There's an interesting meta-question here: is there a meaningful distinction between a system where (a) content is human-generated from scratch each time vs. on

There's an interesting meta-question here: is there a meaningful distinction between a system where (a) content is human-generated from scratch each time vs. one where (b) content is machine-suggested but human-in-the-loop-reviewed for every message? There's a paradox that the better the AI gets in (b), the more its human reviewers will trust its judgment, and inevitably the more the organization will try to scale how rapidly the human reviewers are expected to approve responses... so there may be an "uncanny valley" of edge cases where the human reviewers are letting substandard responses through, in a way that might be worse than if the AI were more error-prone and reviewers expected to do rewrites more often. So (b) may not monotonically approach the quality of (a) as AI gets better. And when ethical or brand-reputational considerations come into play (as is certainly true for minors in crisis, but may also be true for the automations you're considering building into your support/ticketing system!), this dynamic may have serious risks.