low temperature and answers capped at two hundred tokens. Models have changed since, the logic has not. Each block can be swapped independently, and every time you arbitrate between quality, latency and cost.
That breakdown shows where to look when an agent drags. Transcription and endpointing cost a few milliseconds. The model and text to speech share the rest, in almost equal parts. Switching transcription engine to gain latency therefore means optimising the cheapest line of the budget.
Why voice agents finally work in French
Three things moved at once between 2024 and 2026. Models understand a badly phrased sentence better and answer faster. French synthetic voices left the robotic register that gave the machine away in three seconds. And cost collapsed, to the point of making a call that earns nothing directly worth placing.
That shift can be dated. In May 2025, Louis Graffeuil was already writing that quality was finally there in French, after two years of inconclusive tests. The detail of that turning point sits in his field report on what a voice agent can really automate, and the voice to voice mechanics are unpacked in our article on the voice agent pipeline.
Which AI voice agent use cases actually run in production?
Five use cases hold up today, and they share one trait. The conversation is short, the goal is single, and the agent has an exit door towards a human.
- The smart switchboard. The agent understands a request phrased in natural language and routes it to the right team, with no press 1 then 2 then 3. It replaces an interactive voice response menu that annoyed everyone.
- Appointment booking. The agent qualifies the need, checks real availability in the calendar and books the slot. Medical practices, trades and local agencies benefit first.
- The one minute callback. A prospect fills in a form, the agent calls back immediately, asks two or three questions and offers a meeting. The gain comes from speed, not from the pitch.
- Frequent questions outside office hours. Opening times, address, order tracking, status of a file. The agent answers at night and at weekends, when nobody picks up.
- Cleaning a call list. The agent dials a list of numbers, identifies valid lines, switchboards and dead numbers, then updates the CRM.
| Use case | Maturity | Condition for success |
|---|---|---|
| Smart switchboard | Mature | A clear service tree and a human transfer always available |
| Appointment booking | Mature | Write access to the calendar and explicit slot rules |
| Post-form callback | Mature | Consent collected on the form and a trigger under one minute |
| Frequent questions | Mature | An up to date knowledge base and a bounded answer scope |
| Collections and customer follow-up | Intermediate | An active contract justifying the call and a controlled tone |
| Tier 2 technical support | Fragile | Long memory and multi-step diagnosis still unstable |
| Cold prospecting | Avoid | Prior consent mandatory in France since 11 August 2026 |
What does not work yet, and the pivot that saves the project
Cold outbound calling remains the bad bet. In August 2024, Tandem tested a voice agent for a founder with a small team whose best source of customers was cold calling. The verdict after testing was that the voice was not good enough yet and the project would not go into production.
The story does not end there, and that is the interesting part. The same agent was turned into a lead prioritisation assistant. It dials a list of 1,000 landline numbers, categorises each line according to the outcome of the call, and returns a clean database in under two hours for around 30 euros in total. The founder then knows which numbers actually reach the decision maker. The technology did not deliver on the original promise, it delivered on a neighbouring one that paid off immediately.
The other limits are more technical. A noisy line or a strong accent degrade speech recognition. A conversation running beyond a few minutes makes the agent lose the thread. Intonation stays flat facing an unhappy customer, where a human would soften. None of these limits is fatal when the scope is well chosen, and all of them become fatal when the agent is supposed to handle everything.
How much does an AI voice agent cost in 2026?
The direct cost of a minute of calling reads on three lines. The voice platform, models included, comes to around 0.07 euro a minute according to the official ElevenLabs Agents pricing checked on 9 August 2026, that is 0.080 dollar. Telephony adds on top, with an inbound call on a French local number billed around 0.01 euro a minute and an outbound call to a mobile around 0.04 euro, Twilio France rates on the same day. The number itself costs in the region of 1.20 euro a month.
Those figures explain why use cases that looked absurd three years ago now make sense. A qualification call lasting twenty seconds costs about three cents. Cleaning a list of a thousand numbers therefore comes to a few tens of euros, which completely changes the maths against a day of human work.
