Long View

10-80-10 began as a rule for leaders, not for software. John Maxwell wrote it down in 2007 and again in 2014: spend the first ten per cent casting the vision and lining up the resources, delegate the middle eighty, and come back for the last ten to put the finish on it. Dan Martell repeated it for founders in 2023. It held because it answers the question every leader actually has — how much of this should I do myself — with a number.

AI made the number urgent. The eighty is no longer a team you have to hire; it is analysis, processing and prototyping available to anyone, at any hour, and now running as agents that write code, work a browser and carry a plan for hours. When the middle costs almost nothing, the only question left is what you put in front of it and what you do with what comes back. That is the ratio, and it is why the formula puts you in the exponent: AI to the power of human.

The first ten is instinct and context. Most of what you know about your field is not written down anywhere. It is in the customer-service call, the sales conversation, the regulator's aside, the colleague who says “that will not work here.” No model has seen it. It enters the build through your brief and your stop conditions, and it is the single thing that makes your eighty different from anyone else's.

The last ten is judgement and signature. AI's default is the average of everything it has read. Left alone, its output converges: measurably less varied across writers, measurably stripped of the small markers — contractions, first person, the specific noun — that readers use to recognise a person. The last ten puts them back and puts a name on the result. It is where distinctiveness is restored, discoverability is earned, and proof is attached so the work can travel.

The evidence is recent and it is consistent. When people use AI to write, their time moves to the ends on its own: drafting halves, editing doubles (Noy and Zhang, Science, 2023). Specific, task-level correction improves output; global end-of-process judgement often does not, and in a third of cases makes it worse (Kluger and DeNisi, 607 effect sizes). People who stay in the loop outperform those who hand off cleanly (Dell'Acqua et al., the BCG field experiment, 2023). A small, bounded adjustment at the end beats an open-ended review, both for adoption and for accuracy (Dietvorst, Simmons and Massey, 2018). And the combination of a person and AI beats the better of the two alone on creation tasks — building — while losing on classification and decision tasks (Vaccaro, Almaatouq and Malone, Nature Human Behaviour, 2024). The ratio is not a slogan. It is the shape the data takes.

The first failure is doing too much yourself. Most people, given AI, still draft first and ask second, and then use AI as a spell-checker on their own work. They get a faster version of what they already had.

The second is doing too little at the end. In the largest writing experiment to date, 68% of participants submitted AI's draft unedited. Confidence in AI predicts less critical effort. Experienced builders reviewing AI's work on their own code were slower than without it and believed they were faster (METR, 2025). And a finishing pass recovers less of your voice than it feels like: people who edited AI drafts of their own vows and eulogies moved the text toward themselves, yet it stayed far closer to AI's, and they could not tell (Baumler et al., 2026). This is why the last ten is a checklist with five tests and a reader who knows you, not a feeling that it sounds right.

Every correction in the last ten is expertise AI did not have. The builders who keep their corrections — as context the next run starts from, as standing rules, as the sequence of “no, like this” — have a loop that improves; the ones who discard them start every build from zero. The frontier of model training now rhymes with this: supervised reinforcement learning teaches a model step by step from expert trajectories rather than from final answers. Your tens are the trajectory.

Build fast and you will need to know whether what you built is any good. That is R2: whether it is seen and believed, and the growth that produces. Build on what your field actually knows and you will need the sources that hold it: that is Q2, quantitative and qualitative, and why the asked and observed sources matter most. Build something distinctive and you will need to place it where people and models look, through the channels that decide: that is POET. Build it to be read and you will need it provocative, persuasive and positioned, with proof: that is P3. Build to reach the few who move the many: that is the Per Mille Effect. 10-80-10 is the first paper because it is how each of the others gets built.

Human involvement earns trust when it is known and true — and it is now also law and infrastructure: disclosure rules for AI-generated content are enforceable, content credentials travel with files, and search systems reward demonstrated expertise. The provenance line in the last ten is your side of that bargain. Disclosure costs a little trust; being exposed costs more.