AI Smart Assist for Nutrition-Aware Food Ordering on Zomato

✳The problem they went after

New and light Zomato users over-order or build nutritionally lopsided meals for three connected reasons: portion sizes are unclear, there is no nutrition guidance in context, and nothing establishes whether the order is for one person or a group. New users get an amplified version, since discovery leads with indulgent items, no history exists to personalise against, and unfamiliar meal sizing turns early orders into mistakes.

✳What they actually did

AI Smart Assist splits into two modes. Cold Start Mode applies population-level nutrition heuristics drawn from standard Indian meal composition, offering gentle guidance such as balanced thali suggestions without making false personalisation claims. After three completed orders the system graduates to Personalized Mode, using behavioural history for contextual recommendations, portion adequacy assessments and healthier in-restaurant substitutions. Cart intelligence watches item counts and asks users to clarify group size before guidance appears. Targets are a doubling of first-to-second order conversion and a 15 percent cut in new-user cart abandonment.

✳What you can take from it

Worth stealing

Cold start mode uses population heuristics, not fake personalisation
Personalised mode unlocks after three completed orders
Cart intelligence asks group size before giving guidance
Trust treated as the gate on how deep AI goes

Read this if

Designing an AI feature that has to work on day one with no data

The transferable bitAI behaviour has to earn its way into personalisation. Claiming to know a user you have no data on breaks trust immediately, so cold start mode deliberately says less and the personalised mode unlocks on a concrete threshold. Asking about group size before advising applies the same instinct at a much smaller moment.