The short versionAsk what the model was given and who checks what it returns before asking which model it is. The second question has a fashionable answer; the first one predicts the outcome.
A question that sounds technical and is not
“Which model should we use” is a procurement question dressed as an engineering one. Models change every few months and are largely interchangeable for the work a documentation product does. What does not change every few months is whether your terminology is consistent between departments, whether care-team roles are modelled accurately enough to constrain what the software can see, and whether anyone can reconstruct who approved a note six months later.
Those three things decide whether the output is usable. A model given fragmented, duplicated or ambiguous source information produces fluent output that inherits every one of those problems — and the fluency is the part that makes it hard to notice.
Fluent output from a poor record is worse than obviously bad output, because it takes longer to catch.
What governance means concretely
Not a policy document. Four properties of the running system, each of which can be checked by looking at it.
- Shared terminology, so a summary means the same thing when it moves between departments.
- Role-aware access, so the software cannot read more of a record than the person operating it can.
- An audit trail holding who reviewed, what they changed and what they approved — the sequence, not only the result.
- Correction capture, so a clinician fixing the same thing twice becomes a signal rather than a recurring annoyance.
The part that becomes measurable
Governance of this kind has a side effect worth more than the compliance story: it makes adoption measurable with something other than seat counts. Review time, correction patterns, and which templates get amended most are all observable once corrections are captured — and all three speak to whether the thing is actually working.
Usage tells you people opened it. Correction patterns tell you whether it was any good.
What none of it does is make the model right. Governance constrains what the software is given and records what was done with what it returned; it does not check the clinical content of a summary, and nothing here reduces the clinician's obligation to read one before approving it.
Written about Clinical HealthAI
