- Some workflows will not autonomize regardless of capability, because the human presence is the service.
- Death notification and terminal-diagnosis communication are the clearest cases.
- Name the boundary explicitly so it holds as capability improves.
There is a class of clinical work that will not be automated at any level of model capability, and it is worth naming precisely because roadmaps keep implying otherwise. Telling a family someone has died. Delivering a terminal diagnosis. These are not hard problems awaiting better technology. They are the wall.
The workflows that will not move
The common feature is that the human presence is not the delivery mechanism for the information. It is the substance of what is being provided.
A family receiving news of a death is not primarily receiving data. They are receiving the fact that another person came to tell them, stayed, and answered what came next. Automate the information transfer and you have removed the service while completing the task.
The same holds for terminal diagnosis, for consent conversations about high-risk intervention, and for the discussion where a patient decides to stop treatment.
Why this is permanent and not a capability gap
Capability gaps close. This one cannot, because the requirement is not accuracy or fluency. It is accountability and presence — a person who is answerable and who is there.
A system that communicates perfectly still fails the requirement, and improving it makes no progress against the constraint. This is worth stating plainly because it is easy to mistake for a hard problem, and hard problems attract investment.
Recognising it as a boundary rather than a frontier saves that investment.
Some workflows will not autonomize regardless of model capability.
Designing a boundary that holds
Boundaries erode by increment. Nobody proposes automating a death notification. Someone proposes drafting the initial contact message, then handling the scheduling, then preparing the summary the family receives.
So write the boundary down, name the workflows, and make it a governance rule rather than a cultural assumption. Then put your effort into the substantial adjacent work: assembling the record so the clinician walks in prepared, handling the logistics, and carrying the follow-up.
That is a large amount of delegable work sitting right beside something you should never delegate.
What to do next
- 01Name the non-automatable workflows explicitly in your AI governance policy.
- 02Make the boundary a rule rather than an assumption, so it survives incremental pressure.
- 03Delegate the preparation and follow-up around those conversations, which is where the real capacity is.