Speculative · Solo project · 2026 · Product design
Acts first where undo is free. Asks first where it isn't.
A speculative case study: what if quick commerce platforms understood the situation and intent behind what you type? I designed an agentic AI assistant into the Blinkit app, from the first framing question to hi-fi screens, failure cases and ethics.
Hold to talk. Your words stay in the bottom sheet. Products always show in the feed. The whole design is built on this rule.
Role
Solo · product design, end to end
Timeline
June – Sep 2026
Type
Self-directed · speculative
AI pattern
Agentic · you talk, it acts
Scope
Composer · Replenisher · Substituter · checkout disclosure · provenance labels · reversibility rule · failure states
00
Overview
WHAT IT IS
A GenAI assistant that turns intent into a basket
You say what you need in one sentence. You can change every item before you pay. Delivery takes ten minutes, but ordering still takes four to six. This closes that gap.
WHAT I DESIGNED
Four pieces of AI, one system
Composer — turns one sentence into an editable cart.
Replenisher — remembers her habits and offers them, never adds on its own.
Substituter — handles items that run out — before payment in the cart, and after payment on the lock screen.
Disclosure — a label on every AI-added item: Inferred, Bought before, Sponsored.
WHAT’S INSIDE
4 design rules · the information architecture · 10 high-fidelity states · 3 rejected layouts · 3 failure modes · an ethics section on paid placement · the bets that would prove it wrong.
COMING UP
Part 1 — Framing & principles: the problem, the people, the rules.
Part 2 — Hi-fi frames: the rules on real screens.
Part 3 — Failure, ethics & reflection: what breaks, and what it costs.
01
Problem space
Delivery got faster. Ordering didn't.
Blinkit delivers in minutes. Placing the order often takes longer than that: browsing, searching, comparing, adding, checking out. That hurts most when someone is ordering in a hurry.
Now
One item at a time
1
Open the app
Repeat for every item
2
Search “milk”
3
Pick a brand, tap add
4
Check the cart
5
Pay
10 items = 10 searches
With the assistant
One request
1
“Breakfast for 4, we’re out of everything”
2
Assistant fills the basket
Bread, eggs, milk, butter + 6 more
3
Swap the bread, remove eggs
4
Pay
10 items = 1 request
People order in a hurry, with half their attention. So the unit of input shouldn’t be a product search. It should be a situation.
Nobody has asked for an AI to fill their basket. That’s my bet as a designer. Out-of-stock swaps are the opposite: people already complain about them, and they can be measured. So the project also covers that moment people already hate.
decisions · HIGHLIGHTS
Key decisions, each with the option that lost.
These are the seven choices that shaped the study: what lost, what I kept, what it cost, and where to read more.
#
Decision
Discarded → kept
What it cost
Where
01
Where does the AI live?
Chat panel · inline only · dock only
→ Dock for talk, feed for results
Two surfaces to learn instead of one
02
What decides when it may act?
A confidence threshold
→ What the action costs to undo
It must ask before anything it can’t undo, even when sure
03
One input or two?
Search on top, assistant below
→ One field, in the dock
Moves Blinkit’s most-used control and its main ad surface
04
Can brands buy the pick?
Paid picks with no label
→ Only if it fits the ask, labelled
Fewer paid slots; a visible tag invites removal
05
Who places the order?
The assistant checks out too
→ Paying is always the user’s tap
Never fully hands-off; one tap always remains
06
When does it listen?
Wake word, always listening
→ Only while the mic is held
Needs a free hand, even on the move
07
Who added what?
One mixed cart list
→ Manually added / Added for you
A longer cart, split in two
02
People · three attention states
Built for someone distracted, someone too tired to look, and someone on the move.
The work is built around three people. Each shows a different way today’s app lets someone down, just because of the state they’re in.
Attention state 01 · split
Sreya, 29 — Guests in twenty minutes
Thursday, 9:41 pm. Guests in twenty minutes. Nothing planned.
How Sreya orders
Show
Eyes
In glances
Hands
Free
Voice
Speaks, then taps to fix
The hard part
She changes her mind mid-order. Handling that is the main design problem.
What the assistant does
Hears the situation and builds the basket. Voice and touch share one screen — no mode switch.
