A year ago we started deploying AI voice agents on restaurant reservation lines — inbound calls answered by an agent capable of taking a booking, answering menu questions, and escalating to a human when it made sense. Below is what the data now says.
The headline numbers
Across the five Italian restaurant venues where we've run this longest:
- ~72% of calls handled end-to-end by the agent — no human touch beyond the eventual arrival
- ~20% escalated to staff — either the guest asked for a person or the agent hit a topic outside scope
- ~8% dropped before completion — mostly wrong numbers, silent lines, and misroutes
The share the agent handles goes up quarter over quarter as the intent library widens and as staff learn what to route back. Six months in, most venues are past 80% and climbing.
Where the value actually sits
The headline win is not "we saved on labour". It's coverage. The five venues used to lose an estimated 18-24% of reservation demand outside staffed hours — evenings after 22:30, mornings before 09:30, the entire Monday when the venue is closed but bookings for the week get made. The agent recovers that demand. In three of the five venues, this alone justified the deployment.
Labour recovery is real but secondary. The Milanese restaurant automation case reclaimed about 25 hours a week across the group — meaningful, not transformational.
What surprised us
Multilingual outperforms staff on non-Italian calls
An Italian native operator on the phone will slow down for a foreign guest, sometimes switch to a lingua franca, sometimes lose the booking. The agent handles German, English, and French at native cadence. Non-Italian booking conversion on the agent line runs about 1.4× the pre-deployment baseline.
Special-request handling was the failure mode
Allergies, private-room requests, celebration set-ups — early deployments handled these clumsily. The fix wasn't a bigger model. It was a tighter escalation rule: any request containing a listed trigger word ("allergy", "private", "engagement", "birthday", specific dietary vocabulary) routes to staff immediately with full context. The agent stopped trying to be clever.
The recording matters
Every call is recorded, transcribed, and weekly-reviewed by the venue manager. Not to police the agent — to catch the intents the agent doesn't yet know. That review loop is what moves the handled-share up over time. Venues that skip the review plateau at around 65%.
Where it still misses
Three cases the current generation of voice agents doesn't handle well:
- Group bookings above ten covers — the calendar logic is fine, the payment/deposit conversation isn't
- Complaints or refund calls — the emotional bandwidth isn't there; escalate immediately
- Regulars who expect the manager to remember them — the agent doesn't (and shouldn't pretend to)
Each of these is a legitimate reason to route to a human. The deployment succeeds when the operators understand the agent as a first line, not a replacement.
"The best deployments look boring. They pick up the phone, take the booking, and end the call. The cases that need a human get a human. Nobody has to be impressed."
Where to start if you're a venue operator
The three-question audit before any deployment:
- How many calls per week does the venue receive, and what share go unanswered?
- What proportion of those calls result in a booking, and what average cover value?
- What are the top five reasons someone calls beyond making a reservation?
The answers determine whether the deployment saves labour, recovers demand, or both — and how to size the escalation rules from day one. The AI & Automation Systems engagement wraps the build; the Restaurants & Chains engagement wraps the operational side. For a specific venue, start with a 30-minute session.