Two people open the same model on the same morning. Same subscription, same context window, same weights. One of them saves twenty minutes. The other ships a system that saves twenty minutes every day for the next year, for everyone on their team. Nothing about the tool explains the difference.
- Two people with the same model get very different results, and controlled studies show it.
- The gap sits in four places: defaults, one-shot outputs, amnesia and blindness.
- The fix is cheap: anything you do twice becomes a file.
The gap is not access. Access was solved. It is not talent either, which is the comfortable explanation and the wrong one. It is that most people are using roughly one percent of what is sitting in front of them, and paying for the other ninety-nine in a currency that never shows up on an invoice.
That is the 1% Tax. It is what you owe for treating a medium as a tool, and it is the practical consequence of the distinction I drew between author and user.
The same tools
Identical inputs. Divergent outcomes. Every time.
This is historically strange. For most of the history of work, the people with the best outcomes had better instruments, better labs, better capital, better distribution. The instrument was the moat.
That has changed. The frontier model a solo operator can open in their kitchen is the same one a company with ten thousand engineers opens. The instrument has been equalised and the outcomes have not converged at all. They have spread further apart.
When the tool stops being the constraint, the constraint becomes what you do with it.
Which means the difference has moved somewhere less comfortable than budget. It has moved into method.
What the research shows.
The gap shows up in controlled studies. In 2023, researchers ran an experiment with 758 consultants at Boston Consulting Group. On tasks inside what GPT-4 does well, consultants using it finished 12.2 percent more tasks, 25.1 percent faster, with results rated more than 40 percent higher in quality. On a task set deliberately outside its strengths, those using it were less likely to reach the right answer than consultants working without it. The authors call this the jagged frontier: the tool is excellent in some places and unreliable in others, and the boundary is not obvious from outside.1
A separate study followed 5,179 customer support agents after an AI assistant was introduced. Productivity, measured as issues resolved per hour, rose 14 percent on average and 34 percent for the least experienced agents, because the tool passed on what the best agents already knew.2
Same kind of tool in both studies, very different results depending on who used it and where. That is the 1% Tax with data attached.
Where it goes
Four places the other ninety-nine percent disappears.
A chat box is one affordance out of many. If every interaction you have with a model happens in the shape the vendor shipped, your ceiling is the vendor's imagination, not yours.
A good answer is worth one use. A system that produces good answers is worth every future use. Most people generate the first and never notice they could have built the second.
You explain your context again on Monday. You re-derive the prompt that worked. Every correction you make evaporates instead of accumulating, so you start from the same place for years.
You cannot improve a practice you are not measuring. Most people's sense of how much they use AI is a feeling, and the feeling is always wrong in the same direction.
None of these are skill problems. They are structural, which is the good news, structural problems have structural fixes, and the fixes are cheap.
The two postures
What the tax looks like, and what not paying it looks like.
Paying the tax
- Opens a chat box
- Writes a prompt, gets a result, closes the tab
- Re-explains the same context next week
- Keeps the trick that worked
- Feels faster
Not paying it
- Builds the surface the work happens on
- Writes the file that produces the result repeatedly
- Puts the context in a file the next session reads
- Gives the system away so others compound it
- Measures, and knows
The right-hand column is not harder. It is roughly the same amount of effort, spent once, in a durable place. The entire difference is whether the effort lands in a chat log or in a directory.
The receipt
I measure my own practice, because the claim is worthless otherwise.
I track what I actually do with these tools, not how it feels. That is the cheapest of the four fixes and the one almost nobody makes, because measuring your own practice risks finding out it is thinner than you thought.
What it bought is not a number. It is a library of skills I can compose instead of re-derive, a set of agents that run real work, including the one that operates this store, end to end, and a voice that lives in a file rather than in my head, so it survives every new model release.
None of that came from a quarter of trying harder. It came from years of putting corrections into files instead of repeating them. The rule is unglamorous: anything I have done twice becomes a file. That is the whole discipline. Everything else is a consequence of it.
The fixes
One fix for each place.
Each of the four leaks has a cheap fix, and none of them needs a new tool.
Defaults: build one surface that is yours.
A folder, a short instruction file, and a command or shortcut that starts your most common job. The tool matters less than the fact that the starting conditions are yours instead of the vendor's.
One-shot: keep a template for anything you do twice.
When a prompt works, save it with the context it needs and the check you used to judge the result. The second use should cost a minute, not an hour.
Amnesia: put context in a file the next session reads.
Who you are, what you are working on, decisions already made, and what not to redo. When you correct a mistake, correct the file as well as the answer, so the correction is still there next week.
Blindness: keep a weekly tally.
Count the tasks you did with AI, how many produced something reusable, and how many you redid from scratch. After a month you will know which of the four leaks is yours.
A limit worth admitting.
None of this guarantees better output. The jagged-frontier study found that people using the tool did worse on a task outside its strengths. Measuring outcomes instead of activity keeps a good system honest: check results against something that does not come from the same tool.
The bill
Nobody sends you an invoice for this.
That is what makes the 1% Tax so easy to keep paying. There is no line item, no overdue notice, no moment where the cost becomes visible. You simply get a normal amount of value out of an extraordinary instrument, indefinitely, and it feels like using AI.
The tools are available to everyone. The authorship is not. That gap is the whole opportunity, and it widens every week you keep paying.
Sources
- Dell'Acqua, F., McFowland, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., and Lakhani, K. R. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Harvard Business School Technology and Operations Management Unit Working Paper No. 24-013.
- Brynjolfsson, E., Li, D., and Raymond, L. R. (2023). Generative AI at work. NBER Working Paper No. 31161. nber.org/papers/w31161
Writes The Visionary, field notes from the AI-native frontier. Building in public: practical builds, named methodology, receipts, no fluff.