Over the past year, I’ve been rebuilding parts of an incident response stack for a client, and the biggest surprise wasn’t the AI features themselves. It was how much of the underlying workflow had to change to make those features useful. You can’t just bolt an LLM onto a 2015-era ticketing tool and call it AIOps.