Contents7
The phone remains the channel for urgent requests, and the one that is hard to staff at all times: the line rings during a meeting, at lunchtime, or after closing. Voice agents, programs that pick up, listen and answer with a synthetic voice, have become fast enough in 2026 to hold a conversation without awkward silences. This guide from our section on process automation and business bots reviews what they can do, what they do badly, and the rules to know before connecting one to your number.
In short
A voice agent connects to your phone system and handles the call in real time: listen, understand, act, answer.
It does short, predictable requests well: hours, status, appointments, messages, routing.
It does ambiguous requests, complaints and anything that commits the company badly.
Callers must know they are talking to an AI, and cold calling has required prior consent in France since 11 August 2026.
The switchboard and agents remain essential: the voice agent takes repetitive calls off their hands and passes them the others with context.
What a voice agent is, in practice
A voice agent is not the old touch-tone menu (“press 1 for…”): the caller speaks normally, and a model understands the request. Several vendors offer these real-time voice models, OpenAI as well as Google, and connecting them to telephony has become standard. OpenAI’s documentation explains that a call can reach its voice model through a SIP trunk, the protocol of internet telephony; Google’s presents its Live API as a low-latency, real-time voice interface, designed among other things for customer support agents.
The path of a call handled by a voice agent
- The call
It arrives through your phone system, on a SIP line.
- Listening
The model transcribes and understands the request.
- Action
Your rules decide, with the calendar, the CRM or orders.
- The answer
The voice replies, or transfers to a person.
Step 3 is yours, and OpenAI’s documentation says so plainly:
Your backend still owns the incoming-call decision, session configuration, authorization, and business logic.
In other words, the model talks; your system decides what it is allowed to do. The model is chosen on your own calls, not on a demo: our method to evaluate an AI model before production applies to voice too.
What it does well
- Answering frequent questions: hours, address, lead times, documents to provide.
- Giving a status when it is in your tools: an order, a file, a service visit.
- Booking, moving or cancelling an appointment in a shared calendar.
- Taking a structured message (who is calling, why, when to call back) and filing it in the CRM.
- Routing the call to the right department, instead of a touch-tone menu the caller listens to until the end.
- Covering the line outside opening hours, when nobody can pick up.
These are the same conditions as for a written bot: a frequent, simple request whose answer is already in your systems. Our article on WhatsApp and website bots details this sorting.
What it does badly
- Ambiguous requests: the caller who describes a complicated situation before getting to the question.
- Difficult listening conditions: a call from a car or a building site is understood less well than a call in a quiet room, and a misheard proper name throws off what follows.
- Emotionally charged calls: a complaint, an angry customer, a delicate personal situation.
- Decisions that commit the company: a commercial gesture, an exception, a delivery commitment.
On that last point, a voice that promises commits the company as much as a text does. Air Canada was held responsible for what its chatbot had wrongly stated, a case we tell in our article on AI agents in production. On the phone, there is not even a written record the customer can reread: the transcript of every call must be kept.
Watch out
What the agent is not allowed to say
Write down the list of what the agent never promises (a refund, a deadline, a price outside the price list) and enforce it through your rules, not only through the instructions given to the model. Instructions can be worked around; a tool that does not exist cannot.
The rules to follow
Three texts come up as soon as a synthetic voice calls or answers:
- February 2024
In the United States, the FCC classifies AI-generated voices as “artificial” voices under the law on automated calls, which requires the prior express consent of the person called.
- 2 August 2026
In the European Union, the AI Act’s transparency rules apply; a person must know they are interacting with an AI system, unless it is obvious.
- 11 August 2026
In France, telephone cold calling is no longer allowed without the prior consent of the person called.
The sources: law firm Cooley’s analysis of the FCC ruling, Article 50 of the AI Act and Le Club des juristes on the decree of 25 July 2026. In practice, a voice agent introduces itself from the start of the call (“you are speaking with the voice assistant of…”), callers are told if the call is recorded or transcribed, and an agent that calls consumers to sell only calls those who have agreed to it. Our guide to the AI Act and GDPR for a small business covers the rest.
