AI chatbot for restaurants: what your QR code menu still can’t do

Plenty of restaurants digitized their menu with a QR code during the pandemic and never touched it again: a PDF photo of the menu, nothing more. An AI chatbot for restaurants changes what that same QR code can do — instead of just displaying the menu, it answers the customer’s questions (allergens, side dishes, recommendations) straight from their phone, without waiting on a server who is already busy. The nuance matters more than it looks: recent research shows a QR code on its own can actually hurt customer loyalty, while a conversational assistant layered on top of it does the opposite.

A QR code on its own can cost you loyal customers

A study by Ganga S. Urumutta Hewage, L. Boman and S. Lefebvre, published in 2024 in the Journal of Hospitality and Tourism Management based on 325 responses from American consumers, found that a QR-code-only menu reduces loyalty compared with a paper menu — because it removes the human interaction some customers are looking for, an effect that even weighs on the tip left at the end of the meal. A QR code is not a neutral choice: badly used, it is a channel that can cost you regulars, not just save you on printing.

What actually makes the difference: interactivity, not the QR code itself

A second study, published in 2025 on ScienceDirect by Dewan Mehrab Ashrafi, Ali Iskender, Muhittin Cavusoglu and Shafayat Hossain Chowdhury, supplies the nuance the first one was missing: the problem is not the QR code as a medium, but the lack of interactivity and information quality — two factors that, on their own, clearly raise satisfaction and intent to return. A customer who can ask a real question (“does this dish contain nuts?”, “what pairs well with this wine?”) and get an immediate answer gets back exactly the interaction the QR code alone had taken away. That is the gap an AI chatbot closes on the same menu — close to what we built for a restaurant, in our project catalog: a QR-code ordering prototype where the mobile-readable menu answers questions about the dishes in addition to taking the order.

Before you deploy: what to check

• →Ground every allergen answer in your real, verified product sheet, never in the model’s general memory: France has required the display of the 14 regulated allergens for any non-prepackaged food since decree n° 2015-447 of April 17, 2015, under penalty of a fine (€1,500 for an individual, €7,500 for a company) — on this topic, an invented AI answer is not a minor detail, it is a health and regulatory risk. We covered elsewhere on this blog why an AI agent’s hallucination is not a one-off bug but a structural property to design around, and what actually reduces that risk in production.
• →Separate what belongs to the chatbot (answering questions about the menu) from what belongs to an agent (sending the order to the kitchen, updating the till): we explain elsewhere on this blog why that distinction changes the budget and the integration work required, not just the vocabulary.
• →Keep a one-tap handoff to the floor staff for anything outside the menu — a group booking, an event, a complaint: a chatbot that answers off-topic closes the conversation faster than it monetizes it.
• →Test the menu at peak hours before the final launch, with the real questions your customers ask — not just the ones you imagine: that is where the interactivity measured by the two studies above actually plays out, not in the demo.