In late 2024, one trend became impossible to ignore: AI voice agents.
You may have noticed it too: more and more voice solutions are emerging, often verticalized by industry (support, sales, healthcare, real estate...). And that is no accident.
Y Combinator, one of the largest incubators in the world, saw its fall 2024 cohort with 22% of incubated startups working on voice agents. Nearly 1/4.
And through 2025, we saw these companies grow.
Even though we are still early, voice agents are starting to make their way into companies. A bit like chatbots in 2020, but with far greater technological maturity.
What is interesting is that two technological approaches coexist today, for the same use cases, but with very different logics.
- On one side, the “classic” voice agents, which rely on a heavier and more rigid infrastructure with a well defined chain: speech transcription (STT), a language model, then voice synthesis (TTS), all orchestrated by a platform. I come back to this below
- On the other side, a more recent advance: voice-to-voice. Here, you no longer think in terms of a “pipeline” but of direct voice conversation. Voice goes in, voice comes out, with almost no perceptible break. It is the advanced voice mode you may have tried on ChatGPT or other apps.
In this edition, the goal is simple: to give you a clear and actionable read of what is really happening today.
On the agenda, I compare the 2 technical approaches presented above, sharing concrete examples of what a voice agent can do. We will close on what awaits us in 2026 on this topic.
No fantasy. No demo disconnected from reality. Just what works now, and what is coming very soon.
(I have done several editions on voice agents, I invite you to take a look at this one where I explain in detail how to build a voice agent.)
2 distinct approaches to voice agents
Today, when we talk about AI voice agents, it is not a single technological choice but two distinct architectures.
Same business objective. Same use case. But very different foundations.
1/ Approach 1: the "classic" voice pipeline (STT → LLM → TTS)
This is the overwhelmingly dominant architecture in production today. It relies on a heavier infra, but also one that is far more controllable.
Concretely, the chain is as follows:
- Speech-to-Text (STT): the voice is transcribed into text
- LLM: the text is analyzed, reasoned through, and structured to generate a response following a prompt
- Text-to-Speech (TTS): the response is converted back into audio
The big plus here is fine grained control over behavior (prompts, rules, workflows) across the different steps, with the option to add guardrails. It keeps the AI from improvising too much.
On the monitoring side, it is also far more complete: logs, transcripts, replay, debug
This is the approach found in almost all voice agents in production today.
Why? Because it is robust, predictable and suited to business constraints.
On the other hand, since there are 3 steps, that means 3 API calls to different technologies, and therefore higher latency, a sometimes more “segmented” experience and an infra that is harder to maintain.
And as you can imagine, latency is key for a voice agent.
2/ Approach 2: voice-to-voice (real-time, end-to-end)
Recently, another approach has been gaining ground: native voice-to-voice.
You no longer think in terms of a pipeline. Voice goes in, voice comes out, with almost no break.
This approach is driven by players like ChatGPT, Gemini or Grok, through real-time APIs capable of processing audio continuously.
What changes radically:
- latency is very low (near instant conversation)
- better handling of silences, interruptions, hesitations
- more natural intonation, a more human rhythm
- the first building blocks of vocal emotion appearing
And on cost, it is getting serious. Whereas until recently the cost was far too high and complex (input / output cost → very hard to predict), Grok announces 5 cents per minute with its latest real-time API.
In production, the STT→LLM→TTS pipeline remains dominant because it is more robust, scalable and controllable than direct audio→audio transformation approaches
But I am betting on 2026 for the shift in technical dynamics that the stronger arrival of voice-to-voice could bring.
👉 This technology will take a much bigger place in voice architectures, especially for conversational experiences with high perceived value.
What a voice agent can actually do
When we talk about a voice agent, we often think of “answering the phone”.
In reality, it is much broader than that. A voice agent can act before, during and after a call. And that is where the value is created.
Before the call: inject context (and gain relevance)
Even before saying “hello”, the agent can already know a lot.
Thanks to the caller's number, it can query a CRM, find an existing customer, identify a history (last call, open ticket, pending quote...) and sometimes anticipate the reason for the call.
This is key because you have surely noticed it with AI models → the more context there is, the more relevant the response. It is exactly the same for voice agents.
It is simple, but extremely powerful.
Here for example, this block retrieves the caller's context:
During the call: from “single prompt” to real business logic
This is where the change is most striking.
Before, the dominant approach was the single prompt. A big meta-prompt meant to handle every possible case: intents, responses, exceptions. Like here on VAPI.
It worked... but with major limits: hard to maintain, sometimes unpredictable behavior, complexity that explodes as soon as the scenario widens.
Today, we have clearly moved to a multi-prompt logic to go further in the processing.
