What Is a Voice Agent? Cost, Latency and Use Cases
A voice agent answers the phone, holds a real spoken conversation and books the job in your systems. This guide covers how they work, what they cost per minute, the latency threshold that decides whether callers stay on the line, what you are required to tell them, and the cases where you should not buy one at all.
On this page
A voice agent is an AI system that answers or places phone calls, holds a real spoken conversation, and completes the task the call was about. It listens, understands, decides what to do, takes action in your booking system or CRM, and speaks back. No menus, no keypad, no waiting for a human.
The category went from demo to production fast, and the marketing got ahead of the reality. This guide covers what a voice agent actually is, the one metric that determines whether callers stay on the line, what it costs per minute, where it earns money in a service business, what you are legally required to tell callers, and the situations where the honest answer is that you do not need one.
What is a voice agent?
A voice agent is a specific thing, not a synonym for phone automation. Six properties separate it from what came before.
It understands speech, not keypresses. The caller talks the way they would to a person. There is no menu tree to memorize.
It handles interruptions. When the caller talks over it, it stops, listens, and adjusts. This is harder to build than it sounds and it is the fastest way to tell a good agent from a bad one.
It takes action mid call. It checks the real calendar, holds the real slot, writes the real record. A system that just takes a message is an answering service with better vocabulary.
It follows a goal, not a script. This is what makes it an AI agent rather than a script: given the objective of booking a qualified job, it works out which questions to ask in which order for this caller.
It escalates on purpose. A defined boundary sends the call to a human instead of improvising past the edge of what it knows.
It answers honestly when asked. If the caller asks whether they are talking to a person, it says no.
Miss the first two and you have an IVR. Miss the third and you have voicemail transcription. Miss the last two and you have a liability.
How does a voice agent work?
Most production agents in 2026 run what engineers call a cascaded pipeline. Audio arrives from the phone network, gets transcribed, gets reasoned over, gets spoken back. Four components and a timer.
Telephony carries the call. A SIP trunk or a number provider connects the public phone network to your software.
Speech to text transcribes the caller in a continuous stream and, critically, decides the moment they have finished speaking. That decision is called endpointing and it is where most awkwardness comes from.
The language model decides what to say and which tools to call. It is the largest share of the delay, commonly cited at around 70 percent of total latency.
Text to speech generates the reply audio. What matters is time to first audio byte, not how long the whole sentence takes to render.
The alternative is a speech to speech model that skips the transcription and synthesis hops entirely. It is faster and it preserves tone better. It also costs more and gives you less control over the voice and the model. Published 2026 production data puts cascaded stacks at roughly $0.07 to $0.13 per connected minute against $0.18 to $0.21 for speech to speech.
Why latency is the whole product
Everything else about a voice agent is downstream of how fast it replies. Callers do not complain about milliseconds. They say the agent felt slow, or they talk over it, or they hang up.
Two things follow from that table, and both cost money if you ignore them.
First, optimize the tail, not the average. A published 2026 fleet of production deployments reported a median of 680 ms and a 95th percentile of 1,180 ms, and treats under 1,400 ms at the 95th percentile as the real bar. One call in twenty is the one the caller tells people about.
Second, the tail is almost never the model. It is the synchronous lookup. When the agent has to ask your calendar for availability in the middle of a turn, your calendar's response time becomes your agent's response time.
From our work. We built a phone receptionist for mechanic shops that answers inbound calls and books appointments entirely by voice, with one deployment serving many shops and routing by the number the caller dialed. Stack is LiveKit Agents for orchestration, Deepgram for speech to text and speech synthesis, an OpenAI model for reasoning, and Silero for voice activity detection.
Two things we learned the hard way. We had to disable preemptive generation, which is the trick of starting a filler reply while the real answer is still being computed, because it was dropping the tool call afterward and leaving the agent silent on the line. Silence is worse than slowness. And most of our tuning time went into voice activity detection thresholds and tool round trips rather than into the model, which is exactly where the published benchmarks say the tail lives.
Voice agent vs IVR vs answering service
Four options for a ringing phone. They are not close substitutes.
