Voice agent
In one sentence
A voice agent is software you talk to out loud that hears you, works out what you are asking, decides what to do about it, and answers in speech.
Not to be confused with voice bot, or Voice widget.
Definition
A voice agent is software you talk to out loud. It hears you, works out what you are asking, decides what to do about it, and answers in speech.
A recorded phone menu cannot do that. A voice agent holds a real conversation, and it can finish the job before you hang up.
The word hides a pipeline. A voice agent is not one model doing everything, it is four jobs chained together, and each one can be swapped for a different supplier. Voice activity detection and endpointing work out when you have stopped speaking. Speech recognition turns your audio into text. A language model reasons over that text against its instructions and whatever knowledge it just retrieved. Text-to-speech turns the reply back into sound.
Newer speech-to-speech architectures collapse the middle of that chain. Audio goes in, audio comes out, and there is no text bottleneck in between. That buys latency and keeps the tone of what you said, but it costs you visibility: the reasoning step gets much harder to inspect and much harder to control.
What separates an agent from a bot
- A voice bot recognizes intents and walks a script tree.
- A voice agent reasons, calls tools, and can change something in another system.
- In practice the dividing line is tool use, which engineers call function calling. If it cannot do anything except talk, it is a bot.
How it differs from the words sitting next to it
- Voice assistant implies a general-purpose companion on a device, the Alexa shape.
- Conversational AI is the umbrella term and covers text as well as voice.
- IVR is the pre-LLM ancestor: menu-driven, rigid, and unable to cope with anything off the tree.
- Agentic voice puts the emphasis on autonomy and finishing multi-step tasks without being walked through them.
Common misconception
Almost everyone assumes the language model is the hard part. In production the hard part is turn-taking. Knowing when a person has finished speaking, and cutting your own speech the instant they interrupt, is what makes a voice agent feel alive or feel broken. A brilliant model with clumsy endpointing loses to an ordinary model with excellent timing.
Why it matters commercially
Voice is the only interface nobody has to learn. It is also the only one where half a second of delay is noticed by everybody, including people who could not tell you what latency means. That combination makes voice agents unusually unforgiving to build and unusually persuasive once they are built well.
Where AsqVox fits
The Orb is a voice agent that lives on a website rather than on a phone line. It runs true retrieval over documents the business uploads, so the reasoning step is grounded in that business's own material instead of general web knowledge.
Visual
Anatomy of a voice agent
The model is one block of five. The other four decide whether it feels human.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
The AI voice agents market is put at roughly USD 2.54bn in 2025, rising to about USD 35.24bn by 2033.
2.54bn to 35.24bnanalyst forecastGrand View Research, 2025 - Read alongside the Market.us forecast below. The distance between two reputable firms is itself the useful fact.
A second forecast puts the same market at roughly USD 2.4bn rising to about USD 47.5bn by 2034.
2.4bn to 47.5bnanalyst forecastMarket.us, 2025 - The two forecasts differ by more than USD 12bn at the top end. The category is early enough that nobody agrees how big it gets.
A call handled by a person costs roughly USD 7 to USD 12.
USD 7 to 12industry rangeWidely cited industry range, 2025 - Not an audited figure. Variance by geography and call complexity is large, so a business case should use local labor cost.
A call handled by a voice agent costs roughly USD 0.40.
USD 0.40industry rangeWidely cited industry range, 2025 - Same caveat. Treat the ratio as directional rather than the absolute number as precise.
Published platform rates start around USD 0.05 per minute for Vapi, USD 0.07 for Retell and USD 0.09 for Bland.
USD 0.05 to 0.09 per minvendor claimVendor published pricing, 2026 - Base platform rates only. Once speech recognition, the model, synthesis and telephony are stacked, all-in production cost lands closer to USD 0.11 to USD 0.33 per minute.
Conversational AI in contact centers was forecast to cut agent labor costs by USD 80 billion by 2026, with one in ten agent interactions automated.
USD 80bnanalyst forecastGartner press release, attributed to VP analyst Daniel O'Connell, 2022 - Dated 31 August 2022. Cite the date. It is a four-year-old prediction now reaching its target year, which makes it useful context rather than current evidence. The one-in-ten figure is the more revealing half.
ElevenLabs raised USD 500m at an USD 11bn valuation in a Series D led by Sequoia, up from USD 3.3bn a year earlier, on reported ARR above USD 330m.
USD 11bnindependentElevenLabs Series D announcement, 4 February 2026, 2026 - Total funding USD 781m. Deepgram raised USD 130m at USD 1.3bn in January 2026. Funding is a proxy for conviction in the category, not for revenue in it.
Examples
In practice
A dental group runs an inbound voice agent on its main number. A caller says "I think I cracked a filling, can someone look at it this week". The agent transcribes the sentence, reads it as an urgent appointment request, calls a scheduling function against the practice management system, finds two slots, offers them aloud, books the one the caller picks, and sends a confirmation text. No menu, no hold music, and nobody in the practice touched it.
The everyday version
You ring a restaurant at seven on a Friday evening. Nobody picks up, because everyone is serving. A voice agent answers on the second ring, takes your table for four at half eight, mentions there is street parking behind the building, and texts you the confirmation. The restaurant kept the booking and nobody stopped working to take it.
Usage
Who says it
- Engineers use voice agent to mean the whole orchestrated stack, and will immediately want to know which speech recognition, which model and which synthesis you are using.
- Sales and marketing use it as the category noun, often swapping it with voice AI, which reliably irritates engineers.
- Procurement meets it in a scope of work and wants to know concurrency limits and what happens when something fails.
Where it turns up
- Usually in the same clause as concurrency, latency targets, escalation path and data residency.
- A well-written RFP separates voice agent, the product, from voice infrastructure, the telephony or WebRTC layer underneath it. A badly written one runs the two together and produces bids that cannot be compared.
Common misuse
- Calling a scripted IVR a voice agent. If it cannot handle a sentence outside its decision tree, it is not an agent.
- Using voice agent to mean voice cloning or synthetic narration. Those are text-to-speech products. No conversation is taking place.
- Treating AI receptionist and voice agent as the same thing. Receptionist is a use case. Agent is the architecture underneath it.
Questions people ask
What is the difference between a voice agent and a voice bot?
A voice bot recognizes a fixed set of intents and follows a script tree. A voice agent reasons over what you said and can call tools to act in other systems, which is the practical dividing line. If it cannot do anything beyond talking, it is a bot.
Is a voice agent the same as an IVR?
No. An IVR is the pre-LLM ancestor: menu-driven, rigid, and unable to handle anything off its tree. A voice agent handles unscripted speech and can complete a task. Calling a scripted IVR a voice agent is the most common misuse of the term.
How much does a voice agent call cost?
Published platform rates start around USD 0.05 to USD 0.09 per minute. Once speech recognition, the language model, synthesis and telephony are stacked, all-in production cost typically lands between USD 0.11 and USD 0.33 per minute. A call handled by a person is widely put at USD 7 to USD 12.
What is the hardest part of building a voice agent?
Turn-taking, not the language model. Working out when a person has finished speaking, and cutting the agent off the instant they interrupt, is what decides whether the conversation feels alive. Good timing beats a better model with poor endpointing.
Last reviewed 30 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.