.png)
.png)
Don't just design AI; Design with AI.
The Problem
HMW Increase visibility of TMP’s broad range of services and resources so that mentoring programs can independently navigate and find relevant information faster?
Target Users: Leaders in mentoring organizations, typically directors and founders, who are responsible for identifying and utilizing resources to address organizational challenges, such as staff training and DEI initiatives.
The findings so far
There are pain points at Every stage of the user's search journey
The unanswered Question
Given these pain points, when and how should AI intervene?
AI can do a lot, but let's be real—AI is not that smart.
It can make mistakes, has limitations, and sometimes, people simply don’t want to use AI for certain tasks, especially for the mentoring community, where people have long relied on human-to-human interaction.
Qiyu's Prototype
The AI Call Agent, Meet users where they are
Phone calls are currently an important communication channel between TMP and its clients, but many inquiries are repetitive, creating a burden for TMP’s small team. Therefore, we see an opportunity to automate this service.
Transcript:
AI: Hello! You've reached the Mentoring Partnership. I'm Michelle, and AI from TMP and I'm here to assist you with any questions you might have.
Before we start, could you please let me know which language you prefer to communicate in? We support English, German, and Espignol.
Transcript:
AI: Could you please provide your name and the name of your organization?
User: This is Qiyu from the Magic Pals.
AI: Ah, Qiyu from the Magic Pals, welcome back, if I recall correctly during our previous conversation, we discussed resources for mentoring training. How did the echo for you?
Transcript:
User: Mentoring 101 was good. I also wonder if there's any resource for the mentee. We have lots of young people from diverse backgrounds, so any resource for them would be helpful too.
AI: So I understand mentoring 101 was helpful for your mentors for your mentees from diverse backgrounds, we have a resource making the most of mentoring. It's designed to assist youth in building networks and setting goals effectively.
...
AI: Additionally, I noticed your emphasis on diversity. Have you heard about our cultural workshop It's crafted to empower mentors in diverse mentoring relationships, which would be perfect for the magic pals. Would you like more details about this workshop or to schedule one?
behind the scene
Take a closer look at my prompts
Prompts need to be data-driven—not just for personalization, but to scale services effectively.
I transformed TMP’s human-to-human interactions into AI-driven experiences. The AI agent will first ask about the user’s name and then leverage Salesforce data (i.e. TMP's current data storage tool) to retrieve past interactions. This allows the AI to deliver personalized, context-aware responses, ensuring an informed user experience.
Warm language isn’t just about politeness—it also means giving space for users to clarify or correct information. The example above, which simply states "greet them warmly," was only the initial prompt I provided to the AI. From there, I conducted iterative refinements, adjusting the tone for specific scenarios through repeated testing and fine-tuning.
To further define what "warm language" means, I applied insights from my conversational AI research, refining responses with modal particles, humble phrasing, and nuanced wording. This iterative process ensured that the AI’s tone felt natural, empathetic, and contextually aware, rather than relying on generic warmth.4o
More details
why this prototype, why using chatGPT, and the Impacts...
Prototyping goal
Where is the "Low-Hanging Fruit?"
.png)
We often hear the phrase "low-hanging fruit" in AI UX design. In my opinion, it refers to the intersection of user needs, technological capabilities, and business budget.
In other words, it is the "When and How" AI should intervene, based on research-driven insights into user needs and business constraints.
the strategy
Design can be a research tool
My comments on the research protocol, explaining what "research through design" is
“Research through design employs methods and processes to create artifacts that explore, embody, and evaluate ideas, generating knowledge that can be applied beyond the artifact itself.” (Zimmerman, Forlizzi, & Evenson, 2007)
Check out my other research-though-design projects with Prof. Zimmerman
Leveraging my research skills, I defined clear prototyping goals for my call agent, focusing on the following questions about Human-AI interaction:
- Context: When does technology fit into the user’s journey?
- Intervention level: How much support should AI provide?
- Modality: Should the solution be visual, audio, or something else?
The collaboration
Applying Parallel Prototyping to understand human-AI interactions
To explore and compare different solutions efficiently, Our team employed "Parallel Prototyping" to test different prototypes. I brought up "the phone call agent" ideas and therefore was responsible for this prototype.
Efficiency & Quality
Prompt Engineering for Optimization
You don’t need to be an engineer to prototype AI-driven conversations. As a frequent user of ChatGPT’s audio feature myself, I saw an opportunity to leverage it for rapidly prototyping a conversational AI call agent.
Through prompt tuning, I developed a framework that guides users in selecting mentoring resources. Grounded in user research, I iterated on interaction flows to ensure clarity, engagement, and alignment with user expectations.
♾️
Test, Revise, and Iterate
I led a client workshop to gather feedback for the first iteration, using role-playing exercises to simulate real scenarios and uncover edge cases. Key questions explored included:
- What errors occur, and how do users respond?
- What types of conversations occur with AI across the user journey (e.g., browsing, selecting, applying), and which don't?
- How do users perceive AI-generated personalized questions?
The insights
The "knowledge that can be applied beyond the artifact itself"
Exciting discoveries from our testing sessions! This table captures all key findings, but one insight stood out—user autonomy in decision-making. Participants wanted AI as an assistant, not a decision-maker, leading us to a crucial takeaway: avoid over-personalization to maintain user control and trust.
Impact
Fast prototype, Authentic interaction
Save time
with prompt-based rapid prototyping using large models like GPT-4.
.png)
Learns Dynamically
conversational dynamics for faster validation and refinement of ideas
My prototype explores embedding AI into a traditionally non-AI space while navigating the caution of a small mentoring community. Through a research-driven, evidence-based process, I identified key opportunities for innovation and gathered valuable insights to inform future prototypes, ensuring a thoughtful and adaptable approach to AI integration.4o
Colleague's feedback
Qiyu always brings a lot of interesting and creative ideas to the table... I also want to commend her on her willingness to experiment and try new things.
For example, the creation of our third prototype powered by Chat GPT that she led. I feel like that was a new thing she tried out and it really impressed me.
--- Harshini, Project Manager, working on the same team
