Every restaurant guide tells you where to eat. None of them tell you what to order when you get there. That is the gap. So I built an AI voice agent that calls London restaurants and asks one question: if a first-timer walked in today, what is the one dish they absolutely have to order?
Then it listens. And what comes back is extraordinary.
First Order London is a new kind of food guide. No stars. No paragraphs. Just the dish name, recommended by the person who actually works there. 242 completed conversations. 60 restaurants called. 26 neighbourhoods. 31 cuisines. Here are the ten things I learned building it.
1. Independent restaurants answer. Famous ones don't.
Out of 242 conversations, roughly 85% of the big-name restaurants never got past their own phone systems. Dishoom, Hawksmoor, Gloria. IVR menus, voicemail inboxes, AI receptionists. Layer after layer of automation standing between a simple question and a simple answer.
The independent spots picked up. A human answered. They heard the question. They gave a dish. The entire interaction took less than a minute. No hold music. No "press 1 for reservations." Just a person, near a phone, happy to talk about their food.
This is a product insight, not just a data point. If you are building anything that relies on reaching restaurants by phone, target indie venues, not chains. The chains have optimised their phone systems to deflect exactly this kind of call. The independents have optimised for hospitality.
2. The dish name is the review.
That is Mangal 1 in Dalston. No star rating. No three-paragraph critique of the ambiance. Just a person, on the phone, with absolute conviction about the best thing on their menu. The way someone says "definitely" tells you more than any written review ever could.
This is the core thesis of the project. Every food guide in existence is a recommendation of WHERE to eat. Reviews, rankings, lists. But the moment you sit down and open the menu, you are on your own. First Order fills that gap. The dish name, from the person who makes it, is the entire review.
Caravan Exmouth Market told us their cornbread is their signature starter. Sun Kitchen in Hackney said the lamb ribs. Noble Rot said the smoked sole. Each answer arrived in under ten seconds. Each one is more useful than a thousand-word review.
3. One question gets better answers than two.
Version one of the agent asked for two or three dishes, then followed up with questions about why those dishes were special. The completion rate was poor. Staff got confused, or lost patience, or handed the phone to someone else. The conversation sprawled and the data suffered.
Version two was inspired by the Guinndex project, which called 3,000 Irish pubs to ask the price of a pint of Guinness. One question. One answer. Done. I stripped First Order down to the same model: one question, one dish, thirty seconds.
The results improved immediately. The constraint is the design decision. When you ask someone a single, focused question, they give you a single, focused answer. The discipline of asking less got us more.
4. Twenty pounds bought 60 restaurant recommendations.
The entire 60-restaurant dataset cost less than a meal at most of the restaurants in it. Twenty pounds. That covers the ElevenLabs API calls, the phone minutes, the compute. No journalists. No PR agencies. No forms to fill in. No editorial calendars. Just API calls and a well-crafted prompt.
This is what makes AI-native products different from traditional media. The cost of collecting primary source data has collapsed. A food magazine would spend thousands on a single restaurant feature. We spent twenty quid and got 60 genuine recommendations from real staff.
The economics change what is possible. You can call every restaurant in a neighbourhood. You can call them again next month. You can cover 31 cuisines across 26 neighbourhoods for the cost of a sandwich.
5. OpenStreetMap is the best free restaurant database nobody uses.
Before I found the right data source, I tried everything. Yelp blocked the scraper. OpenTable blocked the scraper. DesignMyNight blocked the scraper. Every commercial directory has walls around its data, and rightly so.
Then I discovered the Overpass API, OpenStreetMap's query interface. One query returned 2,993 London restaurants with phone numbers in 30 seconds. Free. Open. No rate limits to speak of. No terms of service that prohibit the use case. The full database had 4,060 venues.
Open data was not just cheaper than commercial alternatives. It was better. More comprehensive coverage of independent restaurants, which are exactly the venues that actually answer their phones. The open-source ecosystem solved the problem that every commercial API made impossible.
6. The agent needs a cover story.
Early calls had a high hangup rate. People heard a voice they did not recognise, asking a question that felt out of context, and assumed it was a scam. Without framing, the question sounds suspicious. Who calls a restaurant to ask what dish to order?
The fix was simple. The agent now introduces itself as putting together a free food guide for London. That single sentence of context reduced hangups significantly. It gives the person on the other end a reason to help. They understand the purpose, and most are genuinely happy to contribute.
Cafe Cecilia's staff initially thought the call was a scam. But they answered anyway and recommended the breaded butter pudding. Even scepticism did not stop them from wanting to share their best dish. The cover story just makes it easier for both sides.
7. 3pm is when restaurants actually talk to you.
Between lunch and dinner service, someone is near the phone with nothing urgent happening. That is the sweet spot. Call during lunch and nobody has time. Call during dinner and they are too busy to care. Call at 3pm and you get the person who is prepping, folding napkins, or just hanging about before the evening starts.
Timing matters more than the script. The best prompt in the world fails if it reaches someone mid-service with 40 covers to manage. The worst prompt succeeds if it reaches someone with five minutes to spare and genuine pride in what they serve.
This is not a technical insight. It is a human one. The product works because it respects the rhythms of how restaurants actually operate. Build around their schedule, not yours.
8. Audio of the call is more compelling than any written review.
The actual recordings are on the site. Go and listen. Hearing a Turkish grill owner in Dalston say "definitely the adana" with genuine pride in his voice hits differently than reading it on a screen. The hesitation before someone commits to their answer. The laugh when they realise the question is fun. The enthusiasm when they land on their favourite dish.
Written reviews strip out everything that makes a recommendation trustworthy. Tone, pace, conviction, spontaneity. The audio preserves all of it. You can hear whether someone is reciting a script or speaking from genuine love for what they cook.
The medium is the message. A text-based food guide is information. An audio food guide is a human connection. First Order is both, but the audio is what makes it real.
9. The venues you have never heard of give the best answers.
Sun Kitchen in Hackney. Not on any best-of list I have seen. Not in any magazine roundup. But when we called, the person who answered said this:
No TripAdvisor scraping surfaces this. No review aggregator captures it. The person who answered the phone IS the source. They are not a critic interpreting the restaurant. They are not a user posting after one visit. They are someone who works there, every day, and knows exactly what is best.
The most interesting recommendations in the dataset come from places with no media presence. The restaurants that do not need PR because their regulars already know. First Order reaches the places that traditional food media cannot or will not cover, and those places have the most to say.
10. The IVR is the review.
Hawksmoor's phone system is a 60-second runaround through hold music, menu options, and transfers. That is a choice. It tells you something about how they treat spontaneous callers. The phone experience is not separate from the restaurant experience. It is part of it.
A restaurant that picks up in three seconds and says "yeah, the adana" tells you something too. It tells you they are the kind of place where a person answers a phone and talks to you like a human being. The speed of the answer, the warmth of the voice, the confidence in the recommendation. All of it is signal.
The restaurants that made it through our process, the ones that answered, engaged, and gave a dish, are self-selecting for a particular kind of hospitality. The IVR filters them out. The human connection lets them in. The phone system is not a barrier to the data. It IS the data.
The numbers
242 completed conversations. 60 restaurants called. 10 genuine recommendations extracted so far. 26 neighbourhoods covered. 31 cuisines represented. Total cost: twenty pounds. And every recommendation came directly from someone who works at the restaurant, not from a critic, not from a user review, not from an algorithm.
This is Week 1. The database will grow. More neighbourhoods. More cuisines. More of those thirty-second moments where someone reveals their favourite dish to a voice on the phone. The interesting part has barely started.
See the live recommendations at First Order London.
