AI Voice Agents: 24/7 Phone Support That Doesn’t Sound Like a Robot

It’s 7:40 pm on a Tuesday. A customer with a machine down calls their usual supplier. It rings five times and goes to voicemail. They hang up, search Google and call the next one. That call never shows up in any sales report: it isn’t a lost opportunity someone can measure, it’s an opportunity that never existed.

The phone is still the channel where a small business has the most money riding on every minute — and the one it handles worst. The classic alternative, a touch-tone menu, doesn’t solve the problem; it redistributes the frustration. “Press 1 for accounts, press 2 for support” is a polite way of asking customers to do the triage work the company should be doing. AI voice agents flip that around: they listen, understand and act.

Let’s be precise about the scope. This isn’t only about taking bookings or orders over the phone — we covered that in detail in AI voice assistants for reservations and orders. This is about a company’s general phone support: who picks up at 7:40 pm, what they can settle without bothering anyone, and exactly when the call has to reach a human.

The problem isn’t call volume, it’s opening hours

In most small companies, the person answering the phone is also invoicing, picking orders and serving the counter. The outcome is predictable: calls cluster at peak times, collide with everything else, and outside office hours nobody picks them up. And someone calling about a breakdown, a delivery date or an order status rarely leaves a message — they call the next supplier.

The market has spent years papering over this with headcount or touch-tone menus, and neither scales well. Expanding reception is expensive and still doesn’t cover nights or weekends; the touch-tone menu saves the company time at the customer’s expense. Gartner estimated that by 2026 one in ten support interactions would be handled automatically, up from the 1.6 % of conversations that were automated when the forecast was made (Gartner, 2022). The same analysis puts the saving in contact-centre labour costs at 80 billion dollars for 2026, in a setting where staff can account for as much as 95 % of a contact centre’s cost (Gartner, 2022).

What AI voice agents are, and how they differ from IVR

An IVR follows a fixed tree: options, tones, a destination. It understands nothing; it only routes. A text chatbot understands, but it lives on your website and expects the customer to type. An AI voice agent combines three pieces — real-time transcription, a language model that decides, and a natural synthetic voice — and adds the one that actually matters: access to your systems while it talks.

That is what separates a pleasant recording from something useful. An agent wired into the ERP can answer “order 4417 shipped yesterday, it arrives tomorrow morning” without transferring anyone. One wired into the calendar can book the technician’s real free slot. One wired into the CRM recognises the caller by their number and doesn’t make them repeat their details. It’s the same logic we set out in enterprise AI agents versus chatbots, applied to the phone.

How you build one, step by step

1. Decide what goes through it. Not every call: start with the repetitive, well-bounded ones (order status, opening hours, lead times, appointments, billing details).
2. Connect it to your systems. The agent queries the ERP, CRM or calendar over an API. Without that connection it’s just an answering machine with good diction.
3. Define the script and the limits. What it may state, what it must never invent (prices, lead times, availability) and what it does when it doesn’t know.
4. Programme the handover. On customer request, on sensitive topics (complaints, serious incidents, cancellations) or below a confidence threshold.
5. Measure and correct. Calls answered out of hours, resolved without transfer, average duration and escalation reasons. Transcripts let you tune the script every week.

Step 4 isn’t an implementation detail: it’s the difference between a tool that helps and one that burns customers. The criteria are the same we apply to any automation that acts on its own, and we develop them in human-in-the-loop: how to supervise AI agents.

What the law requires: tell people they’re talking to an AI

There’s no room for interpretation here. Article 50(1) of Regulation (EU) 2024/1689 (the AI Act) requires AI systems intended to interact directly with natural persons to be designed so that those persons are informed they are interacting with an AI system. The Regulation’s transparency obligations have applied since 2 August 2026, according to the timeline published by the European Commission, so they are already in force.

In practice this is settled with one line at the start of the call (“you’re speaking with the virtual assistant at…”) and a clear route to a person. Far from being a drawback, it helps: the caller calibrates their expectations and stops trying to work out whether anyone actually picked up. For the full picture of how the Regulation affects a small business, we break it down in the EU AI Act for SMBs.

Cost per minute, and how it compares with more reception staff

The advantage of voice is that cost is measurable by the minute. The platforms used to build these agents publish their rates: Vapi charges $0.05 per minute of platform time and, adding transcription, language model and voice, its own calculator puts the total at roughly $0.08–$0.13 per minute (Vapi, published pricing 2026). On those numbers, a thousand conversation minutes a month — around 300 three-minute calls — land in the region of $80 to $130 a month in consumption, on top of setup and maintenance.

The honest comparison isn’t “voice agent versus receptionist”, because they don’t do the same job. It’s: what does it cost to cover the hours nobody covers today? There, the voice agent competes with the answering machine, not with a person. And for the hours that are covered, its role is to filter out the repetitive calls so whoever answers the phone can spend their time on what needs judgement.

When does an AI voice agent make sense?

It pays off when several of these are true:

• Calls come in out of hours and end in voicemail today, or simultaneous calls get dropped at peak times.
• A large share is repetitive: the same question from a different customer (order status, hours, lead times, appointments).
• The answer lives in a system you can reach over an API. If the information only exists in someone’s head, there’s nothing to automate.
• Volume justifies the build: below a few dozen calls a day, a good call-forwarding setup plus a properly staffed WhatsApp usually works out better, as we explain in WhatsApp Business + AI.

And it doesn’t pay off when the call is the product: consultative selling, negotiation, delicate complaints, or any conversation whose value lies in the person listening. There, the voice agent should do nothing more than identify the topic and pass the call on quickly and well briefed.

Frequently asked questions

Can people tell it’s an AI answering the phone?

Today’s synthetic voices sound natural and latency is low enough for fluent conversation, so many callers wouldn’t notice. Legally the question is moot: the AI Act requires you to disclose it either way. What gives away a bad agent isn’t the voice — it’s failing to understand anything off-script, or not knowing how to hand over to a person.

Can a voice agent query my ERP during the call?

Yes, and that’s precisely what makes it useful. It connects over an API to the specific area it needs (orders, calendar, customer record), with read-only permissions for anything it only has to look up. If your ERP has no open API, there are usually alternatives: an intermediate view, a scheduled export or a custom connector.

What happens when the agent doesn’t know the answer?

It should say so and transfer. A well-configured agent never improvises prices, lead times or availability: if the data isn’t in a system, it admits as much and either passes the call on or takes details so someone can call back. That rule is set when you build it, and it’s the most important one of all.

How long does it take to get one running?

An agent that handles a couple of well-bounded call reasons and knows how to transfer takes days, not months. What stretches the project out is integrating with closed systems and the script-tuning phase, which happens by listening to real calls over the first few weeks.

Will it replace my receptionist?

Not typically: it covers what nobody covers today — nights, weekends, peaks — and takes the repetitive calls off the desk. You still need the person answering the phone, but they stop spending the morning repeating the same opening hours twenty times over.

If the 7:40 pm scene sounds familiar, the first step isn’t signing up for a platform: it’s checking how many calls arrive out of hours and what they’re about. With that number on the table, the decision makes itself.

	IVR menu	AI voice agent
Understands natural language	No, tones only	Yes
Queries the ERP or CRM mid-call	No	Yes, over an API
Answers out of hours	Takes a message	Converses and resolves
Hands over to a person	Fixed transfer	By topic or confidence threshold
Cost	Fixed, per phone system	$0.08-0.13/min (Vapi, 2026)

Contact us and we’ll analyse your case for free →