VoicemailIVR menuAnswering serviceVoice agent Caller talks naturallyYes, to nobodyNoYesYes Books the jobNoRarelySometimesYes Available at 2amYesYesUsually, at a premiumYes Handles ten calls at onceYesYesNoYes Cost basisFreeLow, fixedPer call or per minute, human ratesPer minute, cents Fails byCaller never calls backCaller presses zeroTaking a message you still have to actionConfidently mishearing something
The comparison that decides most purchases is the last two columns, and the deciding factor is rarely price. It is whether you want a message or a booking.
What does a voice agent cost?
Three layers, and vendors quote whichever one flatters them.
Per minute usage. Roughly $0.07 to $0.21 per connected minute in published 2026 production data, depending on architecture. Platform list prices are often lower because they exclude telephony, transcription, model tokens, or synthesis.
Build or configuration. A no code platform gets you answering calls in an afternoon. Getting an agent to book correctly into your actual scheduling system, handle your trade's vocabulary, and escalate the right calls is engineering work measured in weeks.
The tail you did not budget for. Call recordings and transcripts need storage, disclosure scripts need maintenance, and someone has to read the failures every week for the first month.
Compare it against the fully loaded cost of the person currently answering, plus the revenue on the calls nobody answers at all. That second number is usually the larger one and almost nobody measures it.
Where a voice agent actually earns its money
Inbound, after hours, high urgency, low judgment. That is the shape of the winning use case, which is why the trades adopted this before the enterprise did.
After hours triage. Emergency calls arrive when nobody is at the desk, and roughly 76 of the 168 hours in a week fall outside standard business hours.
Overflow. The second, third and fourth simultaneous call on a Monday morning that currently go to voicemail.
Booking and rescheduling. The single most reliable voice agent task in 2026, because the goal is narrow and the failure is cheap.
Qualification before dispatch. Capturing the make, model, address, and symptom so the technician arrives with the right part.
Outbound reminders and confirmations. Effective, and the most legally constrained thing on this list. See the compliance section.
Do the math before you buy anything
Be skeptical of the statistics in this market. Published missed call rates for the same sector range from 14 percent to 62 percent, and the sources are mostly companies selling call answering. One of the more credible datasets, CallRail's analysis of 1.1 million leads, reported missed call rates of 32 percent in healthcare, 28 percent in legal, 14 percent in home services and 9 percent in real estate. That is a very different picture from the numbers in most vendor headlines.
You do not need the industry average. Your phone system has your own number in it.
Run that with real figures before you take a demo. If the honest annual leak is smaller than the cost of answering the phone properly, buy nothing and go back to work. If it is large, book a discovery call and bring the number with you.
What you must tell callers
This is the section most vendor guides skip, and it is the one that carries financial risk. General information, not legal advice, and it changes quickly.
For inbound calls the exposure is mostly disclosure and recording. Several US states have bot disclosure or AI transparency requirements, including California's SB 1001 and Texas SB 140, which requires disclosure within the first thirty seconds. Recording consent varies by state between one party and all party consent. In the EU, the AI Act's Article 50 transparency obligations apply from August 2, 2026 and cover systems that interact directly with people.
For outbound calls the exposure is much larger. The FCC confirmed in February 2024 that AI generated voice counts as an artificial voice under the TCPA, which means prior consent is required. Statutory damages run from $500 to $1,500 per call with no cap. A further FCC proposal, from September 2024, would require explicit AI disclosure at consent and at the start of every call. As of mid 2026 that proposal has not been finalized, but enforcement of the underlying rule is active and does not wait for it.
Two rules to build in regardless of jurisdiction. Disclose up front, in the first sentence, because the research on disclosed calls shows completion rates hold up and callers dislike being deceived more than they dislike a bot. And if a caller asks whether they are speaking to a person, the agent says no, every time. That one is not a compliance preference, it is consumer protection law.
How to evaluate a voice agent before you sign
Ask for the p95, not the average. If the vendor only quotes a median, they are hiding the tail.
Ask how latency is measured. The only comparable number runs from the end of caller speech to the first audio byte the caller hears. Component level benchmarks are marketing.
Interrupt the demo. Talk over it mid sentence. Give it a postcode wrong and correct yourself halfway through. Ask it something outside its scope.
