Current AI customer support is genuinely good at one specific job: answering questions where the answer exists in your knowledge base. Stock questions about pricing, features, how-to guides, return policies, integration setups — all near-perfect AI workloads. A well-configured AI agent handles 60-70% of incoming conversations end-to-end, and the customer often doesn't realize they were talking to AI.
But three categories of work don't move to AI well:
1. Complaints and emotional conversations. When a customer is upset, they want acknowledgment and resolution from a human. AI can collect the context but the conversation itself benefits from a person.
2. Decisions with judgment and authority. Refunds, exchanges, contract changes, escalations — anything where the answer depends on context the AI can't reliably weigh ('they've been a customer for 3 years, give them the refund').
3. Genuinely novel troubleshooting. When the answer isn't in any documentation, when a human has to combine 5 pieces of evidence to diagnose — that's still a human job.
What changes with AI: support team headcount stops growing linearly with conversation volume. Agents spend more time on interesting work (the 30-40%) and less on repetitive work (the 60-70%). Job satisfaction tends to *increase* — fewer 'where do I cancel' tickets, more genuinely useful problem-solving.