The breakdown of a real call holds a surprise. The language model, the one everybody talks about, is the smallest line on the bill. Transcription is close to free. Text to speech and above all the orchestration platform absorb most of the cost. Haggling over your model to save three cents therefore makes no sense until the orchestration layer has been looked at.
What the law has required of voice agents since August 2026
Two texts changed the picture this month, and both apply. Since 2 August 2026, Article 50 of the European artificial intelligence regulation requires that anyone interacting with an AI system be informed of it. For a voice agent, that means an explicit announcement in the first seconds of the call, without waiting for the caller to ask. Breaches of the transparency obligations carry fines of up to 15 million euros or 3% of worldwide turnover.
The second text targets outbound calls. Since 11 August 2026, Article L223-1 of the French consumer code forbids calling a consumer for marketing purposes without prior consent. That consent must be free, specific, informed, unambiguous and revocable, and the burden of proof sits with the company. The Bloctel opt-out register disappears, and contracts signed in breach of the rule are void. A voice agent changes nothing about this obligation, it amplifies it, because it makes calling a thousand people in one evening trivial.
Then comes the data side. A call always processes personal data, starting with the voice itself and the calling number. If the call is recorded or transcribed, the caller must be told at the start of the conversation, the legal basis for the processing must be chosen explicitly and the retention period must be bounded. Those three points take half a day to settle during scoping, and cost far more after an audit.
How to deploy an AI voice agent without failing
The method that works fits in five steps, and the first one is not technical.
- Pick a single call flow. Inbound calls outside office hours or post-form callbacks give a clean scope and a measurable result from the second week.
- Write the instructions like an onboarding brief. A role, a mission, steps, business context, sample dialogues and answers capped at two sentences.
- Wire up one tool to begin with. The calendar or the CRM, not both. Every extra tool multiplies the error cases to test.
- Plan the human handover before going live. An agent that knows how to pass the call on is worth more than an agent improvising a wrong answer.
- Listen to real calls every week. Recordings and transcripts are the only material that improves the instructions.
Two settings on that page often go unnoticed and yet decide perceived quality. The output token cap forces the agent to answer in one or two sentences, which avoids the monologues that are unbearable on the phone. And the spoken language instruction, the one asking for contractions and everyday words, does more for the naturalness of the conversation than the choice of voice.
That sequence is exactly the one of a classic automation rollout, and deliberately so. Tandem usually runs it after an AI audit that identifies the call flows that genuinely cost money, then puts it into production as part of an autopilot engagement. When orchestration has to stay in house, it runs on n8n, and our team steps in as an n8n partner agency.
Voice agent or plain workflow, how to decide?
Many projects labelled « voice agent » do not need an agent. An agent is a probabilistic system that decides its own actions, and that autonomy is paid in unpredictability. If your need boils down to qualifying a lead in three questions then writing to a CRM, a framed workflow with one model call in the middle is enough, costs less and can be debugged in ten minutes. That hierarchy of optimisation levels is detailed in our article on workflows, assistants and AI agents.
Voice justifies autonomy in one situation only, the one where the other party speaks freely and no script can anticipate what they will say. That is precisely an inbound call. It is much less true of a scripted outbound call, where rigidity is an advantage.
In practice, the most robust deployments mix both approaches. A single instruction block is enough as long as the agent handles one intent. As soon as several call types land on the same number, a conversation tree with conditional branches takes the lead again, because it can be debugged and changed without breaking everything.
Should your company deploy an AI voice agent?
An AI voice agent is mature in 2026 for anything touching inbound calls with a bounded scope, and Tandem deploys it without reservation on those flows. It stays immature for cold prospecting, now blocked twice over by conversational quality and by prior consent. In between, collections and customer follow-up work when an active contract justifies the call.
The right reflex is not to order a voice agent. It is to count how many calls your company misses each week, what they are worth, and at what time they come in. If the answer fits inside a precise time window, the voice agent is justified. Otherwise, an automated callback and a well placed form will do the same job for less.
