What is an AI voice agent — and how does it actually answer your calls?

An AI voice agent is software that answers your business phone, understands what the caller is asking in plain speech, and then does something about it — routes the call, books an appointment, answers a question, or takes a message and gets it to the right person.

It is not a phone tree. Nobody presses 1 for sales. The caller just talks, and it responds.

I’m Robert Harney. I’ve built phone systems for Chicago businesses since I started Open One Solutions in 2009, and no question has come up more in the last year than this one — usually some version of “is this real, or is it another thing that only works in a demo.” So here’s the plain-English version, including the part where I tell you when not to buy one.

Talk to a real engineerFind out if an AI voice agent fits your call volume

What is an AI voice agent?

An AI voice agent is an automated system that answers your phone using natural conversation instead of a menu. It listens to what the caller says, works out what they need, and takes an action — transferring them to the right person, booking time on a calendar, answering a common question, or capturing a message with the details that actually matter.

An “AI receptionist” is one kind of voice agent — the front-desk version. It is not the only kind, and for a lot of businesses it is not the most valuable one.

The real change from everything that came before is that the caller no longer has to learn your system. They just say why they called.

The useful way to think about it: a digital employee

Forget the technology for a second. What you are actually doing is hiring.

1. You write the job description. What is this agent responsible for, and what is it not?
2. You train it on information you approve — your policies, your FAQs, your documents. Not the open internet.
3. You give it tools. A calendar. A CRM. An intake form. Whatever the job needs.
4. You set permissions and guardrails. What it can do, what it must never do, what it has to escalate.
5. You test it before it talks to a customer.
6. Then you listen, and you coach it. This is the part almost nobody tells you about.

You would not hire someone, train them for a day, and then never check in again. An AI agent is no different. We review real calls, look at where it struggled, and adjust. That is the difference between AI answering your phone and AI representing your business.

The control system around all of this has a name in the industry — the harness. It is what decides what the agent can see, what it can do, when it must stop, and what it is never allowed to say. The model gets all the attention. The harness is what makes it safe to put in front of your customers.

How does an AI voice agent answer a call, step by step?

Here is what happens in the two or three seconds after the call connects.

1. The call arrives on your phone system, the same as any other call. The agent answers on the ring you choose — immediately, after hours only, or when everyone is busy.
2. It greets the caller with your script, in your business’s name. You write that greeting.
3. The caller speaks normally, and the agent converts speech to text.
4. It works out the intent and checks it against what the agent is permitted to handle.
5. It acts — transfers, looks something up, books the appointment, or takes the message.
6. It responds out loud, in a voice you picked.
7. It hands off when it should. To a person, or sometimes to another agent.

The loop takes about as long as a human pause. Configured well, callers stop noticing the mechanics and just have the conversation.

What can it actually do?

More than most people expect, and the honest framing is that it does a defined set of things reliably rather than everything adequately.

1. Answer every call on the first ring, at 2 p.m. or 2 a.m.
2. Route based on what the caller said, not on which number they pressed.
3. Book, reschedule, or cancel appointments on a real calendar with the caller’s details attached.
4. Take a real message — the name, the number, the reason, the urgency — and put it where it needs to go.
5. Answer common questions from information you supplied.
6. Screen calls, so the interruptions that reach your team are the ones worth having.
7. Handle more than one language. More on that below, because it surprises people.

We connect agents to cal.com, Calendly, Google, Microsoft, HubSpot, and Salesforce today. If your system has an API, it can usually be connected with some custom work.

What it does not do is improvise. If you have not given it the information or the permission, it does not invent an answer. That is a design decision, and it is the first thing to ask any vendor about.

How is it different from the old phone tree everyone hates?

An IVR — “press 1 for sales, press 2 for support” — is a decision tree. You built a map and the caller has to walk it. If what they need is not on the map, they press 0 until something happens, or they hang up.

An AI voice agent starts from what the caller said. Someone who calls and says “my phones are down and I have a warehouse full of people who cannot take orders” reaches support immediately. That sentence matches no menu option, and an IVR would have made them listen to five of them first.

Can one call involve more than one agent?

Yes, and this is where it gets genuinely useful.

A good agent does not need to know everything. It needs to do one job well. So instead of one agent that half-knows your entire business, you can run several specialised ones — a scheduling agent, an intake agent, an after-hours agent — each with its own job, its own tools, and its own guardrails.

When a call needs a different specialty, it transfers between agents. Same voice, same conversation. The caller does not know anything happened, and does not need to.

Does it sound robotic, or human?

Better than you are expecting, and not indistinguishable from a person. That is the honest answer.

Current voices are natural enough that most callers stop thinking about it after a sentence or two. But there is usually a small tell — a beat of delay, or a cadence that is a little too even. Anyone who tells you callers cannot tell is selling you something.

We never hide that it is AI

If a caller asks whether they have reached a person, our agent says no and offers to get one. We put disclosure language on the website, and the agent acknowledges what it is.

That is not only an ethics position. The FCC decided in February 2024 that an AI-generated voice on an outbound call is treated the same as a robocall — consent required, caller identification required. That has been in force for over two years. A separate federal rule that would specifically mandate telling callers they are speaking with AI is still proposed rather than final, and some states have gone further on their own. Anyone using AI for outbound calling should confirm what applies to them.

