Every POS vendor in India in 2026 uses the word 'AI' somewhere on their website. Very few can explain what the model is, what it was trained on, or what specifically it predicts. This post is about what AI recommendations actually do in the ordering flow — and what they do not do.
What works today
The most reliable AI application in restaurant ordering is not personalization. It is aggregate behaviour inference — 'what do guests who order Item A also tend to order, at this venue, at this time of day?' This is collaborative filtering, a 25-year-old technique, and it works extremely well for restaurant menus because the menu is small (50–150 items) and the signal is dense (hundreds of transactions per week).
Upsell timing
The critical insight: upsell suggestions work when they appear after the main order is placed, not before or during. A suggestion shown at the wrong moment (before the guest has settled on their main) increases friction and reduces conversion. A suggestion shown after 'Your order is confirmed' — 'Guests who ordered Butter Chicken also love this Peshwari Naan' — converts at 18–24% in our data.
In Indostra's data across 40 venues, post-order add-on suggestions convert at 21% on average. Mid-order banners convert at 7% and increase time-to-order by 40 seconds.
Pairing logic
Beverage pairing suggestions are the highest-margin AI application in the ordering flow. 'A Lassi pairs well with your biryani' — shown after a biryani is added to cart — increases beverage attachment by 14–19% at venues that have implemented it. The margin on beverages is typically 65–75%, so the incremental revenue compounds quickly.
86-aware suggestions
A genuine AI feature that is underappreciated: if an item is 86'd (sold out), the suggestion engine should immediately reroute to the closest substitute and surface it in the UI — proactively, not after the guest hits an error. 'The Raan Gosht is sold out tonight — guests are loving the Dum Mutton instead.' This requires the suggestion model to understand dish similarity, not just co-occurrence.
Personalization limits
True personalization — 'because you ordered X last time, we suggest Y today' — requires the guest to be identified, which requires either login or phone number lookup. Most Indian restaurant guests do not log in. At venues with loyalty programs where 30%+ of guests are identified, personalization works. At most venues, it is aggregate collaborative filtering that does the work.
What is hype
- Dynamic pricing via AI: legally and culturally unacceptable in most Indian dining contexts. Not for now.
- AI-generated menu descriptions: the descriptions are fine but a good writer does it once and the AI does nothing annually.
- Demand forecasting via 'deep learning': simple day-of-week averages with a weather factor outperform most ML models on small restaurant datasets.
- Chatbot ordering: converts very poorly for food ordering in India. WhatsApp-native ordering is more effective.
Builds the KOT routing engine. Believes a kitchen is a real-time system, not a queue.