Attention state 02 · depleted
Amith, 34 — unwell, in bed
Sunday, 11:20 am. Too tired to browse. Wants one delivery and as few taps as possible.
How Amith orders
Show
Eyes
As little as possible
Hands
Barely
Voice
Hold to talk
The hard part
A grid of choices is work he can't do. The screen has to get simpler.
What the assistant does
Rolls his needs into one order and makes the screen simpler. Every item says why it’s there.
Attention state 03 · in transit
Rohan, 31 — consultant, cab to home
9:20 pm. In a cab to Andheri West, home around 9:58. Dinner not sorted.
How Rohan orders
Show
Eyes
In bursts
Hands
One
Voice
Push-to-talk only
The hard part
The order would land seventeen minutes before he does. For him, fastest is the wrong answer.
What the assistant does
Times the delivery to when he arrives. Location only with his yes; ASAP stays one tap away.
One assistant, three sets of rules. How much it does on its own — and how it asks first — follows the state someone is in.
03
Framing · what the AI should do
One test for every AI idea: does it keep ordering fast?
Before designing any screen, I listed everything an AI could do inside Blinkit. Then I sorted the list with one question: does it save the user thinking time? If yes, it fits a ten-minute app. If it makes her stop and think, it works against the product, however clever it is.
Saves thinking time — fits Blinkit
Role
What it does
Note
Composer
Situation → filled basket
Scoped in this project
Substituter
Swaps items that run out after you order
In scope — acts when you can’t step in
Replenisher
Guesses when you’ll run out, reminds you
In scope — it offers, she decides
Translator
Dish or occasion → ingredients and quantities
Needs cooking sense
Explainer
Ingredients, allergens, origin
On tap only
Adds thinking time — doesn’t fit
Role
What it does
Note
Comparator
Brand versus brand
Quick-commerce apps removed this on purpose
Negotiator
Fits the basket to your budget
Makes her trade off items — slows the order
No interface at all
The AI that most improves Blinkit is invisible
Stock forecasting, faster picking, grouping deliveries, better product data, fraud checks. Together these help customers more than any voice button, but none of them has a screen. So I name them here and leave them out of the design.
Serves the platform first
Roles that serve Blinkit first
Ad ranking, dynamic pricing, notification timing. When the AI fills Sreya’s basket, it decides what she sees without her choosing.
04
Principles · the rules behind every screen
Four rules. Each builds on the one before.
Each rule decides something real on a screen, and each can be broken. That’s how I know they matter.
Rule 01
How freely the AI acts depends on how easy it is to undo
The obvious rule is: let the AI act when it’s sure. But a confident mistake is worse than a hesitant one. What matters is how hard it is to undo. Easy to undo: act freely. Hard to undo: ask first. Can’t be undone: get permission before it ever happens. Every choice about how much the AI does on its own comes back to this.
Rule 02
Commit, then correct
When a change is easy to undo, the assistant adds items first and tells you after, instead of suggesting and waiting. Asking for approval at every step kills the point — you end up doing the work anyway. The cart already has steppers (the − 2 + control), so a wrong basket is one tap from right. And adding to the cart isn’t placing an order: she still decides to pay.
Rule 03
Talk stays in the sheet. Results show in the feed.
The first version was a chat panel. I dropped it on purpose: a panel splits the app in two and makes you pick a mode. Instead, the conversation lives in the dock, a bottom sheet that is always there. Everything the assistant finds shows up in the normal product feed, as normal product cards. Questions go in the sheet. Results go in the feed. They never swap.
Rule 04
Each item shows where it came from
Step-by-step AI reasoning looks transparent, but nobody reads it, and it can be made up after the fact. So “Why is this in my cart?” is answered on the item itself, when you tap it, right when you doubt it. The chat log is only a record.
Discarded · offers she accepts
Kept · items that land
Decision · how items arrive
Offers she has to accept, or items added straight in?
The earlier idea used bundles: three cards, an ADD BUNDLE button on each, and an “Add all” bar. It might look convenient, but gives less control and choice.
The version I kept adds items straight to the cart as steppers and tells her after. The cost: it’s less polite, and she has to spot a wrong basket herself. That’s why every added row carries a provenance label (a tag saying where it came from) and can be removed in one tap.
The earlier idea also used its own colours. I moved to Blinkit’s real design system with one added colour.
