Insights · 12 min read
A practical guide to AI voice agents: the real-time stack they need, per-minute running costs, build costs, and the use cases where voice automation pays off.
By GGP Editorial
A voice agent is the thing that answers a phone line and talks like a person. Not a press-one-for-sales menu. Not a recording that says "your call is important to us." A system that hears what a caller says, decides what to do, and speaks back in a natural voice, in real time.
I write this from the vendor side. GlobeSoft is a software company founded in 2018 with 40-plus engineers and more than 300 delivered projects, and we build the AI systems behind customer-facing products. Voice is where a growing share of those projects are heading. This guide covers what a voice agent actually is, the use cases where it pays off, the real-time stack it needs, and what it costs to build and run one.
If you are still deciding whether any AI project is worth the budget, start with how much does AI development cost.
The chatbot comparison confuses more projects than it helps. A text chatbot has seconds to think, can hand the user a button or a link, and nobody notices a two-second pause. A voice agent has a few hundred milliseconds before the silence gets awkward, it has no screen to fall back on, and it has to deal with people interrupting mid-sentence, background noise, and a caller who mumbles a phone number. The text-only version of this product is a different build, covered in how to build an AI customer service platform.
It is also different from a traditional IVR. An IVR plays a fixed menu and routes by key press. A voice agent listens to open-ended speech and can hold an actual back-and-forth. And it is different from the general agents we wrote about in AI agent development, which mostly reason over text and run on a schedule. A voice agent runs on a live call.
The practical consequence is that voice is a latency problem before it is an intelligence problem. The model quality matters, but so does the plumbing that gets audio in and out fast enough.
Voice agents pay off where a call is currently answered by a person doing something repetitive. The pattern is the same every time: a high volume of calls, a limited set of things callers want, and a person on the other end who could be doing something better.
| Use case | What the agent does | Where the savings are |
|---|---|---|
| Inbound support | Answers common questions, looks up orders, resets passwords | Fewer agents on repetitive tickets |
| Appointment booking | Books, reschedules, and cancels over the phone | No missed after-hours calls |
| Outbound reminders | Payment due, appointment tomorrow, prescription refill | Replaces manual dialing |
| Qualification and intake | Collects details before routing to a human | Humans talk to qualified leads only |
| Front desk and reception | Answers, screens, and routes callers | Fewer missed calls, less admin |
| Collections | Reminds on overdue balances, works out simple payment plans | Lower cost per collection |
The worst case for a voice agent is a call where the person needs judgment, empathy, or an exception to policy. The best case is a call with a script that mostly stays the same. Most deployments start with appointment booking or order status and expand from there.
A voice agent is a loop that runs many times per second while the call is live.
The caller speaks. A speech-to-text model transcribes what it heard. That text goes to a language model, which decides what to say next and sometimes calls a tool, like looking up an order or checking a calendar. The reply text goes to a text-to-speech model, and the audio goes back down the phone line.
Four things make this hard.
Latency. From the moment the caller stops talking to the moment the agent starts speaking, you want well under a second, ideally a few hundred milliseconds. Every component in the loop adds its share. This is why voice agents use streaming models that start transcribing and speaking before a full sentence is finished, rather than the batch pattern most text chatbots use.
Interruption. People talk over the agent. The system has to notice the caller is speaking, stop the agent mid-sentence, and switch to listening. This is called barge-in, and getting it right without cutting the caller off at a breath is one of the fiddliest parts of the build.
Tool access. A voice agent that only chats is a novelty. One that can look up an order in your database, check a calendar for open slots, or take a payment is a product. The tool-calling layer is where most of the integration work lives.
Telephony. The agent has to live on a real phone number, which means SIP trunks, call routing, and holding a call session open. This is plumbing, but it decides whether the agent picks up at all.
On top of that, most businesses need call recording and a transcript for compliance or QA, which adds storage and access-control work. The compliance thinking we described for regulated flows in security requirements for FinTech platforms applies just as much to any call that touches payment or personal data.
You have three paths, and they suit three situations.
Assemble on a hosted platform. Vapi, Retell AI, Bland AI, and LiveKit are the names that come up most. You pick a speech model, a language model, and a voice, wire up your tools, and you have a working agent in days. This is the fastest way to prove a use case, and it is what most teams should do first.
Buy raw APIs and build the loop yourself. OpenAI's realtime API, Deepgram for transcription, ElevenLabs or Cartesia for speech. You own the loop, the latency budget, and the integrations. More work, but you get control over cost and behavior at scale.
Build it fully custom. You run your own models, manage the telephony, and own the whole stack. Almost nobody should start here. It makes sense when voice is your core product, or when data has to stay on-premises for regulatory reasons.
The sensible path for most companies is to start on a platform, prove the use case and the call volume, then decide whether the per-minute cost justifies moving the loop in-house. That decision is the same one we walk through in build vs buy AI software.
