Claim
AI ROI is throttled not by AI capability but by three organizational forces: a skill gap limiting productive users to roughly 5%, workflow defense by technical workers protecting existing processes, and salary incentives that reward meeting existing output targets rather than multiplying them.
Mechanism
Three forces compound. The skill gap is real: most knowledge workers lack the judgment to use AI for production-quality output. They cannot tell when AI output is wrong, cannot subtract unnecessary features, and cannot make the iterative corrections that make AI useful. Workflow defense hits hardest in technical roles. Engineers have high-status, well-defined workflows. Switching is socially costly. Non-technical workers such as PMs, designers, and artists have "nothing to defend" and tend to adopt AI at higher rates with higher productivity gains. The incentive ceiling is the most structural throttle. A salaried employee who does existing work in 10% of the time does not produce 10x as much. They deliver the expected output and stop. Solo founders and small incentivized teams show the actual output ceiling because they are paid for output volume, not presence.
Conditions
Holds when: workers are salaried and measured by delivery of defined outputs rather than output volume. Applies across most knowledge work organizations.
Fails when: incentives are explicitly tied to output quantity and teams have autonomy to act on productivity gains. Also fails when an organization restructures workflows around AI-generated output capacity rather than bolting AI onto existing ones.
Evidence
Kim draws on patterns across gaming studio teams and comparisons with solo founders. He notes that engineers reviewing AI output manually cause the AI to "run at human speed" rather than at AI speed. Non-technical employees who have "nothing to defend" consistently show higher adoption rates.
"Most people aren't incentivized to get more output. Look at incentivized teams — small teams, solo founders — they're showing how crazy good the output can get."
Signals
- Teams adopt AI for speedup but report the same annual output volume after a full year.
- Technical workers have longer adoption curves than non-technical workers in the same organization.
- Highest AI output gains cluster in roles without existing high-status tool chains.
Counter-evidence
Kim's framework is pattern-observed rather than rigorously measured. The 5% figure for production-quality AI users is asserted, not derived from a study. Organizations can and do restructure incentives around AI-enabled capacity. Companies like Shopify explicitly restructure headcount decisions around AI productivity, changing the incentive model entirely.
Cross-references
- Agentic systems need a facilitator layer that routes decisions back to humans in four specific scenarios: Ethan Mollick on how the human-in-the-loop role is structurally preserved even as AI automates sub-tasks, complementing Kim's observation about salaried ceilings.