05
Surface · the two-surface rule
A dock for talk, a feed for results.
Four places the AI can show up, each with strict rules on what goes there. Those rules stop the assistant from becoming a separate place you have to go.
How I got here · three layouts for the same job
Hi Neha, time for your weekly grocery shopping?
I have added your usual items, give it a look!
Show me the alternatives
Btw, Oranges are out of stock. shall I add moosambi ?
or take a look at the alternatives:
Hi Neha, time for your weekly grocery shopping?
I have added your usual items, give it a look!
“Show me the alternatives”
Btw, Oranges are out of stock. shall I add moosambi ?
or take a look at the alternatives:
Based on your location and travel,
scheduling delivery in
24 minutes
HOME - Neha, H12, IIT B
CHANGE
V1 · chat panel · discarded
A panel over the app
The shop stays underneath, blurred, and products show inside the chat. The cost: it feels like two apps in one. Every order starts with picking a mode, and the panel turns into a second, smaller shop.
9:41
▮▮▮
Delivery in
8 minutes
HOME · M G Road, Nilambur▾
Search
BRAND BAND
Assistant
Hi Neha — guests in 20 minutes.
Pulled a snack spread from what you usually
reorder. Swap anything.
Composed basket
See all
15%
OFF
8 MINS
ADD
8 MINS
ADD
Also — curd's been out since Tuesday.
ADD
10%
OFF
6 items
₹482
VIEW CART ›
Blinkit
Categories
Cart
V2 · inline only · discarded
Speech as a block in the feed
Nothing gets covered, and the assistant feels part of the shop. The cost: the conversation has no home. Old messages scroll away, and a follow-up has nowhere to go.
9:41
▮▮▮
Delivery in
8 minutes
HOME · M G Road, Nilambur▾
Search
BRAND BAND
Snacks & namkeen
See all
15%
OFF
8 MINS
ADD
8 MINS
ADD
8 MINS
ADD
10%
OFF
8 MINS
ADD
6 items · ₹482
VIEW CART ›
Assistant
Hi Neha — guests in 20 minutes.
Six items ready. Say "show me" or tap to review.
no paneer, they're vegan
Ask or type
Blinkit
Categories
Cart
V3 · dock only · discarded
Everything in a bottom sheet
A steady place to talk and fix things. The cost: about 40% of the screen is taken before she asks for anything, and the cart bar gets swallowed by the sheet.
I kept V2 and V3, split by what they hold. Talk goes in the dock, because conversation needs a home. Results go in the feed, because products belong in the shop. The cart bar sits above the sheet at every height, which fixes V3’s problem. These early frames still use the persona’s old name, Neha.
WHERE EACH SURFACE LIVES
Four places the AI shows up, and what each may hold.
Show the four surfaces
Dock
Bottom sheet — the conversation
What you asked, follow-up questions, the transcript. Temporary on purpose.
84 px idle — just a bar
≈270 px in dialogue
470 px expanded for the audit view
The cart bar floats above the sheet — the chat never hides the total
Inline
Product feed — the results
Everything the assistant finds shows up as normal Blinkit cards in the normal feed.
Composed rows, bundles, substitution cards
Loading placeholders, right in place
Never shifts what you’ve already scrolled past
Marked by a label on the normal card
Push
Outside the app — the edge
The only place the assistant reaches someone who isn’t looking at the app.
Swap choices you can answer from the lock screen
No answer means skip and refund
Used once per order at most
No interface
Doesn't require screen
Faster in-store picking and learning from rejected swaps matter to users, but have no screen. I name them here and leave them undrawn, because there’s nothing to draw.
Blinkit’s yellow and green already mean something: yellow is the brand and helps you find your way, green means confirm. A third bright colour for AI would clash with both. So the assistant gets one quiet colour of its own: aubergine #3E3047, used only for the dock. Aubergine anywhere else would be a bug.
06
Architecture · what can change
An app that adapts still needs a fixed frame.
The information architecture (IA) answers one question: which parts of the app can change, and which never should? Blinkit’s current structure stays almost the same. The redesign adds just one new branch and one overlay layer. Everything else stays where it is — only how each area is used changes.
Principle 1
Fixed edges, flexible middle
Navigation, search, the cart bar and the dock stay fixed on every screen. Only the feed between them changes. So your muscle memory still works, whatever the AI does.
