The useful boundary is not a permanent list of forbidden tasks. It is a set of conditions where accountable judgement matters more than autonomous execution. When the cost of being wrong is material, the action is difficult to reverse or the context is sensitive, automation should support a named decision-maker rather than quietly replace one.
Material commercial decisions need accountable ownership
Pricing exceptions, strategic qualification, negotiation positions, large commitments and decisions that materially change a customer or supplier relationship often depend on context that cannot be reduced safely to one automated rule.
Software can prepare evidence, surface options and enforce process, but a named person should own the consequential decision when the business would still expect someone to explain it if the system were wrong.
Sensitive communication should not lose human judgement
Conflict, complaints, emotionally charged customer situations and messages with material reputational consequence can be assisted by drafting or context retrieval without being fully delegated.
A fast response is not automatically a good response. Human review matters where tone, history, exception handling or relationship judgement changes what should happen.
Do not automate a process the business has not resolved
If competent people still disagree about the normal workflow, the states, the owner or the expected outcome, automation will encode one accidental version of an unresolved process.
Clarify the operating model first. Automating ambiguity makes the problem harder to see because the system can execute inconsistency at scale.
Irreversible or high-consequence actions need stronger gates
Deletion, financial commitment, account changes, public publication, contractual actions and other difficult-to-reverse steps should have authority proportional to failure cost.
The right design may still include automated preparation and deterministic checks, but high-consequence execution should stop at an explicit approval gate unless strong evidence justifies more autonomy.
Weak source data is an authority problem
Automation that acts on incomplete, duplicated or poorly governed source data can be internally consistent and still be wrong. The system should not receive more authority than the information feeding it deserves.
Fix system-of-record quality, access, provenance and exception visibility before turning unreliable inputs into automatic actions.
Human review must be meaningful, not ceremonial
A nominal approval step adds little if the reviewer cannot see the source context, the proposed action, the reason for escalation and what will happen after approval.
Design the human gate so intervention is informed and practical. The purpose is accountable judgement, not an extra click that everyone learns to approve automatically.
Automate repeatable execution aggressively where the operating case is strong; preserve human ownership where consequence, ambiguity or accountability makes judgement part of the product.
Apply the logic to a real process: Score the automation opportunity →