Allie
An AI shopping assistant you can steer, not just ask.
Store chatbots answer with walls of text, forget context, and never ask what matters to you. In a one-week sprint we designed Allie for a fictional app, Skaimart, around one idea: let shoppers tell the AI what they care about.
Full case study coming soon. This is a short version while I rebuild it with the full evidence.
See the full prototype- My role
- Product Designer (collaborating): research support, interaction design, prototyping, testing
- Team
- Nishad Patne (lead) · Gaurav Basantani
- Timeline
- 1 week · Sept 2025
- Tested with
- 20 online shoppers, ages 20–40
- Preferred Allie over Rufus
- 18 / 20
In 30 seconds
Store chatbots take control from shoppers: no memory, no preferences, walls of text.
Give shoppers visible controls: a saved AI Profile, in-chat comparison cards, and Allie on every listing.
18 of 20 testers preferred Allie to Amazon Rufus. All 20 named the AI Profile the feature they wanted most.
How might an AI help people shop without taking over?
The goal was to guide, the way a good in-store associate does, while leaving the choice with the shopper.
Are these bots actually useful?
Do they feel like help in a store, or do they frustrate?
Do they narrow choices and undermine control?
We tested 4 real store chatbots on the same 6 tasks.
Even the best one had gaps
- Tables for product comparisons
- Repeats the same recommendations
- Keeps context across screens
- Occasional hallucinations
- Cites sources, admits uncertainty
- Dense output, not really a conversation
Context is gone the moment you ask a follow-up.
You can't tell the bot what actually matters to you.
Comparisons come as paragraphs, not something you can scan.
The bot lives on one screen, not across the journey.
"Users have to cede control to the bot, rather than collaborating with it."Our audit summary
Go deeper: what 30+ articles and 4 papers saidThe secondary research behind our design direction+
30+ articles and 4 papers pointed the same way.
Personalization can narrow choice
AI-driven personalization can shift decision control away from the shopper, a risk worth designing against.
Great tech, room in the UX
Rufus uses retrieval, fine-tuning, and reinforcement learning. The experience still lags the technology.
Retail is betting on agents
Walmart is investing in agentic AI, and shoppers now trust AI picks as much as influencers.
"Assistants must be helpful and steerable, not black boxes."The principle we designed around
Three principles.
Visible agency controls
Shoppers can steer the bot mid-conversation.
Human-attuned interaction
Fast answers, voice or text, and cards instead of walls of text.
Transparency over constraint
Explain why each pick fits, and always show alternatives.
Tell Allie what matters once. It remembers.
A saved, shareable profile that travels with you across sessions, and even across apps.
Focus chips let shoppers invite surprises, like deals or sustainable picks.
Five-step sliders for price, quality, and brand turn vague wishes into something the AI can act on.
Every recommendation then says why it fits the profile, so shoppers can check the AI's reasoning.
Three ways Allie keeps shoppers in control.
In-chat comparisons
Side-by-side cards inside the chat: specs, pros and cons, price, ratings. Scan, then tap in.
Ask Allie everywhere
A persistent icon on every listing opens a chat that already knows the item you're viewing.
Portable AI Profile
Set preferences once and carry them across apps, like a companion that knows your taste.
18 of 20 shoppers preferred Allie.
20 online shoppers, ages 20–40. Task: find wireless headphones, first with Amazon Rufus, then with Allie.
In plain wordsEach person tried the same shopping task with two assistants, Amazon's and ours, then told us which one they preferred and why.
Preferred Allie's experience
Chose the AI Profile as their most-wanted feature
From audit to hi-fi prototype to testing
"I loved that I could actually tell it what mattered to me. Amazon just guesses."Participant, age 26, A/B test, Sept 2025
Agency beats accuracy.
- Letting people control AI behavior builds more trust than accuracy alone. Nobody wants a shopping assistant that "just guesses".
- On small screens, comparisons need cards, not walls of text.
- A focused one-week sprint can validate an idea faster than long ideation.
ReQuesta
Testing an AI workflow before the AI was ready.
Your move.
Open to Product Designer roles anywhere in the US.
gaurav.inbox29@gmail.com