Principle 2
AI-built screens are states of Home
A feed built by the assistant has no tab, no link and no back-button entry of its own. Closing it brings back the normal grid, so nobody gets stuck somewhere that no longer exists.
Principle 3
Categories stay as a second way in
The assistant searches the categories so you don't have to. Every aisle is still there, still browsable. It's just no longer the only way in.
THE FULL MAP
One new branch, one overlay. The rest stays put.
Show the map
App shell
Bottom nav
home · categories · bistro · cart
Header
location · ETA
Unified intent bar
product names and situations, one input
AI dock
84px idle · present on every screen
Home
Default grid
skeleton, error and reset state
Composed feed state
"guests in 20 min" → basket filled
Ambient suggestion row
why-chip attached · one per screen
Cart & checkout
Provenance labels
inferred · bought before · sponsored
Substitution consent
collected here, per order
Delivery timing
ASAP · arrive-with-you · scheduled
Payment
zero AI presence, structurally
Confirmation
zero cross-sell
Account
Orders & reorder
stable · feeds inference
AI & privacy centre
the one new branch
What the assistant uses
per-signal toggles
Substitution defaults
skip · swap · always ask
Reset personalisation
one tap · immediate
Stable · never changes
Composed · assistant fills
AI overlay · belongs to no branch
Discarded · search on top, assistant below
Kept · one input, in the dock
Decision · one input or two
Keep Blinkit’s search bar, or fold it into the assistant?
Two inputs meant two questions before every order: which one do I use, and what’s the difference? The user should never have to figure that out.
So “search” and “tell the assistant” became one field. Product names (like “Amul Taaza 1L”) get normal search results. Situations (“guests in 20 minutes”) get a feed built for them. The system tells them apart, so the user doesn’t have to.
The cost: this moves Blinkit’s most-used control from the top of the screen to the bottom, and search results are where most of its ads live. A real team would fight this hardest.
Speculative · Solo project · 2026 · Product design
Acts first where undo is free. Asks first where it isn't.
A speculative case study: what if quick commerce platforms understood the situation and intent behind what you type? I designed an agentic AI assistant into the Blinkit app, from the first framing question to hi-fi screens, failure cases and ethics.
Hold to talk. Your words stay in the bottom sheet. Products always show in the feed. The whole design is built on this rule.
Role
Solo · product design, end to end
Timeline
June – Sep 2026
Type
Self-directed · speculative
AI pattern
Agentic · you talk, it acts
Scope
Composer · Replenisher · Substituter · checkout disclosure · provenance labels · reversibility rule · failure states
00
Overview
WHAT IT IS
A GenAI assistant that turns intent into a basket
You say what you need in one sentence. You can change every item before you pay. Delivery takes ten minutes, but ordering still takes four to six. This closes that gap.
WHAT I DESIGNED
Four pieces of AI, one system
Composer — turns one sentence into an editable cart.
Replenisher — remembers her habits and offers them, never adds on its own.
Substituter — handles items that run out — before payment in the cart, and after payment on the lock screen.
Disclosure — a label on every AI-added item: Inferred, Bought before, Sponsored.
WHAT’S INSIDE
4 design rules · the information architecture · 10 high-fidelity states · 3 rejected layouts · 3 failure modes · an ethics section on paid placement · the bets that would prove it wrong.
COMING UP
Part 1 — Framing & principles: the problem, the people, the rules.
Part 2 — Hi-fi frames: the rules on real screens.
Part 3 — Failure, ethics & reflection: what breaks, and what it costs.
01
Problem space
Delivery got faster. Ordering didn't.
Blinkit delivers in minutes. Placing the order often takes longer than that: browsing, searching, comparing, adding, checking out. That hurts most when someone is ordering in a hurry.
Now
One item at a time
1
Open the app
Repeat for every item
2
Search “milk”
3
Pick a brand, tap add
4
Check the cart
5
Pay
10 items = 10 searches
With the assistant
One request
1
“Breakfast for 4, we’re out of everything”
2
Assistant fills the basket
Bread, eggs, milk, butter + 6 more
3
Swap the bread, remove eggs
4
Pay
10 items = 1 request
People order in a hurry, with half their attention. So the unit of input shouldn’t be a product search. It should be a situation.