Ask what happens when your calendar is slow. The answer should involve a filler phrase and a timeout, not silence.
Check the escalation path. Who does it transfer to, when, and what does the human see when they pick up.
Read the disclosure script. If there isn't one, that is your answer.
The twenty minute version
Pull last month's call log and count the unanswered calls, then split them by hour. Multiply the after hours count by your average job value and your inbound close rate. That single number tells you whether this is a business decision or a toy, and it takes twenty minutes to produce.
Where Codeatic fits in
We build voice agents for service businesses, starting with the inbound phone line because that is where the money is provably leaking. Our current work is a phone receptionist for mechanic shops, and we are extending it into a productized offer for Canadian home services businesses, with demo agents for after hours HVAC and plumbing triage, water damage and restoration intake, and parts capture for appliance and garage door work.
That work sits alongside our wider AI automation practice, and the Ziltrix case study covers a voice agent inside a dispatch workflow. Our AI opportunity audit scores your workflows and returns a costed shortlist. We would rather you ran the worksheet above first. If the number is small, we will tell you not to buy.
When not to use a voice agent
Three cases. When the call is the sale and the relationship is the product, as with high value consultative work where a human voice is the differentiator. When your inbound volume is low enough that the owner genuinely answers every call, because then you are solving a problem you do not have. And when the conversation requires judgment with legal or clinical exposure, where the correct build is capture and route, not decide.
The short version
A voice agent answers the phone, talks like a person, and finishes the job in your systems. Judge it on one number, the delay between the caller finishing and the reply starting, and target under 800 milliseconds at the median with the 95th percentile under 1,400. Expect to pay cents per minute and weeks of engineering to make it book correctly. Start with after hours inbound, disclose that it is AI in the first sentence, and give it a human to hand off to. Then measure whether it actually recovered the calls you were losing.
Frequently asked questions about voice agents
What is a voice agent in simple terms?
Software that answers the phone, holds a normal spoken conversation, and gets the task done. It books the appointment or updates the record rather than taking a message for somebody else to action.
What is the difference between a voice agent and an IVR?
An IVR gives you a menu and waits for a keypress or a single word. A voice agent lets the caller talk normally, handles being interrupted, works out what they need, and takes action. Different technology and a completely different caller experience.
How fast does a voice agent need to respond?
Under 800 milliseconds from the end of the caller's sentence to the first audio back reads as a natural pause. Between 800 and 1,200 is acceptable for business calls. Past 1,500 the caller can tell it is a machine and starts talking over it.
How much does a voice agent cost per minute?
Published 2026 production figures run roughly $0.07 to $0.13 per connected minute for a cascaded stack and $0.18 to $0.21 for speech to speech. Platform list prices are often quoted without telephony, transcription or model costs included, so ask what the number covers.
Can callers tell they are talking to an AI?
Often not immediately, which is exactly why disclosure matters. Several jurisdictions require it, and if a caller directly asks whether they are speaking to a person, the agent must answer honestly.
Is it legal to use an AI voice agent on the phone?
Answering your own inbound calls is generally straightforward, subject to disclosure and call recording consent rules that vary by state and country. Outbound calling is far more regulated: in the US, AI generated voice counts as an artificial voice under the TCPA and requires prior consent, with statutory damages from $500 to $1,500 per call. Get advice specific to your jurisdiction before running outbound campaigns.
What happens when the voice agent does not understand?
In a well built agent, it asks once for clarification, then transfers to a human with the conversation context attached. In a badly built one it guesses, which is the failure mode to test for during a demo.
Do I need a voice agent or just better call answering?
If your calls arrive mostly during business hours and someone answers them, better call answering is cheaper and simpler. A voice agent wins when the calls arrive after hours, arrive in bursts you cannot staff for, or need to become bookings without a second human touching them.
Abdul, Founder and CEO, Codeatic
Abdul has spent five years building software, starting as an Android engineer and now leading Codeatic, an AI automation agency working with SMBs and startups across the US, Canada and Saudi Arabia. He holds an MS in Computer Science from PUCIT. His team builds and runs production voice agents, including a phone receptionist handling inbound booking calls for mechanic shops. Connect on LinkedIn.