Shaping the Future of AI at UW
Improving onboarding, usability, and accessibility to support successful first-time use across a campus-wide AI rollout.
"Assume the customer knows nothing and has only tried ChatGPT once before. That is our target. We want to maximize the 'WOW!' factor with Purple so users see how powerful this can be."
— Jared, AI Lead (my manager)
The Context
UW-IT is launching Purple, a secure, university-exclusive GenAI platform, to 100,000+ community members. I joined the team as the sole UX Designer to help its deployment, transforming an early-stage technical sandbox into a trusted service that supports teaching, learning, and daily operations across campus while ensuring strict data privacy.
Purple - UW’s custom AI platform (by 2025.12)
project Goal
To successfully launch Purple to the entire UW community, with a primary focus on building institutional trust and driving user adoption and engagement.
My Contribution
As the UX Designer, I took full ownership of the end-to-end first-time experience strategy and execution. Working as the only designer on a technical team, I bridged the gap between the backend infrastructure and the end-user.
Leading exploratory and evaluative research to identify and prioritize critical experience blockers.
Designing the onboarding strategy and content to ensure "first-time" success for AI novices.
Engineering in-app agents and starter prompts to improve response quality and accelerate user value-discovery.
Collaborating with Communications, Engineering, Web, and Support teams to ensure a unified, supportive rollout.
Key Achievements
Onboarding Design: Delivered a localized onboarding kit (content-oriented) targeting top adoption blockers, reducing first-time friction.
Usability, Accessibility & Compliance: Shipped 20+ accessibility fixes and translated 30+ UI/UX issues into vendor-ready specs, reducing compliance risk and product friction in critical user flows.
Strategic Influence: Conducted user research and translated user insights into actionable themes that helped inform the product rollout strategy and long-term roadmap.
Feedback Program Design: Designed a structured beta feedback program with a human-in-the-loop testing loop to systematically collect and triage issues, increasing actionable user feedback by 8× in one month.
AI Reliability & Grounding: Improved factual grounding and response consistency by iterating system prompts and refining knowledge sources based on recurring failure cases surfaced through human testing.
Role
UX Design, Research, Strategy
Teams
Internal: Data & Engineering, Research, Comms, Web, Documentation & Support
External: Vendor Project Manager
Timeline
Jun - Dec 2025
Skills
AI & Agentic UX
Prompt Engineering
AI Eval
Heuristic Evaluation
Accessibility Audit
Project Timeline
the challenge - The data gap
We lacked insight into general users, our main target.
As we moved from a "Dev Sandbox" to campus-wide production, we faced a major disconnect in our user data:
Existing Data: Early pilot feedback came almost exclusively from power users, who are highly technical and understood AI architecture.
Our target: Our 100k+ rollout targeted general users, many of whom were little-experienced or first-time AI users with different needs and expectations.
Hence, we had a foundational knowledge gap: no representative insights on how non-technical users would perceive or use the platform.
How might we ensure Purple is usable, accessible, and trusted, so general users can succeed on their first visit?
Before the initial rollout, we needed to reduce early friction that would otherwise tank adoption and engagement. Given the limited timeline, I began by identifying existing friction points in the product’s navigation and interaction logic to establish a usability baseline.
Establishing a usability + accessibility baseline before rollout
I assessed Purple’s interface through an accessibility + UX audit and selective benchmarking against familiar AI tools (Claude, Perplexity, Copilot, etc.) to identify the patterns users already expect, and the friction points most likely to derail first-time success.
Turning findings into prioritized vendor action
I translated the audit and analysis findings into a prioritized Accessibility + UX backlog for the vendor, outlining 20+ accessibility fixes and 30+ UX/UI improvements with clear problem statements, recommended solutions, and severity-based prioritization.
Prioritized UX + Accessibility issue backlog delivered to vendor (excerpt)
Outcome:
Reduced launch risk with vendor-shipped fixes
The prioritized backlog helped the vendor move quickly on launch-critical issues. By working with the vendor, we resolved high-priority blockers, improving product usability and reducing compliance risk ahead of rollout.
Understanding first-time user mental models
While the audit addressed technical quality, I needed to gather data on the human experience, so our goals for the launch could be clearer. Specifically, we need insights into how general users would interpret Purple on day one: what they expected, trusted, and found confusing. Hence, I conducted a research study with AI novices to uncover their expectations and concerns.
Method:
Moderated formative usability testing
I conducted 7 1-on-1 moderated sessions combining task-based usability walkthroughs and semi-structured interviews with AI novices across job roles.
Across sessions, I aimed to:
uncover first-time users’ mental models, expectations, concerns for Purple and AI use
evaluate user comprehension on how to select and use specialized agents.
Affinity map of user insights,
including comm reqs, onboarding needs, agent feedback
Insights informed designs, and rollout decisions.