Nobody has asked for an AI to fill their basket. That’s my bet as a designer. Out-of-stock swaps are the opposite: people already complain about them, and they can be measured. So the project also covers that moment people already hate.
decisions · HIGHLIGHTS
Key decisions, each with the option that lost.
These are the seven choices that shaped the study: what lost, what I kept, what it cost, and where to read more.
01
Where does the AI live?
Chat panel · inline only · dock only
→ Dock for talk, feed for results
Two surfaces to learn instead of one
02
What decides when it may act?
A confidence threshold
→ What the action costs to undo
It must ask before anything it can’t undo, even when sure
03
One input or two?
Search on top, assistant below
→ One field, in the dock
Moves Blinkit’s most-used control and its main ad surface
04
Can brands buy the pick?
Paid picks with no label
→ Only if it fits the ask, labelled
Fewer paid slots; a visible tag invites removal
05
Who places the order?
The assistant checks out too
→ Paying is always the user’s tap
Never fully hands-off; one tap always remains
06
When does it listen?
Wake word, always listening
→ Only while the mic is held
Needs a free hand, even on the move
07
Who added what?
One mixed cart list
→ Manually added / Added for you
A longer cart, split in two
02
People · three attention states
Built for someone distracted, someone too tired to look, and someone on the move.
The work is built around three people. Each shows a different way today’s app lets someone down, just because of the state they’re in.
Attention state 01 · split
Sreya, 29 — product manager, Powai
Thursday, 9:41 pm. Guests in twenty minutes. Nothing planned.
How Sreya orders
Eyes
In glances
Hands
Free
Voice
Speaks, then taps to fix
The hard part
She changes her mind mid-order. Handling that is the main design problem.
What the assistant does
Hears the situation and builds the basket. Voice and touch share one screen — no mode switch.
Attention state 02 · depleted
Amith, 34 — unwell, in bed
Sunday, 11:20 am. Too tired to browse. Wants one delivery and as few taps as possible.
How Amith orders
Eyes
As little as possible
Hands
Barely
Voice
Hold to talk
The hard part
A grid of choices is work he can't do. The screen has to get simpler.
What the assistant does
Rolls his needs into one order and makes the screen simpler. Every item says why it’s there.
Attention state 03 · in transit
Rohan, 31 — consultant, cab to Andheri
9:20 pm. In a cab to Andheri West, home around 9:58. Dinner not sorted.
How Rohan orders
Eyes
In bursts
Hands
One
Voice
Push-to-talk only
The hard part
The order would land seventeen minutes before he does. For him, fastest is the wrong answer.
What the assistant does
Times the delivery to when he arrives. Location only with his yes; ASAP stays one tap away.
One assistant, three sets of rules. How much it does on its own — and how it asks first — follows the state someone is in.
03
Framing · what the AI should do
One test for every AI idea: does it keep ordering fast?
Before designing any screen, I listed everything an AI could do inside Blinkit. Then I sorted the list with one question: does it save the user thinking time? If yes, it fits a ten-minute app. If it makes her stop and think, it works against the product, however clever it is.
Saves thinking time — fits Blinkit
Composer
Situation → filled basket
Scoped in this project
Substituter
Swaps items that run out after you order
In scope — acts when you can’t step in
Replenisher
Guesses when you’ll run out, reminds you
In scope — it offers, she decides
Translator
Dish or occasion → ingredients and quantities
Needs cooking sense
Explainer
Ingredients, allergens, origin
On tap only
Adds thinking time — doesn’t fit
Comparator
Brand versus brand
Quick-commerce apps removed this on purpose
Negotiator
Fits the basket to your budget
Makes her trade off items — slows the order
No interface at all
The AI that most improves Blinkit is invisible
Stock forecasting, faster picking, grouping deliveries, better product data, fraud checks. Together these help customers more than any voice button, but none of them has a screen. So I name them here and leave them out of the design.
Serves the platform first
Roles that serve Blinkit first
Ad ranking, dynamic pricing, notification timing. When the AI fills Sreya’s basket, it decides what she sees without her choosing.
04
Principles · the rules behind every screen
Four rules. Each builds on the one before.
Each rule decides something real on a screen, and each can be broken. That’s how I know they matter.
