AI-Powered Workspace
Role: Product designer
Stakeholders: Design systems lead, Product team
Duration: 5 weeks
Deliverables: Research artifacts, hi-fi designs, stakeholder presentations
Background: This case study represents my contribution to a 4-person internship capstone, focusing on developing AI use cases. I collaborated closely with design and product leaders throughout the project. This experience was invaluable for acquiring both technical skills and adeptly handling diverse stakeholder expectations.
The problem
“Current generative AI and other technologies have the potential to automate work activities that absorb 60 to 70 percent of employees’ time today.”
-McKinsey, The Economic Potential of Generative AI, 2023
Generative AI is a technology that could transform our society, but how will it actually interact with the existing digital tools and platforms we use?
This project takes a systems-level approach, asking how might we embed generative AI into existing enterprise systems in a way that is intuitive, powerful, and realistic.
*specific enterprise systems have been anonymized
The people
Judith is a landlord with minimal AI knowledge who has recently gained new properties. She is looking for new tools that will simplify the process of drafting rental agreements without overwhelming her with technical complexities.
Octavia is a small business owner who has experience using ChatGPT for writing emails or creating copy ideas. Tech-proficient and creative, she's in search of a more robust and personalized generative AI tools.
Judith and Octavia are both active end-users of a business management platform, representing the varied experiences with generative AI in its current user base. It is critical that any embedded AI tool will fit both their needs and incoming knowledge.
The Research
I began the process of research by gathering and analyzing over 30 use cases collected by product management teams, using their metrics to create a prioritization matrix and grouping them into overarching themes. Very quickly, I realized that the various use cases could be generalized into two categories: information search and synthesis and content generation.
I then conducted a competitive analysis of current genAI products with a focus on these two tasks, finding several key features: a centralized dashboard, chat management system, response citations, and file upload, and an overall focus on conversational UI.
Additionally, I looked conducted a few exploratory user interviews, as well as looked at academic studies of AI chatbots, to identify common user concerns and preferences.
The Designs
User journey
With this research, I began mapping out what different task flows would look like for Judith and Octavia and how that would translate to both state design decisions and back-end needs.
Some key questions to answer were:
How do users first discover the AI tool and how do they toggle it as they move through the existing platform?
How and where do you communicate data privacy and inaccuracy risks while building user trust?
Where can users provide feedback to be collected and used to further train the large-language model?
Based on the user journey mapping, I broke up my approach to designing the tool into 3 main areas of focus: navigation, chat management, and chat responses.
Styles and Fonts
The platforms I was building out this tool for had an established design system for which I followed closely in determining style. However, in order to distinguish the embedded tool and tie together its different elements, I used a specific purple motif (highlighted).
#EDEAE6
#2F2D46
#5B5397
#E0EBF5
Navigation
#666666
It was critical while working with the platform development team that the AI tool blended with the existing navigation system. A challenge of this requirement is that the tool could not exist on the main navigation.
Considering this, I crafted two primary ways to access the tool: through the search function and a pop-up chat. The search call-to-action leverages the familiar user behavior of seeking answers, thereby enhancing the discoverability of generated AI insights. On the other hand, the pop-up chat offers users a direct route to an AI-powered chat feature, ensuring a straightforward pathway for engagement.
Chat Dashboard and management
Both Judith and Octavia make use of current platforms for various management duties, which can become quite overwhelming. As a result, effectively handling chats and their content becomes crucial. Our competitive research highlighted that a centralized dashboard is a pivotal design choice, differing from a sole reliance on an infinite-scroll pop-up chat.
A dashboard also allowed me to incorporate several other high-impact features, including the ability to upload and converse your own files, as well as pin important points in the conversation.
One of the project’s design priorities was the design of the AI-generated responses themselves. Multiple facets like tone, sourcing, and interaction patterns were intangible and closely tied to model training. Consequently, my designs concentrated on only manipulating elements that have been proven feasible with existing models.
Several design decisions aimed to enhance the generated responses. These encompassed incorporating citations for users, establishing a feedback mechanism, maintaining a professional tone, clarifying the tool's limitations, and facilitating users with pre-structured prompts. These choices collectively bolstered the effectiveness of the AI-generated responses and increase trust.
Response Generation
Final Screens
Reflections
Being a young designer, this project was an amazing learning journey. It marked one of the first times collaborating with a globally dispersed design team. I gained insight into effective stakeholder management, the importance of actively seeking mentorship, and the art of establishing clear goals and boundaries. Moreover, I was able to delve into the intricacies of designing for extensive enterprise systems and observe the critical coordination required among product, development, and design teams to make something special.