Question

Why does AI get you 80% of the way there?

Because the part it does is the part already written down somewhere by somebody, and the rest exists only inside your business. The expensive part is that it cannot tell it stopped.

The line is not difficulty. It is whether anyone wrote it down

The usual explanation is that AI handles the easy eighty percent and the hard twenty is left. That is close enough to feel true and it points at the wrong thing, which matters because it suggests the fix is a smarter model.

Sort the work differently. Not by hard and easy, but by written and unwritten.

Nothing in the right column is harder than the left. Most of it is trivially easy for you. It is simply not anywhere. It never got recorded, because it was never the sort of thing that gets recorded.

So the eighty is not a competence ceiling. It is the edge of the written world.

The part that makes it cost you

Here is the mechanism that turns a limitation into a recurring bill.

The system does not hand you eighty percent of a job and say so. It hands you what it takes to be a complete job. From where it sits, it is complete: it has produced everything its available information supports, and nothing in that information indicates a standard existed and was not supplied.

Which is why the cost always lands on you and never shows up as a line in any tool. Someone has to read the output holding the missing standard in their head, notice the collision, and fix it. That someone is you, on every job, and it is the reason a technology that saves hours can still leave you feeling like the bottleneck.

It also explains the specific texture of the irritation. The work is not bad. If it were bad you would simply reject it. It is good, and wrong in a way you cannot state until you are looking at it, which is the most attention-expensive kind of wrong there is.

Better models move the split without moving you

The reasonable hope is that this shrinks with each model generation. It does shrink, and that is not the relief it sounds like.

A stronger model expands the written world. More of the job comes back done. The remainder gets smaller.

But the remainder is defined as the part that is not written down anywhere, so shrinking it does not dilute it. It concentrates it. What is left after a better model has done its work is more purely you than what was left before, because everything that was merely uncommon has now been absorbed and only the genuinely particular survives.

So the recurring cost does not trend to zero with better tooling. It distils, and the share of your day spent doing things only you can do goes up rather than down. Whether that is good news depends entirely on whether any of it is being kept.

What would change the picture

Not effort at the correction step. You are already good at that step, and it is not where the loss is.

The loss is that the step evaporates. Everything else in the job has a destination. The output is filed. The prompt gets reused. The conversation is scrollable. The judgment applied while correcting has no address, so the next job of the same kind starts from the same place and asks for the same correction again.

The change is that the correction gets kept with its reason, general enough that it applies past the one case. Then the next job starts closer to your line than the last one did. You still correct. You correct further along, and the distance narrows over months instead of resetting every session.

Frequently asked

Where this sits

What happens to the last part after you supply it is at the Final Twenty. The category is defined at what Imprinted AI is.

Related

• The Final Twenty. Where that last part goes after you supply it, which is nowhere.
• Why does my AI output feel generic? The same gap, felt rather than measured.
• Imprinted AI vs RAG. Why retrieval is weakest exactly where you are most needed.
• Articulation density. What changes the ratio, and what does not.

Last updated: 28 July 2026