Embedded Files
In early 2023, generative AI was new and almost entirely uncharted from a design perspective. This project built a secure internal chatbot that gave employees instant, AI-powered assistance with creative, analytical, and technical tasks — drawing on proprietary company knowledge. With no established design patterns to follow, the experience was built from first principles, guided entirely by user feedback. The chatbot's core framework later became the foundation for a major enterprise engagement in the healthcare sector.
ROLES
UX Designer, Presenter
RESPONSIBILITIES
Interaction Design · Information Architecture · UX Writing · Concept Presentation
IMPACT
Reduced employee support tickets for common technical issues through self-service AI assistance
Drove broader AI adoption internally through intuitive, trust-first design
Chatbot framework became the foundation for a landmark enterprise deal in the healthcare sector
QUICK FACTS
Team: Cross-functional: Employees across functions
Context: Internal enterprise tool · Generative AI · 2023
Notable: Core framework adopted in a major enterprise healthcare engagement
Visuals: Wireframes shown. Full designs anonymized.
Problem
Employees in 2023 knew generative AI existed. Fewer knew what to do with it in a secure, professional context — and with good reason. Public AI tools couldn't touch confidential company data. The goal was an internal chatbot that drew on proprietary knowledge to give employees real, accurate assistance — not generic responses from a public model.
The harder problem wasn't technical. It was human. Trust in AI was fragile. Expectations were either too high or too skeptical. And unlike designing a form or dashboard, there was no established design language for enterprise conversational AI. The canvas was genuinely blank.
Process
Iterating with employees, not for them
Rather than benchmarking against competitors, the approach anchored in familiar interface conventions and layered new AI-specific elements carefully on top. User feedback wasn't a phase — it was the process. Continuous testing with employees across functions surfaced two critical insights: employees didn't know what to ask the chatbot, and they had no way of knowing how much to trust what it said. Both became primary design challenges.
Key Design Decisions
Suggested prompts as the entry point
The blank input field created immediate friction — employees opened the tool and froze. Suggested prompts displayed on the initial screen offered a starting point without requiring prior understanding of the tool's range. Each prompt modeled a different type of use: creative, analytical, technical.
A prompt library for deeper discovery
A browsable library of sample prompts organized by task type educated employees about the chatbot's full capabilities over time. Strategically, it also reduced support tickets by proactively showing employees how to solve common problems themselves.
Source citations for transparency
When the chatbot used a company document, a clickable filename linked directly to the source. Employees could verify what they were being told, read the original context, and decide whether to dig deeper. The chatbot stopped being a black box and started showing its work.
A confidence level indicator
Not all responses are equal. A confidence indicator displayed alongside each response communicated how directly and reliably the answer was sourced from company knowledge. This reframed the chatbot from something employees had to trust blindly into something they could use with informed judgment.
Reflection
This project taught me that trust is a design requirement, not a feature. Citations and confidence indicators weren't added at the end to make the tool feel safer — they emerged from watching employees hesitate before acting on early outputs. That hesitation was data.
Working without established patterns was uncomfortable at first and clarifying by the end. When there's no template to follow, you return to fundamentals: who is this person, what do they need, what will make them confident enough to act.
Impact
What started as an internal experiment became something larger. The design framework built here — the trust architecture, transparency features, user-centered approach to conversational AI — proved transferable. When a major opportunity emerged in the healthcare sector, the groundwork was already laid.
That outcome wasn't planned. It was the result of building something carefully, with real users, at a moment when most organizations were still deciding whether to engage with generative AI at all. Getting in early, and getting it right, turned an internal tool into a proof of concept that traveled further than anyone expected.
Google Sites
Report abuse