A Product Design Case Study
ZEPTO — AI INTENT SEARCH
From item lookup to task completion.
North Star
“The best search doesn’t return items. It finishes the task.”
Product Type
AI-Powered Intent Search & Auto-Pack System, Quick Commerce
Context
UX Trilogy Design Challenge, with Rajat Patel · Nov 2025 · 1st Place
Approach
Intent-First · Mode-Based Control · Trust-Layered Personalization
Placement
1st Place
◆ CHAPTER ONE — THE TRIGGER
Ten Minutes Before Guests Arrive
A Bangalore apartment. Guests in 10 minutes. Shelves empty. Searching “biryani” returns rice — but the job was never rice. The job was a complete plan, a complete pack, and no forgotten tea powder halfway through the celebration.
The gap: what a person means vs. what the search box understands. That gap is the product.
CHAPTER TWO — THE PROBLEM
What We Were Actually Solving
Real problem
Search treats every query as a keyword to match, not an intention to fulfil. The work of turning a craving into a complete order falls on the user, every time.
Core insight: users don’t search for items — they search for intentions.
The more human the query, the worse it performs
Precision query: “Amul Toned Milk 1L” → works fine. Human query: “milk,” mid-cooking, brand undecided → falls through.
Four cohorts, one search bar, four jobs
First-time buyers need confidence. Early-life customers need safety. Active users need speed. Power users need personalization. Search fails all four when it returns the same ranked list.
Eight repeating pain points
Multiple searches, forgotten essentials, no scenario recognition, ignored dietary filters, low personalization, weak substitutes, no explainability, and no sense of completion.
CHAPTER THREE — PRODUCT VISION
What It Is / Is Not
AI Intent Search is
A task-completion engine, a scenario-aware prediction layer, a mode-based control system, a transparent pack-builder, and a personalization layer the user can tune.
It is not
A faster keyword search, a recommendation carousel, a silent smart black box, an opaque auto-cart, or a personalization layer that tunes itself silently.
The Six-Stage Intent Loop
Query → Intent Clarifier → Pack → Personalization → Add All → Learning Loop. Listens, clarifies, predicts, adapts — each stage catches the failure mode of the one before it.
CHAPTER FIVE — SCREEN DECISIONS
Control Surfaces Before Automation
Mode Selection
Smart, Budget, Speed, Voice. The user states the priority before the system optimizes — not the reverse.
Intent Chip
“Planning to make paneer butter masala?” appears live, is tappable, dismissible, and commits nothing.
Pack Preview
Top 2–3 items first, then “view all.” This is the trust surface where one search becomes one reviewable solution.
Clarifier Modal
Fires only above ambiguity threshold. Always-on feels incapable; never-on feels reckless.
Three Shelves
Boosters, Your Usuals, Expiry Replacements stay separate so commercial layers don’t contaminate trust layers.
Voice Mode
The same six-stage pipeline, hands-free. An alternate front door, not a separate feature.
CHAPTER SIX — LEARNING LOOP & BUSINESS CASE
Control → Trust → Adoption → AOV Growth
Every add, remove, and swap feeds the next prediction — a visible improvement, not a backend tweak. Scenario-based tasks grow convenience and revenue together: Zepto stops competing on fastest lookup and starts competing on most complete task finish.
Micro-rituals: Breakfast Hero · Health Guardian · Clean House Champion · Snack Ninja — narrow badges tied to one real repeated behavior each.
01
Instead of one optimized mode with hidden trade-offs, I made the user set the mode before search. One extra tap buys control over what the AI optimizes for.
02
Autocomplete suggests more text; the chip suggests what you’re trying to do. A wrong guess becomes visible early, when correction is cheap.
03
Clarifier fires only past a defined ambiguity threshold. Questions feel earned, not like a system admitting it can’t understand.
04
I didn’t load the full recommended cart immediately. Top 2–3 items create a cheap checkpoint before commitment.
05
I kept Boosters, Usuals, and Expiry Replacements separate. A unified ‘recommended’ shelf would make everything feel like a sales tactic.
AI’s job is to reduce effort, never to reduce visibility.
Where AI suggested full auto-add, a unified recommended shelf, or frequent clarification, I chose pack preview, separate shelves, and thresholded clarifying. Visibility won whenever it conflicted with speed.
The unresolved tension
Personalization depends on order history. First-Time Buyers have the least history — and need the most help. What should a stranger’s first pack default to?
◆
End of Case Study