I synthesized the data into an affinity map, clustering findings across product UX/UI, onboarding and training needs, agent comprehension and discoverability, and broader expectations, concerns, and desired use cases for AI.
These insights not only informed design & engineering decisions but shaped our cross-functional rollout plan, providing the evidence needed for our Communications and Training teams to support users on day one.
1. Data privacy and compliance are top-of-mind concerns for users.
Users wanted clear reassurance about data privacy, retention/storage, and usage/compliance guidelines, and were skeptical without institutional confirmation. Trust information should be highly visible and easy to find early.
2. Agent selection creates choice overload and blocks first-time success.
Users struggled to understand how agents differed and which one to choose, often misunderstanding purpose or functionality. Clearer onboarding and discovery guidance would reduce friction and speed time-to-value.
3. First-time users need confidence-building guidance to use agents effectively.
When agent capabilities aren’t explicit, users default to probing and comparing agents to figure out what works, undermining confidence. Onboarding should make agent purpose and first steps obvious to reduce uncertainty and make the first interaction feel safe and easy.
Design solution
Designing onboarding through content
Because the third-party interface offered limited flexibility for UI changes, I leaned into Content Design, and worked in collaboration with the Comms and Documentation teams. I mapped out a series of onboarding touchpoints to guide users through the journey, and each addressed a specific friction observed in testing.
Onboarding Flow
1
UW-IT webpage: “Get Started” KB article
Goal: Help new users form the right mental model, learn the basics quickly, and feel safe to try Purple.
New users would enter Purple from UW-IT announcements and documentation, so the “Get Started” article becomes the first impression for many first-time users. I drafted the article structure and content to answer the questions users had before they feel ready to click into the product.
✍️ What I included and why:
Step-by-step onboarding to compensate for limited in-product walkthrough.
Explain "agent" and their differences to lower barriers and set expectations.
Privacy & compliance notes to address trust concerns early.
Tips for asking good questions to help users get a better first outcome and reduce “blank page” anxiety.
"Get Started" KB Article
Feel free to check out the full article here!
2
In-product usage guidelines
Goal: Make safe-use expectations and data handling understandable in the moment, not behind external links, to help users feel confident using the tool responsibly.
In early versions, the usage guidelines were mostly listing external links to broader GenAI policy pages. In testing, users wanted a clear answer right away—especially around what data is stored, where it lives, and what not to enter.
✍️ What I changed:
Pulled the most important policy information into the product with clear hierarchy, so users don’t have to leave Purple to understand basic safety.
Added plain-language clarity around chat history storage / retention / approved data types, aligned with concerns surfaced in research.
Key policy information shown in the context
3
In-product agent descriptions
Goal: Help users choose the right agent to start chatting with by clarifying scope, strengths, and limitations.
In testing, the unclear descriptions made it confusing for users to choose the right agent for their needs. I collaborated with Comms to rewrite the “on-hover” (directory) descriptions and the in-chat descriptions to be consistent, credible, and easy to skim.
✍️ Design Principles:
Clearly communicate each agent’s purpose, best-fit tasks, and boundaries
Keep tone consistent, credible, and user-friendly across agent
Before
Long, buzz-word heavy, and generic claims
After
More scannable and focused description
4
In-Chat Starter prompts
Goal: Help first-time users get a quick win without feeling overwhelmed or excluded.
In testing, some starter prompts felt too random or too role-specific, which made users unsure whether the agent could help with their needs. Some prompts also triggered overly long responses (1+ page scroll), making “trying a prompt” feel overwhelming. I refined the prompts to demonstrate core capabilities, set expectations, and keep the first interaction lightweight.
Key Constraints:
UI limit: The UI supports max 4 prompt slots per agent.
Audience: Official agents serve broad UW audiences (staff/faculty/students) and general tasks, so prompts should be understandable without assuming a specific role or context.
✍️ Design Principles:
I grounded my design principles in testing insights and prompt UX guidance and best practices (article from NN/G).
Role-neutral & inclusive phrasing so prompts don’t feel “for staff only” or “for students only”
Prompts that show capability and set expectations (what this agent is good for)
Simple examples to keep the first response within ~0.5 page scroll, so “trying a prompt” doesn’t become overwhelming
Before
Too random or role-specific prompts
After
More task-relevant, role-neutral starters that set expectations
For Prompt Iteration:
Given the 4-slot limit, as Purple rolls out to new audiences and more agents are added, we’ll refresh starter prompts based on the most common user intents and early friction we observe, so prompts stay relevant and lightweight.
Result
🎉 Reduced first-time confusion and improved time-to-first-value
After these onboarding touchpoints shipped and Purple entered soft launch and beta, we saw clear qualitative signals that the experience was easier to start:
Agent clarity improved: In earlier testing, 6 of 7 participants found agent purposes unclear; during soft launch/beta, we received almost no feedback about agent confusion.
Smoother time-to-first-value: Compared to early testing, users reported reaching a first successful interaction more quickly.