This is the module's anchoring competency. For each step in a workflow — not the workflow as a whole, and not a single task in isolation — the question is which bucket it belongs in:
| Bucket | Meaning |
|---|---|
| AI-appropriate | Claude can own this step end-to-end |
| Human-retained | This step needs a person, even if Claude assists elsewhere in the process |
| Collaborative | Claude and a person work this step together — neither owns it alone |
"When the wrong steps are automated, it can backfire." A workflow can have several steps that are genuinely well-suited to automation and still fail if even one fragile, ambiguous, or high-stakes step gets automated alongside them. Delegation judgment has to be applied per step, not assumed for the whole process just because most of it is a good fit.
Deliberate delegation is what "turns individual wins into workflow value." Getting a great result from Claude on an isolated task doesn't automatically compound into workflow-level value — that only happens when the delegation boundaries between steps were chosen on purpose, not defaulted into.
This connects directly to Module 3's stakes-based verification calibration — the same underlying judgment (match effort/oversight to actual stakes) is being applied one stage earlier: at workflow-design time, deciding which step gets delegated at all, rather than at output-review time, deciding how hard to check a given result.