Concretely:
- we break the conversation down into nodes
- each user choice moves forward through a business tree
- each node has its precise role, its dedicated prompt, its rules
And in each node, the agent can:
- draw on a knowledge base kept up to date via RAG
- use tools (API, CRM, calendar...)
- send a text message
- check availability (teams, slots, operators)
- transfer the call to a human if needed
All of this live, during the conversation. If you are interested, I made a video to show all of this.
A small detail that changes everything: before using a tool, the agent can use transition phrases:
“Please hold, I am checking the information.”
“I am looking to see if an operator is available, I will get back to you.”
It masks the latency, makes the exchange more human, and avoids awkward silences...
After the call: automate what used to be done manually
Once the call is over, the work is not done.
And that is often where teams lose time.
After the call, a voice agent automatically triggers transcript analysis, task creation in the CRM, a sentiment analysis and anything you can imagine: a notification to a team if no one was available, for example.
We can reproduce complete business logics that, until now, were manual: customer follow-up, internal escalation or scheduled reminders
Example here on Hubspot during opening hours (peak activity) with task creation in the CRM and assignment to the account manager:
The new use cases
Once the voice agent is set up, the real question quickly becomes obvious: “What can it really be used for in a business?”
Today, some use cases are mature enough to be deployed. They often serve as an entry point... before moving toward more innovative uses.
The starting point for a voice agent internally is the smart IVR: no more “press 1, press 2...”. The agent understands the request spoken aloud, handles simple questions via a knowledge base and routes to the right department.
You can even go as far as automatic appointment booking: qualification, slot suggestion, direct reservation in the calendar.
→ Zero missed calls, even in the evening or on weekends.
What is being sold is not replacing humans (we are far from that) but making it possible to avoid missed calls, handle peaks, respond 24/7...
Now, here are the new cases we are seeing emerge:
- Outbound calls for sales teams: it is probably one of the most profitable uses for a company today. The voice agent is used to call inbound leads within the minute, ask 2-3 key questions, qualify the need, propose an appointment and sync the exchange with the CRM
The goal is simple: to reduce the speed-to-lead. Before a competitor calls. Before the prospect forgets.
👉 Result: short conversations (1-2 min), a pickup rate of +70% and a clear increase in campaign conversion rates (Meta, Google, etc.)
- Quote follow-up (and collections). This is another use case rising fast: automated follow-up.
Rather than ignored emails and time-consuming manual calls, the voice agent calls automatically, personalizing the approach to find out whether the quote was received and to transfer to the responsible sales rep.
👉 These are short, structured, efficient exchanges. And above all, scalable, without tying up the teams.
- The recruitment and pre-qualification calls in recruitment. This is more recent, the next big shift.
The voice agent replaces the first screening calls and the repetitive, low-value exchanges. It asks the right questions, evaluates the answers, structures the information, and feeds the HR tools.
In staffing agencies, the feedback is very strong:
👉 more qualified candidates
👉 less time wasted
👉 better overall quality of the profiles passed on
In practice, as soon as you have a flow of calls to handle, customers or prospects to qualify, decisions to speed up, then:
👉 a voice agent can save time, reduce costs and increase conversions.
What next? Asking the question of emotion, of intonation, and of the relationship of trust in the voice.
Toward more emotional voice agents?
For a long time, AI voice was mainly judged on one criterion: “Does it still sound robotic?”
In 2025, the question changed. It is more about: “Does it convey an intention? An emotion?”
Today, we can produce controlled emotion with Text-to-Speech (TTS)
Tools like ElevenLabs make it possible to add emotions directly into the text, via variables, and to guide intonation, rhythm, intensity
Example:
“[smiling] Perfect, I will take care of it right away.”
“[reassuring] Do not worry, I am checking the information.”
We do not change the substance. We change the way it is said.
This is particularly useful at key moments in support or in customer relations.
So we are starting to design emotion into the voice conversation. And it completely changes the perception on the user's side.
Even though there is still a big difference between English and French on the same text:
And the French version:
On the voice-to-voice side, emotion is much more natural. You no longer “act out” an emotion from a text. The emotion is built directly into the model.
Result: more natural intonation, better managed silences, a more human rhythm, smoother reactions to what the other person says.
Example here with the latest version of Grok (latency is super low):
What if AI became better than humans... at emotion?
It is counterintuitive, but several signals point in that direction.
As a16z sums it up very well:
AI does not get tired, it does not get annoyed, it really listens all the way through and can adapt its tone consistently, all the time.
In certain contexts, support, guidance, recruitment, sensitive situations, this emotional consistency becomes a huge advantage.
For now, it is still largely underused.
But it is clearly one of the big topics of the coming years.
