DEV Community: Hostly

2,332 real phone calls later: five things I learned building an AI host for a restaurant

Hostly — Thu, 12 Nov 2026 13:38:01 +0000

I build Hostly (hostly.ie), an AI host that answers the phone and WhatsApp for restaurants in Ireland. Our first venue, a restaurant and lounge in Limerick city, went live on 15 April 2026. By 19 October the host had answered 2,332 calls and taken 249 bookings straight into the venue's OpenTable diary.

A demo of a voice agent is easy. A phone line that real guests ring on a Saturday evening is not. These are the five lessons that cost me the most to learn.

1. The diary is the only source of truth

A voice agent that sounds confident and is wrong is worse than a phone that rings out.

Our worst early fault was quiet. When the connection to the booking system was slow, the host sometimes told callers there was less availability than there really was. Nobody complained, because the caller simply did not book. We only found it by reading the calls against the diary.

What we do now:

The host may only state availability that came back from the booking system during that call.
Every availability check is logged and audited.
When a time is full, the host offers the nearest free time that day instead of a plain "no".

2. Never deploy while someone is on the phone

On the web, a deploy during traffic costs a page reload. On a phone line it hangs up on a guest. We dropped a handful of calls this way before we changed the rule.

Now a release only goes out when the count of active calls is zero. It is a small script and it is the most important one we have.

3. Decide what the AI must not decide

13% of calls (299 of 2,332) were handed to the restaurant's team: private hires, complaints, deliveries, anything involving judgement or money outside a normal booking.

The hand-over matters more than the conversation:

The caller's number and the reason are logged every time.
Every manager gets the alert on WhatsApp at once.
One of them marks it as handled, and the others see that.

Without the last step, five people assume one of the other four rang the guest back.

4. Names are the hardest part of a phone call

Speech recognition is good at sentences and poor at surnames, especially Irish ones on a noisy line. We have had a booking saved under a phrase the caller said in passing instead of their name.

What helped:

Checking the name against a list of things that are obviously not names before saving it.
Recognising a returning caller by their phone number, so the name is already known.
When a booking ends up without a usable name, sending the guest a short WhatsApp message asking for it, and writing the reply back to the booking.

5. Short calls are the goal

The median call was 55 seconds. Callers do not want a conversation. They want a table, a change, or an answer about parking.

Every extra sentence the host says is a chance for the caller to talk over it. Most of our improvements were removing words, not adding them. The host also says it is an AI at the start of every call. It sets the right expectation from the first second.

The numbers, for anyone building something similar

Measure	Result (15 Apr to 19 Oct 2026)
Calls answered	2,332
New bookings by phone	249
Changes and cancellations	166
Handed to a person	13%
Bookings taken 4pm to 8pm	52%
Median call length	55 seconds
Busiest day	83 calls

The full record, including what went wrong, is public: hostly.ie/case-studies/limerick-restaurant-lounge.

If you want to hear it, the demo line is 061 788 376 (Ireland). Nothing is booked on the demo line.

Daniel Silva is the founder of Hostly, an AI host for restaurants and service businesses in Ireland, built in Limerick. hostly.ie

We let an AI answer a Limerick restaurant's phone for six months. Here are the real numbers.

Hostly — Thu, 12 Nov 2026 13:37:16 +0000

Most of what is written about AI phone answering for restaurants is a vendor quoting a survey. This is different: it is the record of one phone line in Limerick, Ireland, taken from the database, including the parts that went wrong.

I am the founder of Hostly (hostly.ie), an AI host that answers the phone and WhatsApp for Irish restaurants. From 15 April to 19 October 2026 our host answered every call to a busy restaurant and lounge in Limerick city centre. The venue uses OpenTable for its diary and has a small front-of-house team.

The short version

Measure	Result
Calls answered	2,332
New bookings taken by phone	249
Covers booked	761
Changes and cancellations handled	166
Calls passed to a person	299 (13%)
Median call length	55 seconds
Busiest single day	83 calls (20 June)

About one call in nine ended in a new booking. The average booking was for three people.

The phone rings when nobody can answer it

This was the finding that mattered most to the owner.

52% of bookings were taken between 4pm and 8pm. That is evening service, when every member of staff is on the floor.
The peak hour for calls was 4pm to 5pm, followed by 5pm to 6pm.
7% of calls came outside 10am to 10pm, and 27 bookings were taken in those hours, when the building was empty.

Before the AI host, a call during service was answered if someone could reach the phone. Otherwise it rang out. Half of the bookings in this record came from exactly those hours.

Month by month

Month (2026)	Calls answered	Bookings by phone
April (from the 15th)	124	11
May	104	7
June	361	34
July	327	40
August	527	70
September	452	53
October (to the 19th)	437	34

Volume more than quadrupled between May and August. Two things happened: the venue moved the line over fully, and the summer season arrived.

What the AI did not handle

13% of calls (299) were passed to the team. These were private hires, complaints, deliveries and anything the host is told not to decide. Each one was logged with the caller's number and the reason. Since October, each one is also sent to every manager on WhatsApp, and one of them marks it as handled.

I think 13% is a healthy number. An AI host that never hands over is making decisions it should not make.

No-shows

Of the 249 bookings, 50 were later cancelled and 17 were no-shows. That is a no-show rate under 7% on bookings the host confirmed by SMS.

What went wrong

In the first months the host sometimes told callers there was less availability than there really was, when the connection to the diary was slow. We fixed the cause in October and now audit every availability check.
A handful of calls were dropped when callers rang during a software update. We now only update when no call is in progress.

Every one of these is visible in the venue's dashboard and was reviewed with the manager.

What I would tell another restaurant owner

Count your evening calls first. If your phone rings most between 4pm and 8pm, that is where the bookings are being lost.
Test on a live line, not a demo. Ask for a table at a time that is full. Ask to change a booking. Ask for a person.
Ask the vendor for their record, not their statistics. Calls answered, bookings taken, calls handed over, and what went wrong.

The full breakdown is on the Hostly site: Six months of an AI host on a Limerick restaurant's phone. You can also ring the demo line on 061 788 376 and talk to the host yourself. Nothing is booked on the demo line.

Daniel Silva is the founder of Hostly, an AI host for restaurants and service businesses in Ireland, built in Limerick. hostly.ie