Gubagoo: Chat Copilot
Before ChatGPT made "copilot" standard product vocabulary, I designed Chat Copilot, my first project building user-controlled AI into a product experience. It looks rudimentary now, but it's interesting to see what's held up and how AI patterns have matured.
tl;dr
- Problem Chat AI handed every lead to a human operator with no way back in, so operators typed contact-info requests and inventory lookups by hand.
- My role Sole designer, taking an ML engineer's technical concept to a piloted MVP alongside a front-end engineer.
- Solution A control bar above the composer that suggests the most relevant task, runs it, and shows progress on each chat card, handing control back to the operator if anything goes wrong.
- Impact In the pilot, an ops manager called it "my own personal assistant all day long" and a dealer BDC manager said teams were "hitting more team goals than ever."
Context
Gubagoo's chat AI greeted every dealership chat lead, then handed the chat to a human operator when it was ready, with no way for AI to reengage. Operators had to handle every chat start to finish, with no way to automate any part of the workflow, so those working several chats at once were typing out contact-info requests and simple inventory lookups by hand. The engineers brought me a technical concept: let operators bring the AI back into the chat. I owned the design from there, working with an ML engineer and a front-end engineer. I turned the concept into a stance: the operator stays in the loop, and the AI works for them rather than taking over.
This was a rough MVP, so I designed from intuition rather than formal research. My feedback loop was the operator team leads, whom I met with to react to the design as it took shape.
In pilot testing, an ops manager described it as having "my own personal assistant all day long." A dealer BDC manager said the AI now took the repetitive work, so operators could focus on the conversations that need a person, and teams were "hitting more team goals than ever."
About the solution
Copilot sits in a control bar above the message composer. It suggests the one most relevant task, runs it with the team's first machine-learning model, and shows its progress on each chat card. If anything goes wrong, it pulls the operator back in. The human stays in the loop and the AI can be called back in: these are the patterns AI copilots use today.