What a restaurant chatbot should actually do
Answer from your menu and your policies, cite where the answer came from, and say so when it does not know. That last part matters more in a restaurant than anywhere else, because the question it will be asked most carefully is about allergens, and a confident wrong answer about peanuts is not a customer service problem, it is a safety problem.
Beyond answering, the useful jobs are taking the order, taking a booking where you offer them, and doing both in the language the visitor wrote in. Everything else is decoration.
Step one: gather what you already have
You are not writing content. You are collecting it. Most restaurants have all of this within reach:
- The menu, in whatever form it exists: a PDF, a price list, a photo of the board. Include the allergen and dietary columns if you print them.
- Opening hours, including the holiday exceptions, because the holiday question is the one that arrives at eleven at night.
- The booking policy: party sizes, deposits, how late you hold a table, the cancellation window.
- The practical page: parking, high chairs, step-free access, dogs on the patio.
Step two: upload it, and let the menu be read
Upload the files, or point the agent at your website if the information is already on it. The menu gets special treatment: upload it as a file and the dishes, prices, descriptions, sizes and extras are read out of it into an editable list. You correct anything that was misread, mark a dish as off for the day when it is, and the ordering side is set up from that list rather than typed in by hand. Nobody sets up online ordering by retyping ninety items, which is the real reason most restaurants never finish setting one up.
Give it a few minutes to read everything, then ask it something only your menu could answer. If the answer carries a citation back to the menu, it is working.
Step three: put it on the site, or skip the site
On a website it is one script tag, pasted before the closing body tag. Whoever built the site can do it in a minute, and most site builders have a place to add a script without touching code.
If you have an Instagram page and no website, skip this step. Every account comes with a hosted page that opens the same agent full screen, and a QR code that points at it. Put the link in your bio and the code on the menu, the window and the receipt.
Step four: ask it the hard questions
Before anyone else does. Ask about an allergen your menu does not mention and check that it says it cannot confirm rather than guessing. Ask in Spanish and check that it answers in Spanish from the English menu. Ask whether you can seat a party of eight on Saturday and check that it quotes your policy. Ask what is in the set lunch. Ten minutes of this tells you what to upload next, and it is the same test your customers will run without telling you.
Keeping it honest
It answers from your menu, not from your kitchen. When a dish changes and the uploaded menu does not, it will confidently give the old answer, because the old answer is what you gave it. Upload the new menu when it changes, and never present the chat as a substitute for asking staff about a severe allergy. A good agent declines the question it cannot answer; it still cannot answer a question whose answer you never wrote down.
What to turn on next
Once the answers are right, the order is the obvious next step: the chat takes it against the menu it read, totals it, and the customer pays by card or at the counter. Orders arrive in your dashboard and, if you turn it on, on a kitchen display on a tablet by the pass. Promotions such as buy one get one, codes and happy-hour windows run from the same menu. And when the kitchen is slammed, one switch pauses online orders and tells the customer so, without losing their basket.