AI Voice Agents: What They Do, Cost, and When to Build One
Your booking line rings 40 times a day. Half those calls are the same three questions: are you open, do you take my plan, can I move my Thursday slot. The other half are real leads that roll to voicemail the second your one front-desk person is already on another call. Buried in that pile is the reschedule that, missed, becomes a no-show and an empty appointment you cannot backfill.
An AI voice agent is the software built to pick that call up, hold a real conversation, and finish the task while the caller is still on the line. It answers the question, moves the appointment, qualifies the lead, and writes the outcome into your systems. Not a phone tree. Not a bot reading a fixed script. A back-and-forth that reaches done before anyone hangs up.
The category got good fast, then got crowded. Vapi, Retell, Bland, ElevenLabs, OpenAI's Realtime API, and a stack of resellers wrapping those same engines all pitch the same promise in the same words. Underneath the identical pitches, most of them are orchestrating the same three or four engines, and the differences that decide whether one is usable are easy to miss until a real caller exposes them.
Under the hood, a voice agent is three models in a loop
Strip the marketing off and a voice agent is a pipeline that runs on every turn of the conversation. Audio streams in and gets transcribed by a speech-to-text model. A language model reads the running transcript, decides what to say and what to do, and calls any tools it needs. A text-to-speech model turns the reply back into a voice and streams it to the caller. Then the caller talks again and the loop repeats.
The whole thing lives or dies on latency. A gap of a second or two between the caller finishing and the agent replying reads as a dropped line, and people start talking over it. This is not a soft preference. ITU-T Recommendation G.114, the telecom standard for voice quality, puts one-way transmission delay under 150 milliseconds in the preferred band and treats anything past 400 milliseconds as degrading real conversation. A voice agent has to run speech-to-text, a model call, and text-to-speech inside a budget that tight, which is why streaming architectures and model choice matter more than the size of the feature list.
Platforms like Vapi and Retell exist to orchestrate that loop for you. They handle the audio routing, the turn-taking, and the tool calls, and they let you plug in whichever speech and language models you want. That is the layer you are actually evaluating when you compare tools, so the questions below are about that layer, not about a specific brand.
The capabilities that earn their keep
Every vendor lists 30 features. Seven of them decide whether the agent survives contact with a real caller. Press on these during a demo and ignore most of the rest.
It handles interruptions without falling apart
Real people cut in. They say "no, the other Thursday" halfway through the agent's sentence. A usable agent detects the interruption, stops talking immediately, and picks up the new input, a behavior often labeled barge-in. The failure mode is an agent that plows through its scripted reply while the caller is trying to correct it. Interrupt the demo mid-sentence on purpose and watch what happens.
It answers from your real information, not a generic script
The agent should respond with your hours, your service area, your pricing, and your policies, pulled from a knowledge source you control, not from whatever the base model guessed. When it does not know, it should say so and offer to route the call, not invent an answer. Throw an oddly specific question at the demo, something only your business would know, and see whether it grounds the reply or bluffs.
It completes the task on the call
This is the whole point. The strong agents check live calendar availability, offer real slots, and lock the booking in before the caller hangs up. Under the hood that is function calling: the language model triggers an action against your calendar, your CRM, or your database mid-conversation. An agent that only "takes a message for someone to call back" has quietly reintroduced the exact delay you were paying to remove.
It writes clean data back into your systems
After the call, the agent should create or update the contact, log a summary, and push structured fields into HubSpot, Salesforce, Pipedrive, or whatever you run. This is the piece that breaks most often and gets checked least. A booking that lands in the tool's own dashboard but never reaches your CRM means someone is still copy-pasting leads by hand. If you already wrestle with records living in two places, our take on a clean HubSpot and Salesforce integration is the same standard a voice agent has to meet.
It knows when to hand a call to a human
Some calls should reach a person, and the agent needs clear rules for which ones. A good setup does a warm transfer to a rep or a cell phone and passes a short summary so the human is not starting cold. Ask whether transfers carry that context or dump the caller into a fresh conversation where they repeat everything.
It works outbound as well as inbound
Answering calls is one job. Placing them is another: appointment reminders, payment-due nudges, lead qualification, a callback when a form comes in. Outbound is where a voice agent stops being a fancy answering machine and starts clearing a task queue your team never gets to. Confirm the platform supports outbound campaigns with your own numbers, not just inbound answering.
It logs, records, and lets you review
You cannot improve what you cannot hear. Every call should produce a recording, a transcript, and a structured outcome you can audit. The first two weeks of any voice agent are spent listening to real calls and fixing the three places it gets confused. Without that review loop, you are shipping a black box and hoping.
If a tool nails interruptions, grounded answers, in-call task completion, a clean data write, human handoff, and call review, it covers what a lean team actually needs. The rest is usually padding on the pricing tier above yours.
