I think the biggest gap in the AI support conversation is what happens after the demo. Resolution rates look great on a slide, but production support is really about failure handling. The strongest stacks I’ve seen tend to have a clear knowledge source, well-defined escalation rules, monitoring for bad responses, and a feedback loop for improving the agent. AI shouldn’t just know when to answer it needs to know when not to answer. I’d also argue that knowledge base maintenance is one of the most underestimated parts. Even a strong agent becomes unreliable when the underlying information is outdated. The interesting metric for me isn’t just “% automated.” It’s how reliably the system handles the edge cases without creating more work for the human team.