Rule 01
How freely the AI acts depends on how easy it is to undo
The obvious rule is: let the AI act when it’s sure. But a confident mistake is worse than a hesitant one. What matters is how hard it is to undo. Easy to undo: act freely. Hard to undo: ask first. Can’t be undone: get permission before it ever happens. Every choice about how much the AI does on its own comes back to this.
Rule 02
Commit, then correct
When a change is easy to undo, the assistant adds items first and tells you after, instead of suggesting and waiting. Asking for approval at every step kills the point — you end up doing the work anyway. The cart already has steppers (the − 2 + control), so a wrong basket is one tap from right. And adding to the cart isn’t placing an order: she still decides to pay.
Rule 03
Talk stays in the sheet. Results show in the feed.
The first version was a chat panel. I dropped it on purpose: a panel splits the app in two and makes you pick a mode. Instead, the conversation lives in the dock, a bottom sheet that is always there. Everything the assistant finds shows up in the normal product feed, as normal product cards. Questions go in the sheet. Results go in the feed. They never swap.
Rule 04
Each item shows where it came from
Step-by-step AI reasoning looks transparent, but nobody reads it, and it can be made up after the fact. So “Why is this in my cart?” is answered on the item itself, when you tap it, right when you doubt it. The chat log is only a record.
Discarded · offers she accepts
Kept · items that land
Decision · how items arrive
Offers she has to accept, or items added straight in?
The earlier idea used bundles: three cards, an ADD BUNDLE button on each, and an “Add all” bar. It might look convenient, but gives less control and choice.
The version I kept adds items straight to the cart as steppers and tells her after. The cost: it’s less polite, and she has to spot a wrong basket herself. That’s why every added row carries a provenance label (a tag saying where it came from) and can be removed in one tap.
The earlier idea also used its own colours. I moved to Blinkit’s real design system with one added colour.
05
Surface · the two-surface rule
A dock for talk, a feed for results.
Four places the AI can show up, each with strict rules on what goes there. Those rules stop the assistant from becoming a separate place you have to go.
How I got here · three layouts for the same job
Hi Neha, time for your weekly grocery shopping?
I have added your usual items, give it a look!
Show me the alternatives
Btw, Oranges are out of stock. shall I add moosambi ?
or take a look at the alternatives:
Hi Neha, time for your weekly grocery shopping?
I have added your usual items, give it a look!
“Show me the alternatives”
Btw, Oranges are out of stock. shall I add moosambi ?
or take a look at the alternatives:
Based on your location and travel,
scheduling delivery in
24 minutes
HOME - Neha, H12, IIT B
CHANGE
V1 · chat panel · discarded
A panel over the app
The shop stays underneath, blurred, and products show inside the chat. The cost: it feels like two apps in one. Every order starts with picking a mode, and the panel turns into a second, smaller shop.
9:41
▮▮▮
Delivery in
8 minutes
HOME · M G Road, Nilambur▾
Search
BRAND BAND
Assistant
Hi Neha — guests in 20 minutes.
Pulled a snack spread from what you usually
reorder. Swap anything.
Composed basket
See all
15%
OFF
8 MINS
ADD
8 MINS
ADD
Also — curd's been out since Tuesday.
ADD
10%
OFF
6 items
₹482
VIEW CART ›
Blinkit
Categories
Cart
V2 · inline only · discarded
Speech as a block in the feed
Nothing gets covered, and the assistant feels part of the shop. The cost: the conversation has no home. Old messages scroll away, and a follow-up has nowhere to go.
9:41
▮▮▮
Delivery in
8 minutes
HOME · M G Road, Nilambur▾
Search
BRAND BAND
Snacks & namkeen
See all
15%
OFF
8 MINS
ADD
8 MINS
ADD
8 MINS
ADD
10%
OFF
8 MINS
ADD
6 items · ₹482
VIEW CART ›
Assistant
Hi Neha — guests in 20 minutes.
Six items ready. Say "show me" or tap to review.
no paneer, they're vegan
Ask or type
Blinkit
Categories
Cart
V3 · dock only · discarded
Everything in a bottom sheet
A steady place to talk and fix things. The cost: about 40% of the screen is taken before she asks for anything, and the cart bar gets swallowed by the sheet.
I kept V2 and V3, split by what they hold. Talk goes in the dock, because conversation needs a home. Results go in the feed, because products belong in the shop. The cart bar sits above the sheet at every height, which fixes V3’s problem. These early frames still use the persona’s old name, Neha.
