top of page
Data collection for genAI features in Google Workspace
I led designs on this project for Google Workspace from 0 to 1. This 6-month long initiative, spanning multiple 1P teams, legal, and customer feedback teams, resulted in a standardized feedback flow that will be adopted by all Workspace mobile genAI features. Initial designs are live as of August 2024 🎉.
Main contributions
+ UX design lead
+ Mobile first experience
+ Led all reviews with legal
+ Partnership with UXR for testing and improvements
Deliverables
+ End to end mobile UX designs
+ High fidelity end to end prototypes
+ A11y documentation
+ Share-able component
In a nutshell
I led the design and development of the first mobile genAI feedback flow in Google's Workspace Suite, initially launching on Google Keep. This initiative enabled users to provide feedback on new genAI features and contributed to training Google Workspace LLMs (large language models).
Success metrics
-
Final data donation flow designs saw an increase of 67% in prompt donations, leading to significant improvements to Workspace LLMs. ✅
-
Decreased time to completion by consolidating 5 screens from previous feedback flow into 1 screen. ✅
-
Components are fully re-usable. Google Chat, and Gemkick teams are already using new components. ✅
The problem
As Google Keep was launching Magic List, we needed a reliable way to collect real-time user feedback on new genAI features. Although there were other Workspace products with existing feedback flows, they weren't optimized for genAI features.
As a starting point I completed an audit of the existing feedback flow on Gmail, and looked at areas that could be improved.
I collaborated with my PM partner to generate a set of 'How Might We' questions...
HMW...
Collect feedback efficiently, offering both quick and comprehensive input?
HMW...
Collect the necessary data in full compliance with applicable laws and regulations?
HMW...
Collect data specifically designed to inform and drive product enhancements?
From here we arrived at a set of feature goals:
Business goals
.png){kind=small}
-
Feedback flow needs to be re-usable by any other 1P teams
-
Final designs need to adhere to new Gemini design patterns
User goals
-
Users should be able to quickly submit feedback
-
Users should understand how their donated data will be used
Initial designs
Exploration 1 - Modular Card Experience (no data collection)
Pros
✅ Closely aligns with existing Gemini feedback flow patterns
✅ Modular approach works for panel, and full screen experiences
✅ Modular approach scales responsively between web and mobile experiences
Cons
❌ New card component is a custom component that will cost extra eng time to build
❌ Limited information can fit in card space
Exploration 2 - Full Screen Experience (collect both input and output data)
Pros
✅ Easier to build, and can save eng cost
✅ Additional vertical space allows for more legalese if needed
✅ Prompt donation is clear and users can see exactly what they are donating
Cons
❌ Full screen experience might feel heavy for users
How much data should we collect?
Final designs
Bite sized experience, for actionable results
1. Users can quickly select feedback category to further enhance quality of data donated
2. Submit button is above the keyboard on mobile if it were to be opened
3. All T&S's are approved by stringent legal standards
4. We automatically collect prompt data, and users can opt-out of donating output data.
Solution is scalable and responsive for multi-modal experiences
Feedback flow designs on Gemkick panel in Gmail
Final thoughts 💭
When building new genAI features, including methods to collect feedback from users is just as important as the feature itself. Developing genAI features without user feedback is like playing golf n the dark - impossible. As my team worked on this feature there were numerous requests from other teams (Gmail, Calendar, Chat) that wanted to re-use this component. My team's proactive approach resulted in a valuable data collection tool now leveraged across Workspace, demonstrating the power of user-centered design.
bottom of page