An AI assistant operations teams could trust

The problem. Operations staff needed answers buried in a large, scattered library of SOPs and policy documents.

What I did. I was handed a blank project connected to a search index, and I built everything else: the chatbot, a reporting view for leadership, and testing with one person from each team before it went wider. Then it launched in beta to the operations teams.

What went wrong. Once real people used it, an answer that was right for one team could come back confidently wrong for another. I traced it to one shared index mixing content across countries and audiences. I tuned retrieval until it was usable, then designed the next version, where a coordinating agent sends each question to an agent that only knows that team's documents.

What changed. Operations staff had one place to ask instead of digging through folders, and it stayed in use after I left.

What I'd do differently. I'd sort out how documents are labeled and who can see them first. Retrieval comes second, and the prompt comes last.