Taking an LLM and putting it to work on real support cases looks simple… until it refunds an order that never existed, performs actions that can’t be undone, or drifts off task the longer the conversation runs. There’s a world of difference between an impressive demo and an agent that actually holds up in production.

In this talk I’ll walk through how we took an AI support agent to production at Groupon: the real problems (the ones demos never show), dynamic turn-by-turn control of context and tools, and why the key piece wasn’t a better prompt but deciding what the agent is allowed to do at each moment. I’ll also cover the other half of the problem — who builds and maintains these agents — and how to let non-technical people operate them without writing code.

An honest talk about what worked, what broke, and what we learned along the way.