08
Failure types
Three failures. Three safeguards.
F1 · State 05
Interrupted
Keep completed additions labelled; stop the rest without undoing anything.
F2 · State 06
Wrong basket
One “Undo” restores the basket to its pre-assistant state.
F3 · Lock screen
Irreversible action
Show a timer, give both choices equal weight, and default safely if unanswered.
Thursday 17 September
9:54
BLINKIT
2 min ago
We swapped your tender coconut
It went out of stock and you didn’t answer in time, so we added Cocojal Tender Coconut Water. Same size, ₹4 less. Your order is on its way.
Rejected · Silence means swap
Thursday 17 September
9:52
BLINKIT
now
Tender coconut just went out of stock
Closest match: Cocojal Tender Coconut Water. Same size, ₹4 less, and you bought it in June. Everything else is packed.
Swap
Skip and refund
1:42
Couldn't answer? We skip it and refund ₹245.
Chosen · Silence means skip and refund
Decision · F3
No response? Skip and refund.
The item is unavailable, payment is complete, and the rider leaves in two minutes.
I chose skip-and-refund over an automatic swap. An unwanted replacement can’t be recalled once dispatched; a missing item can be reordered.
The trade-off: Lower order value when prompts go unanswered—and pressure from the business to change the default.
Open questions
How long should undo remain available? Can the refund default withstand business pressure? How should sponsored replacements be labelled?
09
The trade-off
An AI-filled basket is ad space too.
Search results are where Blinkit shows its ads. A basket filled for the user decides what she sees without her choosing anything. So the AI doesn’t remove ads from the app — it moves them into the basket, where there’s no label to show them.
The displacement
Fewer searches means less ad space
Every situation that turns into an AI-filled basket is a search that never happened, and a paid slot that never showed. Convenience and ad revenue pull in opposite directions, and I haven’t seen a concept deck admit it.
The pressure
Paid spots move into the basket
The business answer is obvious: let brands pay to be the assistant’s default pick. It’s the only way to win back the lost ad space, and it will get built whether it’s labelled or not.
The position taken
Label it on the item, or don’t do it
So there’s a third label: Sponsored, next to Inferred and Bought before. A paid item is allowed in an AI-filled basket only if it fits what the user asked for, is labelled on the item, and can be removed in one tap. An assistant that can be quietly bought isn’t an assistant.
10
THE ASSUMPTIONS
What has to be true,
and what would prove it wrong.
Bet 1 · composition
Fixing a basket is faster than building one
True if: fixing a few wrong items is faster than searching.
Bet 2 · substitution
Asking first reduces complaints
True if: checking before a swap means fewer support tickets than today.
Drop if: skip-and-refund loses more order value than the tickets it saves.
Bet 3 · disclosure
People accept labelled ads
True if: a Sponsored tag doesn't make people strip the item out.
Drop if: Sponsored items are removed about twice as often as others.
The known cost
Less ad space
Fewer searches means fewer paid slots. That gap is a real cost of this feature.
11
Reflections
What designing something that acts taught me, that designing a tool didn’t.
The reframe
Ease of undo decides how much the AI does
I started by asking how sure the assistant should be before acting. That was the wrong question, and it led to a worse design — a confidence score just excuses confident mistakes. Asking what it costs to undo led to an assistant that can do opposite things in two places, for a reason a user can actually understand.
The hardest trade-off
Acting without asking, without being pushy
Every confirmation step I removed made the product better and the assistant ruder. The fix was making “no” easy, labelling each item with where it came from, and asking permission earlier for the one thing that can’t be undone.
How I worked with AI
Built with Claude and Figma
I used Claude with Figma to move faster. The design decisions are mine.
