We are adding AI to our workflows, our IDEs, our incident response, and sometimes even our production investigation process. But AI does not see production truth. It sees the signals we give it: logs, metrics, traces, code, dashboards, and runbooks. In this talk, I’ll share a real production story where a feature looked available, the deployment was green, and logs existed — but the feature delivered no value. When we gave the available signals to AI, it produced a confident but wrong direction. We’ll use this story to rethink observability for the AI era: not as “more logs,” but as a way to reduce ambiguity, expose hidden domain context, and help both humans and AI tools reach better hypotheses faster.
Software Engineer