WHERE EACH SURFACE LIVES
Four places the AI shows up, and what each may hold.
Show
Dock
Bottom sheet — the conversation
What you asked, follow-up questions, the transcript. Temporary on purpose.
84 px idle — just a bar
≈270 px in dialogue
470 px expanded for the audit view
The cart bar floats above the sheet — the chat never hides the total
Inline
Product feed — the results
Everything the assistant finds shows up as normal Blinkit cards in the normal feed.
Composed rows, bundles, substitution cards
Loading placeholders, right in place
Never shifts what you’ve already scrolled past
Marked by a label on the normal card
Push
Outside the app — the edge
The only place the assistant reaches someone who isn’t looking at the app.
Swap choices you can answer from the lock screen
No answer means skip and refund
Used once per order at most
No interface
Doesn't require screen
Faster in-store picking and learning from rejected swaps matter to users, but have no screen. I name them here and leave them undrawn, because there’s nothing to draw.
Blinkit’s yellow and green already mean something: yellow is the brand and helps you find your way, green means confirm. A third bright colour for AI would clash with both. So the assistant gets one quiet colour of its own: aubergine #3E3047, used only for the dock. Aubergine anywhere else would be a bug.
06
Architecture · what can change
An app that adapts still needs a fixed frame.
The information architecture (IA) answers one question: which parts of the app can change, and which never should? Blinkit’s current structure stays almost the same. The redesign adds just one new branch and one overlay layer. Everything else stays where it is — only how each area is used changes.
THE THREE PRINCIPLES
What the AI can move, and what always stays put.
Show
Principle 1
Fixed edges, flexible middle
Navigation, search, the cart bar and the dock stay fixed on every screen. Only the feed between them changes. So your muscle memory still works, whatever the AI does.
Principle 2
AI-built screens are states of Home
A feed built by the assistant has no tab, no link and no back-button entry of its own. Closing it brings back the normal grid, so nobody gets stuck somewhere that no longer exists.
Principle 3
Categories stay as a second way in
The assistant searches the categories so you don't have to. Every aisle is still there, still browsable. It's just no longer the only way in.
THE FULL MAP
One new branch, one overlay. The rest stays put.
Show
App shell
Bottom nav
home · categories · bistro · cart
Header
location · ETA
Unified intent bar
product names and situations, one input
AI dock
84px idle · present on every screen
Home
Default grid
skeleton, error and reset state
Composed feed state
"guests in 20 min" → basket filled
Ambient suggestion row
why-chip attached · one per screen
Cart & checkout
Provenance labels
inferred · bought before · sponsored
Substitution consent
collected here, per order
Delivery timing
ASAP · arrive-with-you · scheduled
Payment
zero AI presence, structurally
Confirmation
zero cross-sell
Account
Orders & reorder
stable · feeds inference
AI & privacy centre
the one new branch
What the assistant uses
per-signal toggles
Substitution defaults
skip · swap · always ask
Reset personalisation
one tap · immediate
Stable · never changes
Composed · assistant fills
AI overlay · belongs to no branch
9:41
▮▮▮
Delivery in
8 minutes
HOME · Kadappakada, Kollam ▾
Search
BRAND BAND
Snacks & namkeen
See all
15%
OFF
8 MINS
ADD
8 MINS
ADD
8 MINS
ADD
10%
OFF
8 MINS
ADD
6 items · ₹482
VIEW CART ›
Assistant
Hi Neha — guests in 20 minutes.
Six items ready. Say "show me" or tap to review.
no paneer, they're vegan
Ask or type
Blinkit
Categories
Cart
Discarded · search on top, assistant below
Kept · one input, in the dock
Decision · one input or two
Keep Blinkit’s search bar, or fold it into the assistant?
Two inputs meant two questions before every order: which one do I use, and what’s the difference? The user should never have to figure that out.
So “search” and “tell the assistant” became one field. Product names (like “Amul Taaza 1L”) get normal search results. Situations (“guests in 20 minutes”) get a feed built for them. The system tells them apart, so the user doesn’t have to.
The cost: this moves Blinkit’s most-used control from the top of the screen to the bottom, and search results are where most of its ads live. A real team would fight this hardest.