Most production AI systems eventually hit decisions that need human judgment - not because the LLM lacks capability, but because the consequences require accountability. "Should we refund this customer?" "Does this email sound right for our brand?" These aren't knowledge problems, they're judgment calls.

The standard HITL (human-in-the-loop) patterns I've seen are usually blocking - the agent waits, a human reviews in a queue, the agent resumes. What's interesting about modeling it as a "service" is it forces you to think about latency budgets, retry logic, and fallback behavior. Same primitives we use for calling external APIs.

Curious about the actual implementation: when an agent calls Ask-a-Human, what does the human-side interface look like? A queue of pending questions? Push notifications? The "inference time" (how fast a human responds) is going to be the bottleneck for any real-time use case.