What 'human-in-the-loop' should actually mean
It's not a confirmation dialog. It's that no drafted content reaches a patient, a payer, or a chart without an authenticated staff member explicitly approving it — and that approval is logged with timestamp, actor, and reasoning.
What automation does not do
It does not finalize diagnoses, adjust dosages, screen for self-harm risk, send messages, submit packets, or close cases. Those decisions stay with named clinical staff. Publishing this list openly matters more than hiding it in a settings menu.
Source-linked everywhere
Every draft shows where it came from — chart entry, fax, lab result, payer policy. Reviewers shouldn't have to take anything on faith. If a suggestion can't cite its source, it shouldn't ship as a suggestion.
Reject + Reason as a first-class action
Rejecting a draft should be one click, and the reason should feed model behavior tuning per clinic. The rejection signal is more valuable than the approval signal — it's how the system learns your team's actual standards.
An audit trail you can read in plain language
Who drafted, who reviewed, who approved, who edited, who submitted. Each step with a timestamp and the version of the content at that step. Clinics that get audited shouldn't have to translate cryptic logs into a story.