What the tools did
Wrote the prototype code from my direction
Laid out the pages and screens in Figma
Drafted the copy and cut it down
Gathered references and generated options
What I decided
The reversibility rule the whole design follows
Words stay in the sheet, products show in the feed
If she doesn’t answer, skip and refund
A paid item needs a Sponsored label
END OF THE CASE STUDY
Part 3 of 3
PART 1
Framing & principles
00–06 · problem, rules
PART 2
Hi-fi frames
07 · the assistant, drawn
PART 3
Failure, ethics & reflection
08–11 · you are here
NEXT PROJECT
WeatherEx — UX audit & redesign
Turning a dense weather platform into a decision tool
An assistant is only as good as its undo.
Self-directed · speculative
saurabhkg.framer.website
ux.saurabhkg@gmail.com
08
Failure types
Three failures. Three safeguards.
F1 · State 05
Interrupted
Keep completed additions labelled; stop the rest without undoing anything.
F2 · State 06
Wrong basket
One “Undo” restores the basket to its pre-assistant state.
F3 · Lock screen
Irreversible action
Show a timer, give both choices equal weight, and default safely if unanswered.
Rejected · Silence means swap
Chosen · Silence means skip and refund
Decision · F3
No response? Skip and refund.
The item is unavailable, payment is complete, and the rider leaves in two minutes.
I chose skip-and-refund over an automatic swap. An unwanted replacement can’t be recalled once dispatched; a missing item can be reordered.
The trade-off: Lower order value when prompts go unanswered—and pressure from the business to change the default.
Open questions
How long should undo remain available? Can the refund default withstand business pressure? How should sponsored replacements be labelled?
09
The trade-off
An AI-filled basket is ad space too.
Search results are where Blinkit shows its ads. A basket filled for the user decides what she sees without her choosing anything. So the AI doesn’t remove ads from the app — it moves them into the basket, where there’s no label to show them.
The displacement
Fewer searches means less ad space
Every situation that turns into an AI-filled basket is a search that never happened, and a paid slot that never showed. Convenience and ad revenue pull in opposite directions, and I haven’t seen a concept deck admit it.
The pressure
Paid spots move into the basket
The business answer is obvious: let brands pay to be the assistant’s default pick. It’s the only way to win back the lost ad space, and it will get built whether it’s labelled or not.
The position taken
Label it on the item, or don’t do it
So there’s a third label: Sponsored, next to Inferred and Bought before. A paid item is allowed in an AI-filled basket only if it fits what the user asked for, is labelled on the item, and can be removed in one tap. An assistant that can be quietly bought isn’t an assistant.
10
THE ASSUMPTIONS
What has to be true,
and what would prove it wrong.
Bet 1 · composition
Fixing a basket is faster than building one
True if: fixing a few wrong items is faster than searching.
Bet 2 · substitution
Asking first reduces complaints
True if: checking before a swap means fewer support tickets than today.
Drop if: skip-and-refund loses more order value than the tickets it saves.
Bet 3 · disclosure
People accept labelled ads
True if: a Sponsored tag doesn't make people strip the item out.
Drop if: Sponsored items are removed about twice as often as others.
The known cost
Less ad space
Fewer searches means fewer paid slots. That gap is a real cost of this feature.
11
Reflections
What designing something that acts taught me, that designing a tool didn’t.
The reframe
Ease of undo decides how much the AI does
I started by asking how sure the assistant should be before acting. That was the wrong question, and it led to a worse design — a confidence score just excuses confident mistakes. Asking what it costs to undo led to an assistant that can do opposite things in two places, for a reason a user can actually understand.
The hardest trade-off
Acting without asking, without being pushy
Every confirmation step I removed made the product better and the assistant ruder. The fix was making “no” easy, labelling each item with where it came from, and asking permission earlier for the one thing that can’t be undone.
How I worked with AI
Built with Claude and Figma
I used Claude with Figma to move faster. The design decisions are mine.
What the tools did
Wrote the prototype code from my direction
Laid out the pages and screens in Figma
Drafted the copy and cut it down
Gathered references and generated options
What I decided
The reversibility rule the whole design follows
Words stay in the sheet, products show in the feed
If she doesn’t answer, skip and refund
A paid item needs a Sponsored label
END OF THE CASE STUDY
Part 3 of 3
An assistant is only as good as its undo.