What 16,000 Hotel Phone Calls Taught Us About AI Voice Agents

When you deploy an AI voice agent across a hotel group, you expect to learn about guest behavior. What you don't expect is to discover that your old phone system was hiding thousands of lost callers — and that removing it would make things look worse before they got better.

Over the past six months, we analyzed 15,910 incoming phone calls handled by Una across a European hotel network — 15 AI voice agents, 175 days, roughly 91 calls per day. The data revealed patterns that no one anticipated, and lessons that apply to any hotel considering AI for their phone lines.

Here's what we found.

The Vanishing Caller Problem

Every hotel tracks "empty calls" — calls where the guest hangs up before having a real conversation with the AI. For months, the network's empty call rate hovered steadily between 12% and 15%. Perfectly normal. Nothing to worry about.

Then, in mid-February, the hotel group made a seemingly minor change: they removed the IVR menu — the "Press 1 for Czech, Press 2 for English" system that greeted every caller. The idea was simple. IVR menus are annoying. Guests hate pressing buttons. Let the AI handle it directly.

Within days, the empty call rate jumped to 19.5%.

The initial reaction was alarm. Had the AI gotten worse? Was something broken?

Neither. The AI hadn't changed at all. The data had.

What Really Happened: The Hidden Filter

Here's the insight that changed how we think about voice AI deployment.

The old IVR system wasn't just routing calls — it was silently filtering them. Callers who didn't understand Czech, couldn't figure out the menu, or simply lost patience would hang up during the IVR stage — before ever reaching the AI agent. Those abandoned calls were invisible in the AI's statistics. They were ghosts.

When the IVR was removed, every caller went straight to the AI agent. The same people who used to drop off at the IVR stage now dropped off at the AI stage instead. They became visible in the data for the first time.

The IVR didn't prevent caller loss. It just hid it.

The real empty rate had always been higher than 15%. The hotel group just couldn't see it.

Spa Hotels vs. City Hotels: Two Different Worlds

The most striking pattern emerged when we split the data by property type.

Spa and wellness hotels maintained a steady empty call rate of around 11.7% across the entire period. Even after the IVR removal, their February numbers barely moved — just 12.7%. These properties serve a loyal, often local audience. Guests call specifically to book treatments. They're patient, motivated, and frequently speak the local language. The phone is their primary booking channel, and they're willing to work with an AI to get what they need.

City and business hotels told a completely different story. Their overall empty rate was already higher at 16.7%, but in February it spiked to 26.5%. One in four callers hung up without speaking.

The reason is straightforward: city hotels serve a much higher proportion of international travelers. Business guests, tourists, convention attendees — people who are less likely to speak Czech, less patient with technology, and more likely to simply open Booking.com on their phone the moment something feels off. For them, hearing a greeting in an unfamiliar language isn't a minor inconvenience. It's a stop signal: "They won't understand me here."

The Real Cost of a Dropped Call

When a guest calls a hotel and hangs up, they don't just disappear. They go somewhere else. They book on an OTA — and the hotel pays 15–20% commission. Or they choose a competitor whose phone was answered in their language.

At a rate of 91 calls per day across the network, even a 5 percentage point increase in empty calls translates to roughly 4–5 lost conversations daily. Over a month, that's 120–150 potential bookings that never happened — not because the AI failed, but because the first three seconds of the call sent the wrong signal.

The Fix: Not Back to IVR, Forward to Intelligent Greeting

The solution wasn't to bring back the IVR menu. Guests genuinely dislike pressing buttons, and the data confirmed that IVR created its own friction. Instead, the answer was a bilingual AI greeting — a short, natural prompt in both the local language and English:

"Dobrý den, welcome to [Hotel Name]. How can I help you?"

This does everything the IVR used to do — signal that English is available, give the caller a clear entry point — but without the mechanical "Press 1" experience. Instead of buttons, the AI listens to which language the guest responds in and continues in that language automatically.

The target: bring the empty rate below 12% across the entire network — including the previously invisible callers that the IVR was hiding.

Three Takeaways for Hotel Operators

1. Your current metrics might be lying to you. If you have an IVR or any pre-filter before your AI agent, you're probably not seeing the full picture. Callers lost at the IVR stage don't show up in AI statistics — but they're still lost revenue.

2. The first three seconds define everything. A phone greeting isn't just hospitality courtesy. It's a conversion mechanism. If an international guest hears a language they don't understand, they hang up — and they won't call back. They'll book elsewhere, usually on an OTA at your expense.

3. Different property types need different strategies. A wellness resort with 90% domestic guests can afford a single-language greeting. A city hotel near an airport or convention center cannot. Segment your approach the way you segment your marketing.

Looking Ahead

We're now running A/B tests on the bilingual greeting across the network and tracking the results daily. Early signals are promising — but we'll share the full data when we have it.

What this experience reinforced for us at Polydom is that deploying AI in hospitality isn't just a technology challenge. It's a data interpretation challenge. The numbers you see are shaped by the systems around them, and sometimes the most valuable insight is discovering what your data has been hiding from you.

If you're running AI voice agents at your property — or considering it — start by asking a simple question: What am I not seeing?