The interesting question about automation is not what can be automated. It is what should stay human, and why.
Four patterns are worth leaving alone.
Decisions that are mostly judgment
If two experienced people could reasonably decide the same case differently, the process is judgment wearing a process costume. Automating it means encoding one opinion as law and hiding the debate that made it good.
Processes that change every time
Automation rewards repetition. A workflow that runs twice a quarter with different steps each time has nothing stable to encode. The drafting assistance helps; the pipeline does not.
Anything where the failure is invisible
If a wrong output would not be caught until it compounds, the automation needs telemetry before it needs autonomy. No visibility, no pipeline.
Work that is the relationship
Some emails, some reviews, some calls exist to maintain a human connection rather than to transfer information. Delegating them saves minutes and costs a currency that is harder to earn back. Keep the pen on those.
The right posture is not automation-first or human-first. It is ask-first: does this process get better when a machine runs it, or just faster? Faster is not always better, and saying so out loud protects the cases where it is.