AI voice agents: what you can really automate in 2025

You may have had the chance to try OpenAI's voice mode? Voice agents are the same thing, except the agent follows precise instructions, has tools at its disposal and has been optimized to help users in a few specific cases.

I've already shared concrete examples on this topic in previous newsletters. But back then, the technology in French wasn't very mature yet: the voice (especially the French one) was still too robotic and the latency was a bit too high.

This is only the beginning of voice agents in France

For a few months now, a shift has been happening: the quality is finally here.

Latency is reduced, they understand context, handle silences, follow up, shorten their answers, know what to do, and in French.

Why this change now?

I see 3 major reasons:

• The models have made a leap : Understanding is getting sharper, responses faster and higher quality.
• Voice generation has improved dramatically (ElevenLabs, Cartesia, …) with turnkey tools to build them like VAPI. You can deploy a voice agent in a few clicks without complex infrastructure.
• The cost is becoming affordable. For example, OpenAI has cut its prices by 60 to 80%. We're talking a few cents per call.

And for many companies, there is an immediate ROI.

A plumber who doesn't have time to pick up the phone

A customer support team drowning in calls?

The AI agent filters, answers, books an appointment, 24/7. And it handles call scalability (especially load spikes for companies with volatile volumes)

How it works

An AI voice agent is a pipeline that turns the human voice into instructions processed by an AI, then back into synthetic speech.

Here are the 3 main steps shown in the diagram below:

1. STT (Speech-to-Text) : the agent converts the user's voice into text.
   Current solutions: Whisper (OpenAI), Deepgram, Azure, AssemblyAI
2. LLM (Large Language Model) : the text is analyzed by an AI model that generates a suitable response.
   Models used: GPT-4o, Gemini 2.0, …
3. TTS (Text-to-Speech) : the text response is turned into audio with a realistic voice.
   Current solutions: ElevenLabs, PlayHT, Cartesia, …

All of this goes through an orchestration layer, here in the example it's via a voice agent building tool (VAPI server), which links the different modules together (transcription, model, voice…) and manages the calls via Twilio or WebRTC.

Twilio is the solution I use for telephony, it's fairly simple to get French phone numbers by filling out the form (24 to 48h lead time), a KBIS is required.

All the solutions on the market use the same technologies mentioned above, the differentiation comes from the call orchestration behind the scenes and above all from the features they can offer.

Here, I'll zoom in on VAPI, which remains one of the most advanced technologies on the market.

Building your voice agent step by step

You don't need to be a developer to launch an AI voice agent. To do this, we go through VAPI (or Rounded - a French solution similar to VAPI). The interface is clear, modular, and lets you chain all the steps without writing a single line of code.

Here's how to proceed:

1. Choose the AI technologies: STT, LLM, TTS

As soon as you create the assistant, you define:

• the voice transcription (Deepgram, Whisper, etc.)
• the conversation model (GPT-4o, Gemini, etc.)
• the voice synthesis (Cartesia, ElevenLabs, PlayHT, etc.)

Depending on your choice, the technology used will impact cost and latency. The more powerful the technologies you use, the higher the latency and cost will be.

For a good quality/price ratio, you can go with the following setup: Deepgram Nova 2 General in French, Cartesia Sonic v2 and GPT-4o-mini (if the conversation stays basic).

2. Structure the instructions

Next, we can move on to the instructions: the system prompt. This is where you explain its role, its tone, its mission, the information to collect, and so on.

