A useful question to ask about LLMs in healthcare is not "what can they do" but "what can they do without crossing a regulatory line you have not budgeted for". The answer right now is: a lot, as long as the output is administrative rather than diagnostic. Summarising consultation notes, drafting referral letters, translating between clinical and patient-facing language, structuring unstructured intake — these are all valuable, low-risk applications that real teams are deploying today.
Where the ceiling sits
The moment an LLM output starts to influence a clinical decision, you are in medical-device territory under MDR. That does not make LLMs unusable for clinical workflows — it just means the integration has to be designed so that the human remains the decision-maker, with clear audit trails for everything the model surfaces.
Administrative drafting is broadly safe, with clinician approval before sending.
Patient-facing chat needs strict scoping and escalation paths.
Anything that classifies a vital, image, or symptom needs a regulatory pathway.
Anything that personalises treatment needs more than a pathway — it needs evidence.
“The interesting LLM work in healthcare is the work that makes a clinician faster, not the work that pretends to replace one.”
— DIETRICH CHARISIUS — CTO