Last week I booked a table at a restaurant. A friendly voice picked up, asked how many of us, what time, any allergies, smoking or non-smoking — and ended the call. I wrote it in my calendar and silently thanked them for not keeping me waiting.
It only hit me afterwards that it hadn't been a human.
There's no sci-fi to it. A standard cascade voice AI, probably Vapi or Retell behind it, local-language TTS, sub-one-second latency. On the restaurant's side, nobody stayed up until half past ten to take my booking. Next morning the owner opens the backend and finds 14 new reservations for that night.
This is where we are now. And plenty of companies are still thinking "maybe next year".
Why it works now and didn't ten years ago
"AI phone bot" used to be a synonym for "annoying IVR menu". Then three things lined up:
Speech recognition in real time is now 95–98% accurate. Language models in streaming mode respond in under 300 ms. Synthetic voices have reached the point where most callers don't notice.
Together that gives you sub-one-second response times. Meaning: a natural conversation. Not "please hold while I process your request" — but "alright, that's a table for four on Saturday at eight".
This isn't another tech curiosity. It's a strategic tool.
What it's for — and what it isn't
Let me reassure everyone: no, AI isn't going to call your sick grandmother in hospital.
The one rule I keep telling people: if the call is emotional or trust-based, don't automate it. If it's informational or procedural, do automate it.
Where it works well:
- Repetitive inbound: opening hours, bookings, package status
- Informational outbound: reminders, satisfaction surveys
- Lead qualification ahead of sales (30% of inbound leads aren't relevant anyway)
- 24/7 coverage in smaller towns, or in a second language where you don't have a native-speaking agent
Where not to put it:
- Crisis lines
- Complex legal or medical advice
- Highly emotional complaint calls
- Product configuration with 15 branching paths
That line shouldn't be drawn from the tech side ("can we do it?"). Only from the customer's side: what do they want to hear on the other end?
The three most expensive myths
"I'll automate it 100%." You won't. Realistic containment lands at 70–85%. The 100% target is a myth, and the people chasing it lose the whole thing on the last 5% — because that's where an AI ends up brushing off a frustrated customer. Plan the human handoff, and pass the context along so the human agent doesn't have to "start over".
"It'll be cheaper on direct cost." Maybe not. A solid premium stack ($0.15–0.30/min) at a mid-sized service desk can look slightly more expensive at first glance than running it with people only. But if you're only looking at direct cost, you've missed the game. The return comes from 24/7 availability, shorter wait times, +10 NPS points, lower agent churn, and calls you no longer lose at peak hours. It turns positive in 3–6 months if you measure it.
"I don't have to disclose it's AI." Yes, you do. The EU AI Act has required it since 2025, and research shows disclosure doesn't hurt containment. What does collapse is trust, when customers realise after the fact that you misled them. The "pretend it's a human" play kills slowly, but reliably.
What to do now
If you're just starting out, three things:
- Look at your call statistics. Top 10 call types. Which 60% can go to AI in the first round?
- Test 2–3 platforms (Vapi, Retell, or LiveKit if you want a brand-consistent voice). Two weeks.
- Launch with one use case, in one language. Pilot with 50 real calls, opt-in. Scale from there.
The "big-bang rollout" almost always fails. Pilot → measure → refine → scale rarely does.
If you want to go deeper: the full architecture, the 5 strategic decisions, the ROI model details, the industry benchmarks (restaurant, healthcare, banking, insurance), the 7 most common rollout mistakes, and the 4–6 month roadmap are all in the Voice AI agents — phone assistants in practice knowledge base piece.