How a London restaurant group handled 18,000 calls a month with an AI voice agent — and the one decision that ended it

An AI voice agent answers the calls your team can't — and for one London restaurant group, that meant handling 18,000 calls a month across five restaurants, 24/7. Pachamama Group ran exactly this setup for about three months: an AI agent on ElevenLabs picked up overflow calls, sorted out what each caller wanted, and passed a clean call card to the contact centre. It worked at scale. Then a single data-access decision ended it — and that lesson is now the first thing we check before any contract.

What does an AI voice agent actually do on a restaurant phone line?

Picture a Friday at 9:40 PM. A guest is trying to book a group banquet for fourteen people — a table worth somewhere between €2,000 and €5,000 once you add wine and service. The line rings out. No one picks up, because the floor is full and the host is seating a queue. The guest calls the restaurant down the street instead. That booking is gone, and you will never see it in your reports, because a missed call leaves no trace.

This is the quiet leak in almost every restaurant. At peak hours, 30–40% of calls go unanswered (QSR Magazine 2025: 32% during the 5–8pm window; Hostie AI 2024–25 study of 500,000+ calls: 36% missed at peak), and the most expensive losses happen exactly when no human can get to the phone: the evening rush and the after-hours window. You can't fix this by hiring — you'd be paying someone to sit idle most of the day and still drown at 9 PM on a Friday.

An AI voice agent fills that gap. It answers when your people can't, understands what the caller wants, and either handles it or routes it to a human with the context already attached. (For the full how-it-works mechanics — the audio pipeline, intent handling, the call-card flow — see the cornerstone breakdown: "How a voice AI agent handles inbound calls — the Pachamama case" (in Russian).) Here we'll stay with the restaurant story, told honestly: what worked, what ended it, and what it means for you.

The Pachamama case: 18,000 calls a month, told straight

Pachamama Group runs a set of well-known London restaurants. The problem was the classic one at scale — too many calls, not enough people to answer them all, and the worst misses landing in the highest-value hours.

The setup was deliberately simple. An AI voice agent built on ElevenLabs answered the overflow — the calls the team couldn't reach — and orchestration ran through n8n. For each call, the agent did four things:

• Answered 24/7, including the after-hours window where bookings were otherwise lost outright.
• Classified intent — was this a new booking, a cancellation, a lost-item enquiry, a supplier, or a question about the menu or location?
• Transferred where it made sense, for example handing supplier calls through to the right line.
• Sent a structured call card to a Telegram group for the contact centre: the caller's number and exactly what they wanted, so a human could call back without starting cold.

Across five restaurants, this handled roughly 18,000 calls a month, running for about three months. The agent never replaced the contact centre — it caught what would otherwise have rung out, and turned each caught call into a clean, actionable card. The humans stayed in the loop on everything that needed judgement.

So why did it end? The OpenTable decision

Here's the part most case studies would quietly omit. The agent didn't wind down because the technology failed. It wound down because of a data-access decision outside the restaurant's control: OpenTable closed API access to booking data in London.

That single change broke the loop. Without live access to the reservation system, the agent lost real-time visibility into bookings. It could no longer reliably check a reservation or process a cancellation against the live book. The consequence was concrete and painful: diners who tried to cancel through the agent could still be charged no-show fees, because the cancellation never reached the system of record. A voice agent that can hear a cancellation but can't write it back to the booking platform isn't a convenience — it's a liability.

So we wound it down. That's the honest end of the story.

The lesson we now bake into every contract: a pre-sale data-access audit

The Pachamama experience changed how we sell. Before any contract, we run a pre-sale data-access audit of your stack — your POS, PMS, CRM, and booking API. The single question we answer first: can the agent read and write the data it needs, reliably, through a stable interface?

If the source is closed — if your booking platform won't expose an API the agent can depend on — we tell you that before the contract, and we don't start. It's a strange thing for a vendor to say out loud, but it's the only responsible position. A voice agent is only as good as its access to your source of truth. We'd rather lose the sale than sell you a system that charges your guests no-show fees on bookings it couldn't cancel.

What this means for your restaurant today

If you're evaluating a voice agent now, here's the practical frame.

The pain is real and measurable. With 30–40% of calls missed at peak (QSR Magazine 2025: 32% during the 5–8pm window; Hostie AI 2024–25 study of 500,000+ calls: 36% missed at peak), and the evening and after-hours bookings being the ones that actually cost you revenue, the question isn't whether you're losing bookings — it's how many.

There's a volume threshold. We don't take projects under 500 calls a month. Below that, the economics don't justify the setup, and you'd be better served by simpler fixes. The honest fit is a restaurant or group with real call pressure.

The numbers. Setup starts from €3,000; managed service from €800/month plus a per-minute markup. Go-live is 14 days for a standard restaurant deployment (28 days for enterprise).

Latency is no longer the question. In 2026, sub-500ms speech-to-speech is normal — ElevenLabs ConvAI v2 runs around 400ms. That's table stakes, not a selling point. The agent will sound responsive; what matters is whether it's wired correctly into your booking system and whether it routes the hard calls to a human cleanly. That's where the Pachamama lesson lives.

Where a voice agent doesn't help — and where a human takes over

Be clear-eyed about the limits. A voice agent is excellent at high-volume, well-defined calls: bookings, cancellations, opening hours, location, routing suppliers. It is not the right tool for a distressed guest, a complaint that needs empathy and authority, a complex group event with custom catering and negotiation, or any edge case that falls outside what it was scoped to handle.

The correct design isn't "AI instead of people." It's AI on the front line for volume, with a clean human fallback for everything that needs a human. In the Pachamama setup, the agent's whole job was to hand humans a better starting point — the caller's number and intent already captured — not to be the last word. If a vendor tells you the agent handles everything, walk away.

One compliance line you can't skip

From 2 August 2026, the EU AI Act, Article 50 requires that callers are told they're speaking with an AI at first contact. We deliver this as a hard-coded, deterministic opening line — not something the model improvises — so the disclosure is consistent on every call. Transparency breaches under Article 50 carry penalties of up to €15M or 3% of turnover. For the full compliance breakdown, see our dedicated guide: "EU AI Act compliance for voice agents" (in Spanish).

See it on your own line

The fastest way to judge a voice agent is to be the caller. Enter your restaurant's URL and the AI agent will call you back in 3 minutes — you'll hear exactly how it handles a booking, a cancellation, and a question, on your own restaurant's context.

Prefer to call in yourself? Reach the live demo line directly: +44 [TODO: data-point — UK demo number] (UK) or +34 [TODO: data-point — ES demo number] (Spain).

Expect a call from your AI agent within 3 minutes.

<!-- asset EN-1, synthetic personas not used as testimonials, byline=Vladimir Nagin, keyword volumes pending Рита validation -->