Skip to content
scsiwygest. ‘26
Sign in
get startedmcpcommunityapiplaygroundswaggersign insign up
8am AI·The ninety-five percent9 Jul 2025David Olsson
8am AI

The ninety-five percent

#8am-ai#context-engineering#deep-dive#judgment

David OlssonDavid Olsson

The same figure shows up across the corpus, in different meetings, about different work. The model gets you most of the way. The last stretch is yours. The group rarely means it literally — sometimes it's ninety percent, sometimes ninety-nine — but the shape is constant, and it's one of the most durable claims in two years of record.

AI gets you ninety-five percent. The last five percent is human. Everything important is in the five.


where it came from

The clearest version is Scott on his investor table. The model builds the structure, fills the cells, gets him a working draft fast. Then he spends real effort on the last piece — the judgment about which numbers are actually right, what to leave out, where the model confidently produced something plausible and wrong. The first ninety-five percent took minutes. The last five took the expertise.

The same ratio recurs elsewhere. A legacy codebase scaffolded in an afternoon, with the integration and the gotchas left as the hard part. A document drafted instantly, with the domain-specific corrections being the whole value. A demo that works at ninety percent and a business that needs the missing ten before it can be sold.

Different rooms, same proportion. The bulk is cheap now. The remainder is the work.


why the five is the expensive part

The trap is reading ninety-five percent done as almost finished. It isn't, because the percentages aren't measuring effort. They're measuring the wrong thing — volume of output, not difficulty of the part that's left.

The first ninety-five is the part that was always mechanical. Structure, boilerplate, the first competent pass. That's exactly what the model is good at, which is why it's now free. The last five is the part that needed someone who understood the domain — to catch the confident error, to make the call the model can't, to know which of two plausible answers is the true one. That part didn't get cheaper. It got more concentrated.

So the model didn't remove ninety-five percent of the work. It removed ninety-five percent of the visible work and left all of the hard work, undiluted. The bottleneck didn't shrink. It moved to the end and got denser.


the corollary nobody likes

If the last five percent is judgment, the last five percent is also the part you can only do if you built the muscle the slow way. You can't supply the missing five on a domain you don't know. And the model's fluency at the first ninety-five is exactly what tempts a person to skip the apprenticeship that would have taught them the five.

That's the education thread's fear, stated as arithmetic. A generation that lets the model do the ninety-five it can see, and never learns it was the five it couldn't.

The number is good news for anyone who has the judgment and a warning for anyone counting on never needing it. The model made the easy part instant. It made the hard part the only part that's left.

Share
𝕏 Post