Designing a Conversational chatbot for UNIQLO
Developing an AI Chatbot for UNIQLO's discovery and order user flows through VoiceFlow. Adding intents for the LLM to develop, the project created a working conversational system.
4 weeks
E-commerce
2 CUX Designers
Voiceflow, Figma, Google Suite
My Role
UX Design
Conversational Design
Impact created
Creation of a new Conversational system
Personalized discoverability of products
Streamlined userflows for selection and check-out
Natural user utterances for better engagement
3
Focused user flows
15%
Increase in user retention
60%
Increase in time spent on website
*compared to previous version
Overview
The project focuses on developing a functional conversational AI agent equipped with 'Intents' to initiate specific conversation flows based on natural user utterances. By integrating with a large language model (LLM), we enhance the contextual understanding of terminology, specifically tailored for UNIQLO. The design prioritizes intuitive user paths, enabling users to engage with the chatbot in a natural and meaningful way through voice user interface (VUI) design. Additionally, we developed the chatbot's conversational user interface (CUI) to seamlessly integrate with UNIQLO's website and app, ensuring a cohesive user experience across platforms.
The chatbot hence intends to enhance specific user flows to improve discoverability of products, minimising cognitive load.
UNIQLO Chatbot: Effortlessly stylish, playfully engaging, and always reliable…
We detailed out the characteristics of the chatbot in alignment with UNIQLO's brand philosophy and the target audience's communication style.
Tone: Clear, concise, and upbeat, reflecting Uniqlo's minimalist, playful brand.
Text Quality: Warm, inviting, and professional, with precise, well-organized information to minimize cognitive load.
Communication: Confident, structured responses; clear hierarchy of information for easy readability.
Response Time: Optimized for fast, efficient replies and issue resolution.
Interests: Incorporates relevant knowledge of fashion, tech, and sustainability.
Behavior: Proactive, personalized assistance with clear guidance to reduce effort for users.
Reliability: Accurate, dependable responses with a focus on clarity and efficiency.
Task Flow to User 'INTENTS': Research for Identifying major high traffic and complex user flows
I started by interviewing users regarding their most used searches on clothing apps, and what they normally use the apps for.
Contextual inquiries with users when they online shopped helped uncover patterns highlighting the recurring user flows of:
Casual browsing (Broad): "I want to see womens' clothing,", "I want to see latest arrivals", "something new and warm"
Product finding (Specific with multiple layers of specificity): " I want to buy a warm black dress in my size available this weekend for a pick-up." " It should be less than 80 USD."
Order Management: "i want to return my second-last order", "Can I still exchange this?"
Purchashing (Cart to order placement): "I want to place this order for delivery." "I want to pick my order from my nearest store."
Three of these user intents were branched to sub-intents and Level 1 Direct Intents for charting out options present within the UNIQLO ecosystem with all its touchpoints to cater to these intents.
Intents to UTTERANCES:
Observing how people normally communicate to and fro, or convey their actual needs in a natural manner…
how many touchpoints would the chatbot system need to trigger specific flows?
It makes sense for a chatbot to enable the user to ask questions directly rather than going through a flow like you would on a website or an app. Hence, for each possible query a user could ask for, keywords and sentences with multiple keywords in them were added at multiple steps for each Intent such that
The agent can recognize the intent of the user at a main or a sub-level category.
The agent can re-direct to that particular moment in the flow and give a direct, focused reply to the user.
Multiple behavioral ways of user's asking, talking, writing can be accommodated by the agent with an accurate response.
Integrating behavioral traits, mannerisms of communication by crafting typologies of 'Utterances'.
I realised it was better to be extremely specific from the beginning for the LLM to trigger the correct agent response flows.
I re
How can I make the user experience of the conversation be more natural, helpful and intuitive?
Strategies incorporated on user testing and desk research:
Naturalness of conversation and accuracy provided to users: In order to reduce cognitive overload and keep users constantly informed in their conversation with the chatbot, I integrated codes for the Cart total to show up in conversation as and when needed.
Users want to be constantly aware of their Cart expenses.
Transparency builds trust.
In order to reduce cognitive overload and keep users constantly informed in their conversation with the chatbot, I integrated codes for the Cart total to show up in conversation as and when needed.
This reduced chat drop-rates by 30% in users trying to go to the Cart manually to see their total. It also made them trust the chatbot more and hence, willing to use it more.
Drop-off rates decrease when users feel supported in tasks they are unable to perform.
Subtle hints and help provided improve user satisfaction and speed.
Recall to when you are unable to find your order number or your account number, and you do not even recall where to find them. I added error pathways for particular conditions to handle exceptions and guide users through issues smoothly, ensuring a seamless experience even when problems arise. I believed doing this properly was critical for reducing user frustration.
Additionally, I felt it was important to add in subtle ways to help users through some tasks, as with the example shown below. I added supportive information to guide users through tasks using the 'No match' flow text and structure.
The prototype worked across multiple user task intents to provide accurate responses with its designed characteristics.
I believe additional rounds of user testing always helps in the final touches. I shifted some utterances for some specific user behavioral styles during user testing.
In observing and talking to people use the prototype, I took the approach of not providing any tasks. I wanted to observe how different users wish to engage with the chatbot, and I realized multiple patterns and behavioral styles.
Some users had a free-style way of conversation, in long text responses describing what they wanted in great detail.
" I want a long-sleeved blazer that is grey in color and is warm." The chatbot failed to give accurate product results in these contexts. I realised I need to go back in and input detailed layer of product description to the LLM for it to internally categorize the products in different layers. Here, I did a manual coding to put specific intents in the scope of the project.
Some users consistently chose to type in responses even with clear CTA buttons in the chatbox. I had not expected this mental model and hence added other sub-intents that connected to the utterances in addition to the CTA connections.
Some options that I intended to minimise some steps or confusion were taken out as I realized they were not naturally expected by the user.
Envisioning how the Conversational User Interface will look like on UNIQLO…
Taking into account Uniqlo's visuals and interaction styles, I envisioned how the interface would look like on mobile app, tablet and desktop website versions.
How I can develop this further
My key take-aways
Specific Intents matter: Using multiple, specific intents ensures more accurate chatbot responses compared to joint intents, leading to better user interactions.
Error Pathways enhance UX: Adding error pathways and subtle tips for No Match conditions helps users recover quickly, improving the overall flow and user satisfaction.
User-centered design: A focus on guiding users with clear, actionable steps at every point keeps the user experience smooth and intuitive.