In the prompt, I recommend:

1. Clearly define the agent's role. Start with a simple sentence:"You are Julie, a sales advisor for a home repair company."
   The agent has a clear identity and a tone to adopt (warm, professional, informative…).
2. Use simple, natural language"Use a simple level of language. Use professional, natural language, with words that make the conversation more lively and not robotic."
   The agent is smoother, more human, and avoids convoluted phrasing.
3. Give the agent a precise goal"Your role is to understand the client's need and offer an appointment if relevant."
4. Limit the length of responses"Your responses must be short: 1 to 2 sentences maximum."
   Avoids long AI monologues and smooths out the voice exchanges.
5. Structure the prompt into logical blocks. Clearly separate:The role : "You are Julie, an advisor…"The mission : "You must qualify the request…"The steps : "Start by greeting, then ask for the first name, then the address…"The business context : "The company offers fast interventions, 24/7."
6. Give realistic dialogue examples. This guides the style and the flow.Caller: "Hello, I'd like a quote"
   Agent: "Of course, can you give me your first name?"
   Caller: "Marc"
   Agent: "Thank you Marc. Is it for a home or a business premises?"
7. Add a "Notes" section to fine-tune the behaviorTo place at the end of the prompt for specific rules:"If the user is vague, rephrase the question."
   "Don't reformat the date. Only trigger the booking function once."

A good prompt is like a brief given to a new employee. It must be clear, complete, considerate, but firm on what it must do (and must not do).

3. Choose the operating mode (single prompt or workflows)

When you create an agent, there are two ways to structure the conversation logic.

The most common one, the single prompt

It's the simplest way to get started. You write one big block of instructions (the system prompt) that handles all the possible resolution paths.

And in cases where it can't answer, it will refocus the conversation or transfer to a human.

Basically, it's a kind of intelligent IVR that categorizes the incoming call and answers simple tickets from the context it has.

It's enough for simple agents such as:

• Appointment booking
• Answering FAQs
• Lead qualification in 3-4 questions

✅ Advantages: quick to set up, smoother to test
❌ Limits: hard to maintain as soon as the conversation flow gets more complex

The other option is to define a workflow

As soon as you have several types of calls, or conditional branching, the workflow mode becomes necessary.

VAPI offers a visual editor (like n8n), which lets you:

• have a general prompt (information about the agent)
• create multiple steps (sub-prompts)
• add conditions: "if the client is an existing one, say this…"
• trigger specific actions (API, transfers, tools…)
• organize the conversation like a logic tree

It's ideal when the conversation scope is broad and you want the agent to handle the call autonomously, often used for after-sales service and multi-service support.

What's interesting is being able to have a main voice agent that can redirect to other assistants. Each assistant will have tools at its disposal (often max 3)

✅ Advantages: ultra-modular, more reliable (deterministic workflow), clearer to debug
❌ Longer to design (but much more robust)

4. Add tools to your voice agent

Having a conversation is nice, but it doesn't let you go very far in call management. What's interesting is letting your agent perform actions and thus truly solve a need over the phone.

For this, solutions tend to offer turnkey tools. For example, on VAPI, you'll find the ability to:

• Send an SMS
• Make an API call to scenarios (book or check a slot on Google Calendar)
• Leave a voicemail
• Transfer to a human agent

Here, we can distinguish native features like "leave a voicemail" or "transfer to a number" that can be enabled in 1 click.

Or more advanced features such as "voicemail detection". Very useful when running outbound call campaigns with a voice agent. It lets you leave a voicemail if you reach an answering machine.

And the addition of more custom tools. You can then make an API call to any scenario on an automation tool. I'll give you examples further down of what I've been able to do.

(VAPI is also starting to add third-party tool integrations like Slack)

As you can see, building a voice agent on VAPI is about assembling all the building blocks. And above all, testing, iterating, adjusting. By the way, I haven't mentioned monitoring, but all the logs are available with plenty of data including the transcripts (and the option to record the calls).

Demo of a French voice agent

Now, here's the demo of a voice agent I configured on VAPI. The goal is mainly to show you how the agent improvises from the instructions I gave it.

Another demo on an intelligent IVR equipped with SMS sending:

In the demo, the agent fails to send me the SMS the first time because I cut it off just before it finishes. A chance to show that it did have the information that the SMS wasn't sent (and that it then sent it).

The possible use cases

Once your voice agent is configured, the real question becomes "what can it really be used for in a business"? Today, what's actually running in production?

