Can AI Help Users Configure Complex Products?
Exploring conversational interfaces for 3D product configurators.
Role: Product Designer
Timeline: November 2023 - April 2024
Topics: Workshop Facilitation, Usability Testing, Prototyping, AI
Overview
As an agency, we had built several 3D product configurators for clients. These tools simplify the sales process for companies selling customizable products, such as kitchen ranges, garage cabinets or made-to-order pillows, by allowing users to specify their preferences online.
– Example of a kitchen configurator
Intro
We wanted to explore ways in which AI could be used to enhance the online configuration experience. My team was tasked to reimagine the online shopping experience of custom-made or customizable products, through conversation. This is what we learned from our early experiments, using AI to help users create unique products online.
Process
At first, we didn't know if it would be possible to integrate a LLM (Large Language Model) in the context of 3D product configurators. My team created a proof of concept, a basic chatbot for the L'Atelier kitchen configurator.
– Proof of concept: configuring using natural language
I worked together with the UX Researcher to conduct usability testing on the early proof of concept. These were the main findings:
⚡Insights
- Users had negative perceptions of chatbots.
- The chat interface lacked clarity and affordances.
- The value proposition wasn't clear.
- Some users were impressed by the functionality.
- There were mixed feelings, including excitement and fear, towards AI.
– A few examples from usability testing
Creating a desirable experience
Following the proof of concept, our aim was to develop a desirable AI experience. Simplifying the process was crucial since our primary users were older, wealthy Americans. Additionally, architects and designers at physical showrooms used the tool to create quick proposals during client appointments.
– Example of a kitchen built by L'Atelier
Simulating a Sales Representative
I helped lead a workshop with key stakeholders, to better understand the offline sales process. The outcome was a hypothesis: simulating a sales representative would assist users in creating their dream kitchen. We wanted to replicate the initial conversation customers might have at a physical showroom.
– Example of a conversation flow, created to inform the behaviour and goals of the AI assistant
To test this, I created realistic prototypes where test subjects could type their answers. Sophie, our virtual sales representative, asked them four questions to gather requirements and create an initial recommendation.
– A prototype I created to test generating the initial product recommendation
What we learned
Despite numerous iterations, we consistently found:
⚡Insights
- Users preferred conventional interfaces
- They wanted to interact with the product
- Text input was inefficient
- Users desired to be in control of the experience
- Interaction with the chat lacked affordances, visual cues to help guide the experience
– The main insight shared by our UX researcher
Finding purpose
Based on these findings, we aligned with stakeholders. We agreed that the AI capabilities would be best utilized for generating recommendations, answering questions, and assisting users as needed. We shifted away from a fully guided design process and focused on providing value at key moments.
Next steps
For the initial release, we simplified the interaction with the virtual sales representative, Sophie, to a single dialog. Users could choose to design independently, seek help from Sophie, or use text input for a starting point. We planned to gather real data to inform further decisions.
In the meantime, we continued exploring more contextual approaches that would better integrate into the whole design experience. The first results from unmoderated testing looked very promising - for the first time in usability testing, we saw people intuitively interact with the AI feature and ask questions about the product. We'll be exploring this contextual direction further.
– Prototype I created to test the contextual approach.
As the user interacts with the app, suggestions
adapt to their actions.