Voice agents are priced per minute, and the number is smaller than most people expect, which is why the category is growing.
Here is a realistic split using published list prices from the major platforms as of this writing. These move, so treat them as a guide, not a quote.
| Component | Typical cost | Notes |
|---|---|---|
| Speech to text | $0.01 to $0.02 per minute | Streaming transcription |
| Language model | $0.01 to $0.05 per minute | Varies by model and reasoning depth |
| Text to speech | $0.02 to $0.04 per minute | Higher for premium voices |
| Telephony | $0.01 to $0.02 per minute | Per leg, inbound or outbound |
| Platform hosting | $0.00 to $0.05 per minute | Zero if you run the loop yourself |
All in, a platform-based voice agent runs roughly $0.10 to $0.25 per minute of talk time. A call center paying a person $20 an hour is paying about $0.33 per minute plus benefits, desk space, and scheduling, so the economics usually work even before you count after-hours coverage. A high-volume deployment can push the per-minute cost lower by running parts of the loop in-house and buying telephony in bulk.
The number that matters more than the per-minute rate is containment. That is the share of calls the agent finishes without handing off to a person. A cheap agent that escalates every third call to a human saves less than a slightly more expensive one that closes most calls on its own.
If you assemble on a platform, the build cost is mostly your time and the integration work. A first voice agent for one use case, wired to one or two tools, can go live in a few weeks for a few thousand dollars in setup and integration.
If you build the loop yourself or need a production system around the agent, the numbers look more like a normal software project.
A focused agent for one use case, with the loop running on managed APIs, a telephony connection, call recording, and a simple admin view, typically lands between $30,000 and $70,000. A production platform with several use cases, deep integrations to your CRM or order system, a review dashboard for transcripts, and compliance controls runs $100,000 to $250,000 or more. These are ranges, not quotes, and the only number that matters is the one from a scoped proposal. We wrote about getting a real number in how to estimate a custom software project.
The cost drivers, roughly in order of impact:
A platform-based agent for one use case can go from idea to live call in two to four weeks. A custom-built loop with integrations, call analytics, and compliance controls runs three to five months. The schedule is usually set by the integrations and the compliance review, not by the speech models.
Plan to spend the first week on the call scripts and the tool calls, not on the model. If you do not know what the agent should say when a caller asks something off-script, the model will not figure it out for you.
Chasing the demo over the calls. A two-minute demo with a perfect voice is easy. Real calls are full of interruptions, mumbling, and callers who go off-script. Test on real calls early and you will find the gaps before launch.
Ignoring barge-in. If the agent cannot be interrupted gracefully, callers hang up or shout over it. This single detail separates a pleasant call from an infuriating one.
No human fallback. Every voice agent needs an escape hatch to a person. The question is when and how it hands off, and how much context goes with the caller.
Treating voice like text. The prompt that works in a chatbot does not work spoken aloud. Long answers, lists, and disclaimers sound terrible on the phone. The script has to be written for the ear.
Skipping containment tracking. If you do not measure how many calls the agent closes on its own, you cannot tell whether it is saving money or just adding a phone bill. Log every call, tag the outcome, and watch the handoff rate.
It is software that answers a phone call, understands spoken language, and replies in a natural voice, in real time. It can look up information, book appointments, take payments, and hand off to a person when needed.
Roughly $0.10 to $0.25 per minute of talk time on a hosted platform, depending on the speech model, voice, and telephony. Running parts of the loop in-house can bring that down at higher volume.
A focused agent for one use case on managed APIs typically lands between $30,000 and $70,000. A production platform with integrations and compliance controls runs $100,000 to $250,000 or more. These are ranges, not quotes.
Start on a platform. Prove the use case and the call volume first. Move the loop in-house later if the per-minute cost or the level of control justifies it.
It is the ability for a caller to interrupt the agent mid-sentence and have the agent stop and listen. Getting it right is one of the hardest parts of a voice agent and one of the biggest differences between a natural call and a frustrating one.
Yes, but payments and personal data add real compliance obligations, including consent prompts, call recording rules, and data handling requirements. Treat that as scope from the start rather than an add-on.
Two to four weeks for a platform-based agent with one use case. Three to five months for a custom-built loop with integrations and compliance controls.
GlobeSoft is a software development company founded in 2018, with 40-plus engineers and more than 300 delivered projects for over 100 clients across the US, Brazil, South Africa, and Singapore. We build custom software, AI products, SaaS platforms, FinTech systems, ERP and CRM, and mobile apps on a stack that includes Java, Spring Boot, Go, Python, React, Vue, and Node. Voice is a natural extension of the conversational AI and automation work we already ship, which you can see in our AI development services.
If you are planning a voice agent, a broader AI project, or a customer platform, talk to us about your project. Tell us the call volume, the use case, and the systems it needs to touch, and we will help you work through the platform-versus-custom call, the stack, and what to ship first.
Tell us what you are building and where you are today. We typically reply within 24 hours.