I know a company that is proud that when they use AI to call old clients about mortgage rates, most people cannot tell. That is not a feature. It is a liability, and it is deceptive.

If it is AI Bob on the phone and it sounds like me, the caller needs to know it is AI Bob and not the real McCoy. You do not build a business on fooling the people who call you.

What happens when it does not know the answer?

This is the question that separates a system that helps people from a system that traps them, and it is the one most buyers forget to ask.

First, the agent should say it does not know rather than guess. That sounds obvious. It is not — a language model’s natural failure mode is to produce a confident answer whether or not it has one. You control that three ways: narrow the job, ground the agent only in sources you approve, and limit which tools it can reach. Then you require it to say “I do not have that” and hand off.

Ask any provider to state their handoff rules in one breath. If they cannot, they have not thought about it, and you are the one who will hear about it from your customers.

Multilingual staff in one agent

Our agents do not ask anyone to press 1 for English. The caller talks in whatever language they speak, and if that language is in the system, the agent detects it and holds the whole conversation in it.

Think about what that replaces. You are not hiring for language coverage, not routing callers to the one person who speaks Polish, and not losing the call because nobody could help. If your customers speak more than one language, this is usually the feature that decides it.

Is it secure — and who owns your data?

Ask this before you ask about anything else, because almost nobody does.

Every call your agent handles is your business, your customers, and your information. A lot of the AI tools you can sign up for online keep ownership of that data, and use it to train their models. Which means your call data could end up improving a product your competitor is using.

So ask any provider four questions:

1. Is this a closed system, or is it running on the open internet?
2. What model is behind it, and where does it run?
3. Is my data used for training?
4. Do I still own it?

Our answers: your data stays private, and you own it. We are adherent to PCI and HIPAA requirements, and the system is closed rather than open to the public internet.

If you are in healthcare, legal, or financial services, this is not a detail. It is the whole decision.

Four questions worth askingGet straight answers on security and ownership

What does it cost — and the question nobody asks

There is usually a one-time setup fee — ours start around $100, depending on how much configuration and integration the job needs — and then usage-based pricing on top of your phone service.

Everybody in this market quotes a per-minute rate. Expect somewhere between twenty-five and fifty cents a minute, call it forty on average.

That is not the question you should be asking.

Do you bill by the minute, or by the second?

Most providers round up to the full minute. A sixty-one second call bills as two minutes. Three separate twenty-second calls — one minute of actual conversation — bill as three.

We bill by the second.

So our per-minute rate can look slightly higher than someone else’s. Then you run the analysis against real call data, and we come out twenty to forty percent less on the same volume.

It is the same thing I tell people about taxes and fees on a phone bill, and about renting versus buying your desk phones — see our business phone system cost breakdown for more on that. The advertised rate is not where the money is. The billing mechanics are where the money is.

So ask the question: how am I billed?

How long does it take to set up?

Anywhere from the same day to several weeks. A straightforward agent — answer, route, take messages, book on a calendar we already connect to — can go live quickly. What extends the timeline is custom integration work into systems that need it, plus testing, which is not the part to rush.

Where an AI voice agent is not worth it yet

I sell these. I am also going to tell you when not to buy one.

Start with the question nobody asks: do your customers want it?

I know a medical practice specialising in elder care that put an AI agent on the phone and lost patients over it. Their people did not want to talk to a machine, and the system they had bought was half-baked and never properly tested. Both things were true, and either one alone would have hurt.

Know your audience before you automate the front door. Some customer bases will take to this immediately. Some will not, and no amount of good engineering fixes that.

Then the easier ones

1. Low call volume. Fifteen calls a day answered on the second ring is not a problem that needs solving.
2. No repeatable category of call. If every call is genuinely different, there is nothing for an agent to take off your plate.
3. Nothing to connect it to. The value multiplies when it can write to a calendar or a CRM. If those do not exist yet, fix that first.
4. No owner. It needs someone who can say “change that.” Without one, it drifts.

Where it earns its keep fast: a busy front desk, callers who speak more than one language, after-hours calls you are currently losing to voicemail, appointment-driven scheduling, and small teams getting interrupted all day by calls that could have been routed without them.

If that is you, it pays for itself quickly. If it is not, I will tell you.

How Open One builds it

There is a version of this product where you fill out a web form, pick a voice, and hope. That is not what we do.

An engineer sits down with your actual call flow. Who answers what today. What the common questions really are — which is almost never what the owner thinks. Which calls have to reach a human inside a minute. What happens at 5:01 p.m. We configure against that, test with real calls before it touches a customer, and train your team on site so somebody in the building understands how it works.

Then we stay. We review calls, watch for the places it struggled, and coach it. And when the business changes — a new location, a new service, a new person answering — you call us and a person picks up. Same as everything else we do.

What to do next

If you want to know whether this fits your business, the fastest way to find out is a short call where I ask about your call volume and what your team gets interrupted by. Ten minutes, and you will get a straight answer either way — including “not yet,” if that is the answer.

Frequently asked questions