Skip to course content
Free CrewAI course

CrewAI for Multi-Agent Automation

Unit 08.04: Knowing when to stop automating

The value curve of automating more cases flattens and then goes negative. Knowing where is part of the design.

Four slices, decreasing return

The routine majority, minor variations, real exceptions, and genuinely novel cases. Each needs more guardrail, more review, and covers fewer items.

The code lays out the four with a verdict on each.

STAGES = [
    ("the 60% that are routine",       "high",   "clear win"),
    ("the next 25%, minor variations", "medium", "worth it, more review"),
    ("the next 10%, real exceptions",  "low",    "review costs more than doing it"),
    ("the last 5%, genuinely novel",   "none",   "automating it creates work"),
]

print(f"{'slice':36} {'value':8} verdict")
for slice_, value, verdict in STAGES:
    print(f"{slice_:36} {value:8} {verdict}")

print("""
The curve flattens and then goes negative. Past a point, each additional case
handled by the crew needs a more careful reviewer, a more specific guardrail,
and a longer exception path -- all to cover fewer items.

Automate the routine slice, route the rest to a person deliberately, and write
down where you stopped and why. An automation with a stated boundary is
maintainable; one that quietly attempts everything is not.
""")

The last slice is where automating creates work: a more careful reviewer, a more specific guardrail, a longer exception path - all to cover five percent of volume, each item of which is different from the last.

The third slice is the judgement call, and it is where most projects overreach. Review costing more than doing the task manually is a clear signal, and it is measurable with the numbers from the previous unit.

The mistake this prevents

The mistake is treating remaining manual work as a defect in the automation. Routing the hard cases to a person deliberately, with a stated boundary, is a finished design. An automation that quietly attempts everything is one that fails unpredictably on the cases nobody chose to include.

Takeaway

Automate the routine slice and route the rest to a person on purpose. Write down where you stopped and why - a stated boundary is what makes the system maintainable.