A useful AI efficiency assessment should be able to say “no”. The easiest way to make AI look impressive is to assume that every task can be automated and every generated output is a saving. In real work, that is not true.
“Does using AI here create more value than complexity?”
Seven situations where restraint matters
- 1. Verification-heavy work — generation is fast, but qualified checking is expensive.
- 2. Poorly defined work — the real task is resolving strategic uncertainty, not automating its wording.
- 3. Rare work with high setup cost — a once-a-year task may be cheaper to do manually with AI assistance.
- 4. Work where relationships are the product — negotiation, sensitive leadership or high-trust advisory moments.
- 5. Automation that creates more administration than it removes.
- 6. Unstable or inaccessible inputs — incomplete, scattered, legally restricted or rapidly changing data.
- 7. Decisions where accountability cannot be delegated.
If AI produces an answer in one minute but a qualified person needs twenty minutes to verify every statement, the gross generation speed is misleading. The same is true when the objective, inputs and success criteria cannot be made explicit.
A task that occurs once a year may not justify an elaborate workflow. Sometimes doing it manually with AI assistance is economically better than automating it. And in negotiations, difficult leadership conversations or high-trust advisory moments, replacing the human interaction itself can destroy value.
AI can prepare evidence, compare options, identify inconsistencies and simulate outcomes. But there are decisions where a person still needs to own the choice. That is not inefficiency. It is governance.
The goal is not to maximize the share of work touched by AI
The goal is to maximize net leverage. Some work should be fully automated. Some should be AI-assisted. Some should remain human but be better prepared by AI. And some should simply be left alone.
A good AI strategy is partly a strategy of restraint.