Less repetitive operations. More capacity to think about the system.

The Design System team answered the same design and technology questions about the system and about accessibility every day. That was about 200 requests a week: capable people stopping their work to repeat answers that already existed in the documentation.

A chatbot that serves anyone, on any Design System and accessibility topic. It runs on an internal LLM platform, queries MCPs linked to it and uses the system's corporate documentation as its knowledge base.

My part was product and design: how it would look, who it would serve and what the tone of voice of the conversation would be. The technical side was handled by other people.

1. Question
2. Knowledge base and MCPs
3. Answer

The most common mistake was context. When it didn't understand who was on the other side, dev, QA, product or design, the chatbot hallucinated and almost always treated the person as QA. We started teaching good practices for opening a conversation: introduce yourself, say your area and give context about the screens.

It logs errors and conversations, so we can see repeated questions that caused problems and fix things from there. The team that uses it knows it can make mistakes, and that helped it evolve faster.

• ~90%drop in requests to the team, which were about 200 a week. There were several changes in the period, and the chatbot was one of the fronts that contributed the most.

I use AI where it removes friction, increases capacity and frees people for higher-value decisions. Less repetitive operations, more capacity to think about the system. When it doesn't solve a real problem, it doesn't need to be there.