And the switchboard?
It stays, and its work changes. The voice agent takes the repetitive calls and the off-peak hours; the person at the switchboard takes the calls that need judgement, key customers, complaints, and all those the agent transfers. The condition is the same as for a written bot: the transfer arrives with its context. Callers do not repeat their request, and the receptionist sees who is calling and why before picking up.
From the field
We do not run a voice agent in production today, and we would rather say so. Our expense-entry agent, in service on Telegram since July 2026, does follow the rule we would apply to voice: it proposes, and a person approves with one button before anything is written. And a client learned it the hard way on their website: contact requests went out by email, and a failed send went unnoticed. Every request is now saved to the database before the email is sent; a message taken over the phone must follow the same path.
Before you start: listen to your calls
- Sort a month of calls
By reason, with their length and time of day.
- Pick one reason
That is frequent, simple, and answered in your tools.
- Write down what the agent can do
And above all what it never does.
- Plan the way out
To a person, at any point in the call.
- Read transcripts
Every week during the first month.
Not every step needs AI: once the request is understood, booking the appointment or creating the record is a matter of rules, as our guide to automation without AI explains. That is the split we apply in our business bots on Telegram or WhatsApp and in our n8n automation, with or without AI.
For neighbouring topics in the section: customer support where AI sorts and people review, automated supplier invoices, employee onboarding, reports without copy-and-paste, n8n, Make or Zapier, monitoring n8n workflows, automating software without an API, the n8n licence and AI and young people’s first jobs.
Frequently asked questions
Can a voice agent replace a switchboard?
No. It can take repetitive calls and those outside opening hours, but complaints, ambiguous requests and decisions that commit the company stay with people. Set up well, it relieves the switchboard and passes calls on with their context; it does not remove it.
How much does a voice agent cost?
The cost depends on three items: the voice model, billed by conversation volume, the telephony, and the work of connecting it to your tools. The only reliable figure is the one you measure on your own calls, with a per-call meter from day one. Be wary of a per-minute price quoted without the rest.
Do you have to tell callers they are talking to an AI?
Yes. Since 2 August 2026, the AI Act requires that a person be informed they are interacting with an AI system, unless it is obvious. On the phone, a synthetic voice can pass for a human one: the greeting must say so clearly, and callers must be able to ask for a person.
Can you do outbound sales calls with a voice agent?
Only to people who have agreed to it. In France, since 11 August 2026, telephone cold calling requires the prior consent of the person called. In the United States, since February 2024, the FCC has applied the rules on automated calls to AI-generated voices. A voice agent is first a way to answer, more than to call.
Sources
- Can reach its voice model through a SIP trunk (new tab)platform.openai.com
- Live API (new tab)ai.google.dev
- Law firm Cooley’s analysis of the FCC ruling (new tab)cooley.com
- Article 50 of the AI Act (new tab)ai-act-service-desk.ec.europa.eu
- Le Club des juristes on the decree of 25 July 2026 (new tab)leclubdesjuristes.com
Jim · founder of Brain Plus AI
Also in automation and bots
Choose and price7 min
WhatsApp and website bots: saving time or wasting it
A bot saves time when the request is frequent, simple and its answer already exists; it wastes time as soon as it blocks the way to a person. Where it helps, where it costs, and what changed on WhatsApp in 2026.
Practical guide12 min
Automate customer support: AI sorts, humans review
Incoming tickets and emails read, sorted and prepared by AI, then reviewed by your team when the reply commits the company: how to split the work, what can go out on its own, and why sorting costs a few cents when each step has the right model.
Choose and price9 min
Automation without AI: when a rule is enough
Not every automation needs AI. A fixed rule costs less and always gives the same result: here is how to spot the steps where AI really pays off, and those where it only adds cost.