Here are the mature use cases:

1. Intelligent phone switchboard (AI IVR)
   No more "press 1, press 2…". The agent understands the client's spoken request and redirects them to the right person or department.
   Ideal for SMEs, medical practices, agencies…
2. Automatic appointment booking
   The agent picks up, qualifies the need, offers a slot and books it directly (via Google Calendar or another tool).
   Ideal for dentists, plumbers, hairdressers, local agencies…
   Result: 0 missed calls plus better-filled time slots. It can answer on weekends and outside opening hours
3. Outbound calls for customer (or prospect) follow-ups
   A typical example: debt collection. The agent automatically calls a list of customers.
4. Qualification of inbound leads from Meta
   A Facebook lead comes in, the agent calls within the minute, asks 2-3 key questions and offers an appointment. Everything is synced to your CRM.
   Goal: reduce the "speed-to-lead" before a competitor calls or the prospect forgets they clicked on the ad.
   Results: a higher conversion rate on your ads
5. Reactivation of inactive customers (or prospects)
   The agent calls former customers (or dormant prospects) to create engagement. We personalize the call with as many variables as possible so it isn't seen as spam.

In practice: as soon as you have a flow of customers to call, filter or route… a voice agent can save you time, cut costs and boost your conversions.

Automation on your agent

Above, I mention the tools the agent can have at its disposal. Here are examples of automation I use on Make to make agents more powerful. These scenarios link the voice call, the tools (Google Calendar, Brevo, Sheets…) and the business workflows.

Smart appointment booking with Google Calendar

During a call, the voice agent checks a slot's availability, creates the event in your calendar, then receives a confirmation it can relay back to the caller.

The flow in detail:

• Webhook triggered from the voice agent
• Availability check or booking in Google Calendar
• JSON response sent to the agent via Webhook
• Confirmation message or alternative proposal in real time

Automated SMS sending based on the call context

During the exchange, the agent can trigger different SMS sends:

Tech used: Brevo, but OnOff is a more cost-effective alternative for mass sending.

In the SMS, I send URLs that let callers be autonomous.

On-the-fly outbound calls from a Google Sheet

For follow-up cases, debt collection for example, the automation relies on a Google Sheets database. In 1 click, it triggers calls for the rows in a certain status, and the status of the exchange is updated at the end of the call.

The steps:

• Search in Google Sheets
• Call triggered by API
• Row update with status (call successful, no answer, etc.)

Goal: launch 500 or 5000 calls without any manual intervention, with tracking that's readable for the teams.

These automations transform the voice agent: it doesn't just answer calls, it performs actions, relays information, updates your tools, and moves your processes forward.

What's next: what voice agents will transform further

What we're seeing today with AI voice agents is an inflection point. The technology is ready, concrete use cases exist, but in my view, we're still only scratching the surface of the potential.

You've surely noticed, there are still limits:

• Handling noise or accents depending on the Speech To Text (STT) technologies
• Emotions in the voice, intonation, rhythm
• Latency on some models (we're still on classic models like GPT-4o)
• The limits on long conversation memory

But progress is fast. Time is melting these barriers away every month.

On the other hand, the market is completely fragmented on the B2B side:

We have both horizontal platforms (like VAPI, Rounded, Retell, Synthflow, Bland, … bottom right on the mapping): they aim to serve as many cases as possible in no-code or API-first.

And at the same time, vertically specialized solutions, either on the industry side like healthcare, or on the function side with 11x on the Sales side or Decagon on the support side.

💡By the way, for those who are interested, we deploy these last 2 solutions on the French market.

And on the B2C side? Voice will surely become widespread.

On the end-user side, voice will also become a way to interact with your apps, your tools, your services. A voice coach in your fitness app. A voice assistant to manage your personal accounting. A voice agent in a store or in a mobile app.

Voice will no longer be just a channel. It might be an interface ?

In short, AI voice agents won't just answer calls, they'll also settle into every professional and personal tool. Always available. Always connected.