Glossary
The Final Twenty
AI does the first eighty percent in seconds. You do the rest out of judgment. Then the session ends and you do it again on the next job.
The Final Twenty (noun). The portion of a piece of work AI cannot complete, which the operator supplies from judgment, and which is lost when the session closes. Named for the fact that it gets paid repeatedly across jobs rather than once.
The sequence
Almost everyone using AI seriously has run this loop, most of them today.
- You ask for something. The first eighty percent arrives fast, and it is genuinely good.
- It is not right. Not wrong exactly. It misses in ways that are hard to state until you are looking at them.
- You correct it. This part takes real attention, because you are applying standards you have never written down.
- The result is now good. You ship it.
- The session closes.
- The next job begins. The same kind of miss happens. You correct it again.
Nothing in that sequence is broken. Every step works as designed. And the loop runs indefinitely, because the only thing that would break it, the judgment applied at step three, is the one thing that goes nowhere.
Everything in the workflow has storage except the part that was yours
This is the observation worth carrying.
| Produced during the job | Where it goes |
|---|---|
| The finished output | Saved. Filed, shipped, in the drive |
| The prompt | Saved, often reused |
| The conversation | Saved. Scrollable, searchable |
| The source material | Saved. It was already a file |
| The judgment in your correction | Nowhere |
The transcript looks like a counterexample and is not. You can scroll back and read what you typed. What is not in there is the standard the correction came from, stated generally enough to apply to a different case next week. You wrote make the second paragraph less certain. The rule behind it, something like never let the promise get ahead of the proof, was in your head and stayed there.
So the artifact of the most valuable minute of the job is a line of text that only makes sense next to the thing it was correcting.
Why the split matters more than the numbers
The eighty and the twenty are a shape, not a measurement, and the ratio changes by task. The asymmetry underneath it is stable.
The part that arrives quickly is the part that is well represented in what the model learned, which is another way of saying it is the part many competent people would have produced. It is the commodity end of the job, and getting it in seconds instead of an afternoon is a real gain that everyone now has.
The part left over is the part where being specifically you changes the answer. Your standard, your context, the thing you know about this client, the reason the obvious move is wrong here.
Which produces an uncomfortable read on model upgrades. A better model moves the split. It does not move which side of it your judgment sits on. The commodity portion expands, the remainder gets smaller and more concentrated, and what is left is more purely you than it was before. So the recurring cost does not shrink toward zero with better tooling. It distills.
What paying once looks like
The fix is not more effort at step three. It is that step three stops evaporating.
A correction captured with its reason, where the reason generalises past the one case, means the next job of that kind starts nearer your standard than the last one did. You still correct. You correct something further along, and the distance you close by hand narrows over months instead of resetting every session.
That is the entire difference between a tool that saves you hours and a system that accumulates. One returns time. The other turns the calls behind your best work into something the business can use when you are not there.
Frequently asked
What is the Final Twenty?
The part of a job AI cannot finish, which you complete out of judgment. AI gets the first eighty percent in seconds. The remaining portion is where your standards, your context and your call get applied, and it is the only part of the work that was ever specifically yours.
Why does it get paid twice?
Because the correction has nowhere to live. The output is saved. The file is saved. The conversation is saved. The judgment you applied while correcting is not stored anywhere, so the next job starts from the same place, makes the same kind of mistake, and you supply the same judgment again.
Doesn't a better prompt fix this?
It fixes the part you could anticipate, which is the part you already knew how to state. The Final Twenty is mostly made of corrections you could not have written in advance, because you did not know they were needed until the specific case was in front of you. A prompt library grows to cover the predictable end and hits a ceiling at exactly the point the work gets interesting.
Is the eighty-twenty split literal?
No, it is the shape rather than a measurement, and the ratio moves by task. What holds is the asymmetry: the portion AI produces quickly is the portion many people could have produced, and the portion left over is where being specifically you changes the answer. Better models move the split without changing which side of it your judgment sits on.
What would it look like to only pay it once?
The correction gets captured with its reason, and the reason is general enough to apply beyond the one case. Then the next job of the same kind starts closer to your standard than the last one did, and the gap you keep closing by hand narrows instead of resetting.
Where this sits
Where a correction goes when it is not discarded is described at Imprint. The reason it could not have been written down in advance is at tacit knowledge in business. The category is defined at What is Imprinted AI.
Related terms
- The Assembler. The position that pays this cost indefinitely without registering it.
- Imprint. Where a correction goes when it is not discarded.
- Tacit knowledge in business. Why the correction could not have been written down in advance.
- Articulation gap. A way to see, before the work runs, how much of this you are about to pay.
- Imprinted AI vs AI memory. Why saving the conversation does not save the judgment in it.
Where the term comes from
This is one entry in the vocabulary of a longer argument. The full glossary has fifteen terms. The case they belong to runs about 23,000 words, it is free, and there is no email gate on it.
Read The A.I. Business Manifesto
Nothing on this page is for sale. Quote it, argue with it, or pass it on.
Last updated: 28 July 2026