Demo 11 - The Agent Was the Easy Part: Putting an AI Analyst in Front of a Whole Company
AirHelp is a flight-compensation company with a small data team and a wide variety of data-hungry stakeholders, all asking the same kind of question: "what was X last month, split by Y?" We put an AI analyst — AIDA — into Slack so the whole company could ask directly. The talk is about the part that isn't the agent. We deliberately did not build text-to-SQL. AIDA cannot write a query against the warehouse. It can only compose answers from metrics a human has already defined, tested and version-controlled in our semantic layer. That constraint is the product. Text-to-SQL demos work on stage and fail in production, because no warehouse is self-describing and the model does not know what your company means by "active customer". I'll cover the architecture, real examples of how it changed the way people get answers day to day, and the culture shift it forced — inside the data team, where the job moved from answering questions to defining what the answers mean, and outside it, where people had to learn what a governed answer is and when the system should say no. The takeaway for anyone leaving to build this on Monday: your agent is a function of your definitions. Spend the money there.