Where should AI stop and human support begin?

AI support tools increasingly promise to resolve most customer questions without human involvement.

That works well for simple questions, but it becomes more complicated when a conversation involves billing, account access, an upset customer, or something the knowledge base does not cover clearly.

After years working in software support, I have started to think the best approach may be a combination:
AI handles repetitive questions automatically, drafts responses for uncertain cases, and sends sensitive or complex issues to a person.

I am currently building a support product around this model, but I am still trying to understand where founders draw the line.

For those of you using AI in customer support:

Which tickets are you comfortable letting AI resolve automatically?

Which tickets should always require human review?

Has AI improved the customer experience for you, or has it mostly reduced workload?