Support agents that actually close tickets vs ones that just deflect, what's the difference?

​ Been digging into AI agents for customer support, and there's a big gap between agents that actually resolve tickets and ones that just deflect people until they ask for a human. From what I've seen, the ones that work have a few things in common: They can call real tools, like pulling order tracking, checking account status or processing a refund, instead of just answering from a knowledge base. They sit on top of the existing helpdesk instead of replacing it. Intercom Fin and Zendesk AI both do this, and Aissist too, which works across Zendesk, Gorgias, Intercom and Freshdesk. They have a clear escalation rule, and the human gets the full context so the customer doesn't repeat themselves. They start narrow, like WISMO (where's my order) and resets, and expand once resolution rates hold up. The handoff seems to be where most setups break. For anyone running agents in support, what's your actual end-to-end resolution rate, and what made the biggest difference?