Summary
One assistant, built to be simple for 80% of the company on day one
The goal was an intuitive AI platform for every Tempus employee — one that defaults to GPT functionality out of the box, integrates seamlessly with Tempus-specific AI agents, and keeps proprietary data handled securely throughout. It shipped as Tempus One Internal: empowering innovation with secure, in-house conversational AI.
I worked the problem from requirements through usability testing: brainstorming user types with Engineering, pitching the resulting segmentation to leadership, building and testing two competing prototypes, and coding twelve rounds of usability feedback into the themes that shaped the shipped design.
01 — Defining the product
Two brainstorms, one reality check from the CTO
Before any screens, the team had to agree on who this was actually for. That took a joint brainstorm with Engineering, then a pitch to leadership to confirm the split held up.
Brainstorm: types of users and their needs
We ran brainstorming sessions with Engineering to identify the types of users who would interact with One Internal — and how much of the tool each type would actually need.
Outcome
- Most users — need to be able to ask questions to an LLM in a secure environment.
- Advanced users — need to be able to adjust the settings and data sources of One Internal to suit their needs.
"We presented the outcome of the brainstorm with our CTO and they agreed on the users and needs, and provided business context that we should optimize One Internal for the most basic users — around 80% of our users — and 20% being advanced."
Pitch to leadership
That split became a literal capability ladder: what a basic chat can do versus what requires stepping into agent building, laid out from the most casual entry point (a Slack conversation) to the most advanced (rearranging blocks, hooking an agent to a production system, or opening a pull request).
Mocks and prototyping
From the tier map, I sketched the two entry points that mattered most for the 80% majority: the home-screen-to-chat path, and a lightweight way to ask for an agent that didn't exist yet.
I created two different prototypes to test these concepts out with users in order to validate our ideas:
02 — Testing process
Twelve usability tests, coded into themes
Ran usability tests with 12 users and then coded themes where we saw them in the recorded transcripts.
The prompt behind every session
"What are your first impressions of the screen that we're showing you?"
Opening question, every usability session
Key takeaways — how might we
Users had confusion around: what an Agent is, why Agents appear directly on the homepage, and the concept of an "active" Agent and how it can be leveraged.
What we needed to address
- Most users did not understand the concept of what an Agent is
- Most users did not understand the reason for having Agents directly on the homepage
- Most users did not understand the concept of an active Agent and how that can be leveraged
How we responded
- Eliminate the need to understand what an Agent is upon first login
- Make the homepage as simple and intuitive as possible to guide new users
- Eliminate the need to toggle an Agent on or off — just suggest a relevant one based on the question asked
03 — What shipped
Empowering innovation with secure, in-house conversational AI
The how-might-we above resolved into three concrete capabilities, then a revised homepage that put them all in one place.
Core capabilities — voice, categories, and getting suggested the right agent
Revised homepage — the design that shipped
04 — Impact
Usage metrics after 10 months