What an AI voice agent actually costs
Voice agents bill by the minute, and the headline number on a platform's homepage is almost never the number you pay. Most orchestration platforms charge a thin platform fee and then pass through the cost of the speech-to-text, language model, text-to-speech, and telephony you stack on top. A few bundle it all into one blended rate. Here is the current landscape from vendor pages, in USD.
| Platform | What you pay | Model | |---|---|---| | Vapi | $0.05/min platform fee, plus STT, LLM, TTS, and telephony at cost | Unbundled | | Retell AI | From $0.055/min for voice infrastructure, components on top | Unbundled | | Bland AI | $0.11 to $0.14/min depending on plan | Bundled | | OpenAI Realtime API | $32 per 1M audio input tokens, $64 per 1M audio output tokens | Token-based |
Two things to read from that table. First, an unbundled platform fee like Vapi's $0.05/min is a floor, not a total. Add a transcription model, a language model, a voice, and a phone number, and the real all-in rate climbs well above the sticker. Second, the model layer is where cost swings hardest. OpenAI prices its Realtime API by audio token, and in its own launch explainer OpenAI put that at roughly $0.06 per minute of audio input and $0.24 per minute of audio output. A chatty agent that talks a lot costs more than a terse one that mostly listens.
The practical read: at a few thousand minutes a month, per-minute pricing is fine and the platforms above are cheap enough to test. The cost that actually hurts is not the minutes. It is the integration and tuning work to make the agent reliable, and that is a one-time build, not a per-minute line. We break down that build-versus-buy math in automation: build vs buy vs hire.
The features you can skip for now
A lot of what pads the enterprise tier does nothing for a 15-person operation. Skip the 40-language voice matrix if every caller speaks one or two languages. Skip on-prem and self-hosted deployment unless a compliance rule genuinely forces it, because it triples the maintenance burden for a benefit you will not feel. Skip the sprawling analytics dashboard with sentiment heatmaps. In the first month you want three things you can read: did it book the call, did the data reach the CRM, and where did it get confused. Everything past that is a slide in a sales deck, not a capability your team will use this quarter.
How bottta builds a voice agent that fits your stack
Standing up a demo agent takes an afternoon. Getting one that answers a real line, books against your actual calendar, writes clean records into your CRM, and hands off gracefully when it should takes real integration work, and that is the part the off-the-shelf tools leave to you. It is also the part that determines whether the agent earns its keep or becomes a brittle thing nobody trusts to pick up.
That build is what bottta does. Our AI Automation work is exactly this: model routing, the language layer that decides what to say and do, and the agent logic that completes a task instead of reading a script. Our Integrations work wires the outcome into the tools you already run, so a booked call updates the calendar and the CRM without anyone touching a keyboard. We recommend starting with a fixed-scope build over stitching a platform together yourself, because the wiring is where DIY voice projects stall.
For most teams that lands as the $4K project: a defined scope, the integrations included, and 30 days of post-launch support to tune the agent against real calls, which is exactly when a voice agent needs the most attention. If you want us to keep monitoring it, adjust prompts as your business changes, and run a couple of other workflows alongside it, the $3K/month retainer covers ongoing hours across up to three active workflows with weekly calls. The platforms are a fair starting point if you have an engineer with spare cycles and a taste for maintenance. If you do not, working with a studio that builds and then watches the thing is the faster path to a line you can actually leave unattended. This pairs naturally with an AI receptionist for a small business or a customer service chatbot when the same conversation needs to happen over chat, and it is one piece of the broader case for AI automation in a small business.
Frequently asked questions
Is an AI voice agent the same as an AI receptionist?
An AI receptionist is one job a voice agent does: answering inbound calls and booking appointments. The voice agent is the underlying technology, and the same stack also handles outbound calls, lead qualification, payment reminders, and callbacks. If your need is purely answering the front desk, start with our guide to an AI receptionist for a small business.
Can an AI voice agent book appointments and update my CRM?
Yes, and this is the capability worth testing hardest. Through function calling, the agent checks live calendar availability during the call and writes the contact and call summary into HubSpot, Salesforce, or Pipedrive afterward. The integration is where most projects get sloppy, so confirm it writes the specific fields your team uses, not just a generic contact record.
How much does an AI voice agent cost per minute?
Platform fees start around $0.05/min on Vapi and $0.055/min on Retell, with the speech, language, and telephony costs stacked on top. Bundled services like Bland run $0.11 to $0.14/min. The larger cost is usually the one-time build and integration work, not the minutes themselves.
Will callers know they are talking to an AI?
Modern voice agents sound close to human when latency stays under the threshold that keeps conversation feeling live, per the ITU-T G.114 delay guidance. Many teams still choose to disclose it up front, and some regions require disclosure. The honest move is to tell callers, since a caller who feels tricked is a worse outcome than a caller who knows and gets their problem solved anyway.
Should I build a voice agent myself or hire it out?
If you have an engineer with spare time and appetite for ongoing maintenance, the platforms above are a reasonable DIY path. If you do not, the integration and tuning work is where solo projects stall, and a fixed-scope build with someone who owns the wiring gets you to a reliable line faster. Weigh it with our build vs buy vs hire breakdown.
A line that never stops ringing is a liability and an opportunity in the same breath, and the build is what turns one into the other. Start a project with bottta when you want that line answered without adding a person to answer it.