Locale: en
While a resale specialist spends about 30 minutes inspecting physical items one by one, someone else is lying in bed, speaking into a phone, and keeping an article factory moving. If one instruction can permanently affect roughly 1,500 future articles, the unit of work changes from “produce one article” to “rewrite the production rules.”
One day, a resale specialist was outside checking unwanted items to see what could still be sold. Inspect the condition. Touch the object. Check the model, scratches, and wear. Compare it with the market. Physical work takes time because reality refuses to become a spreadsheet just because we would prefer it to.
Inside, a much stranger scene was happening. A person was lying in bed and apparently doing almost nothing except speaking into a smartphone. Yet those spoken decisions triggered research, drafting, revisions, and quality checks across a digital article-production system.
Visually: doing nothing.
Operationally: changing the manufacturing conditions.
Doing nothing (while actually doing something)
With AI-enabled workflows, “I barely did anything” and “a lot of work got done” can both be true.
When human effort is the main production input, that sounds contradictory. But once research, drafting, transformation, and checking can be delegated, the human role shifts from performing every operation to controlling the system: deciding what matters, defining rules, and handling exceptions.
Commute: roll over. Meeting room: phone microphone. Factory floor: AI.
This is not zero work. The operator still decides what to make, what to change, and which rules deserve to become permanent. What changes is the relationship between physical effort and output volume.
The headquarters is a bed, and the meeting room is voice input
Voice input sounds trivial, but it removes a surprising amount of friction.
Sit at a desk. Open a computer. Find the right file. Convert a thought into polished typed instructions. None of those steps is enormous, but repeated hundreds of times they become startup costs.
With voice, a thought can become an instruction almost immediately: “fix this,” “add this perspective,” “use this criterion from now on.” The important point is not that work can be done while lying down. It is that the distance from thought to execution becomes shorter.
The same 30 minutes can have different density: physical work is often serial, digital work can be parallel
A physical inspection does not take time because the inspector is inefficient. There is an irreducible step: somebody has to look at the actual object.
Digital work is different. Search, copying, transformation, comparison, and many checks can be duplicated or run in parallel. One human decision can fan out into multiple processes.
So two people can both spend 30 minutes while producing very different amounts of downstream activity.
That is not a ranking of the people. It is a difference in the medium of work. Physical inspection deals with atoms. An article factory deals mostly with bits. Atoms are famously unimpressed by a well-timed roll in bed.
Real scale comes from changing the rules that make articles
Having AI write one article is useful, but structurally it resembles faster outsourcing.
The bigger shift occurs when an instruction becomes persistent: “use this quality criterion in future articles,” “reject this structural defect,” or “replace this confusing pattern whenever it appears.”
Then a decision no longer ends with one document. It can affect articles written weeks or months later.
If the system eventually produces about 1,500 affected articles, the value of a five-minute decision cannot be measured by those five minutes alone. It must be measured by the radius of propagation.
That is leverage.
When one instruction reaches 1,500 articles, the bottleneck changes
In a manual-worker model, the bottleneck is often the number of hours the person can work.
In a control-layer model, the bottleneck becomes judgment: Which rule should be permanent? Is it correct? What exceptions will it break?
Higher productivity therefore does not merely make work easier. It raises the cost of a bad decision.
If a good rule improves 1,500 articles, a bad rule can distribute damage to the same 1,500 articles with admirable consistency. The real risk is not that the factory manager is in bed. It is that a sleepy rule gets deployed to every production line.
Research supports “AI can speed up specific tasks,” not “one person beats a company”
There is empirical support for part of this productivity intuition.
In a 2023 preregistered experiment published in Science, Noy and Zhang assigned occupation-specific writing tasks to 453 college-educated professionals. The group using ChatGPT completed the tasks 40% faster on average and received quality scores 18% higher.
In field research by Brynjolfsson, Li, and Raymond on customer-support workers, access to a generative-AI assistant increased productivity, measured as issues resolved per hour, by 14% on average and by 34% for novice and lower-skilled workers.
Those findings do not prove that an AI-enabled individual is generally more productive than a company. Organizations also provide sales, legal functions, accounting, customer support, redundancy, continuity, coordination, and failure recovery.
The evidence supports a narrower claim: for some well-matched tasks, generative AI can improve speed and quality. The distinctive leverage here comes from adding persistent rules that propagate into future outputs.
If Google and research are the source, treat search as discovery—not proof
Using research and external evidence as the basis for permanent rules is stronger than relying on intuition alone.
But a search result is not itself evidence. Search is a discovery layer. For research, go to the paper. For regulation, go to the official document. For a product specification, go to the manufacturer.
There is another failure mode: the source can be correct while the implementation is wrong. A principle that worked on average in a study can be misapplied to every article. An SEO convention can be mistaken for a universal rule of good writing.
A useful operating sentence is:
Trust the source, verify the implementation.
Good rules scale to 1,500 articles. Bad rules do too
In a scalable system, improvements and mistakes travel on the same highway.
The minimum safety mechanism is straightforward: test on a small representative sample, run quality checks, inspect real outputs after rollout, and preserve versions so the change can be reversed.
This is not caution for its own sake. These are guardrails that allow the system to stay fast.
Checking every article manually would erase much of the leverage. The goal is to move human attention to the points that matter: the rule-change entrance, representative outputs, and the rollback path when something goes wrong.
It looks like being a CEO, but the more precise role is the control layer
Lying down while people and AI systems execute tasks certainly has CEO aesthetics.
Structurally, though, “control layer” is more accurate.
The physical layer handles physical inspection. The information layer handles research, synthesis, writing, and transformation. The control layer sets goals, priorities, rules, and exceptions.
Once these layers are separated cleanly, the operator no longer needs to perform every task personally.
A system can change its behavior while the human appears to be doing almost nothing.
Conclusion: the person is lying down, but the production conditions are being updated
The funniest part is the contradiction between appearance and effect: someone looks spectacularly inactive while modifying rules that can govern months of future production.
Scaling work no longer has to mean scaling working hours.
The more important questions become: How many times can one decision be reused? How far does it propagate automatically? Can evidence-based rules spread through the system without spreading their mistakes just as efficiently?
At that point, productivity is no longer only “how many articles were written in one hour.” It becomes “how many future outputs improved because of one good decision.”
The headquarters can be a bed.
Just do not deploy sleepy permanent rules to production.
References (3)
- Shakked Noy, Whitney Zhang, “Experimental evidence on the productivity effects of generative artificial intelligence,” Science 381(6654), 2023. https://www.science.org/doi/10.1126/science.adh2586
- Erik Brynjolfsson, Danielle Li, Lindsey R. Raymond, “Generative AI at Work,” NBER Working Paper 31161, 2023; later published in The Quarterly Journal of Economics 140(2), 2025. https://www.nber.org/papers/w31161
- Published QJE version: https://doi.org/10.1093/qje/qjae044
