Why does AI give average results, and what fixes it?
Three ideas worth more than any prompt trick: own both ends of the work, stop reading confidence as correctness, and remember which world the training data came from.
AI tools produce average work delivered with total confidence, and people accept it because confidence reads as correctness. The fix is to own the ends: set the outcome yourself, let AI do the repetitive middle, then edit the result hard. Average output only ships when nobody checks the last ten percent.
Three ideas are doing most of the work in how I use AI right now. None of them are about tools or prompts. They are about where you put yourself in the process, and they are the difference between output that looks fine and output that is actually good.
The 10/80/10 rule
The rule comes from flying. A pilot is fully engaged for takeoff and for landing; most of the hours in between run on autopilot. AI projects work the same way. You own the first ten percent — the idea, the outcome you are actually trying to reach, what the market wants. AI handles the eighty in the middle, the repetitive build you have already proved you can do. Then you own the last ten percent: look at what came back and fix what is wrong.
Most people skip both ends. They type a half-formed prompt, read the output, decide it is good enough, and ship it. That is the whole gap. The ends are cheap to own and almost nobody owns them, which is exactly why owning them separates your work from everyone else's.
Be a thought partner, not a thought leader.
That line is from Jeff Woods' book The AI-Driven Leader, and it is where the first ten percent starts. You hold the idea. You know what the outcome should be. Then you work backwards into the project. We applied the same rule to a specific setup in how to set up Claude for your business; here it is the foundation the other two ideas sit on.
Average answers, delivered with total confidence
Large language models produce average output with enormous confidence, and that combination is the real problem. People are wired to read confidence as correctness. When an answer arrives with no hedging and no visible seams, the instinct is to accept it. Nothing in the tone tells you whether the substance underneath is strong or mediocre, because the delivery is identical either way.
Which is why "double-check the output" is correct but useless on its own. Checking only works if you already know what good looks like. That is what the first ten percent buys you. If you decided the outcome before you asked, you can read the result and say this part is wrong, this part misses the point, this part is fine. If you did not decide anything, you have no standard to measure against, and confident average work sails straight through.
Everything these models learned came from a pre-AI world
This is the one that is harder to shake. It came up at an event in Phoenix last week and I have been thinking about it since. Everything these models learned was written in a pre-AI world. The tactics, the strategies, the business principles, the assumptions about how long work takes and where leverage comes from — all of it was recorded before any of this existed.
Think about teaching someone to drive from material written in a horse-and-carriage world. It can still tell you real things about transportation, about wheels, about keeping the thing maintained. But roads, engines and mechanics are a different body of knowledge, and nobody had written it down yet. We are in the age of cars now.
The practical version: when you ask a model how to grow a business or structure a team, you get a well-synthesized answer from a world where the cost of doing work was much higher than it is for you today. That does not make the answer wrong. It makes it dated, in a way the confident tone will never signal to you.
What this looks like in practice
Building a site used to be hours of clicking. Fonts, sections, images, layout. Now the idea, the market and the branding go in up front, and AI produces the whole structure in one pass. That is the eighty. The ten at the end is reading it and saying this section is wrong, this copy misses, change that.
A client training deck we built recently came back usable on the first pass. It took about two hours of refinement to be good: cut this, add that, this argument does not follow, this one does. Anything client-facing earns those two hours every time.
- Decide the outcome before you open the tool. Write it down if you have to.
- Let AI do the middle — the part you have already proved you can do and should never do again.
- Read every word of what comes back before anything client-facing ships.
- Treat confident answers about strategy as drafts from an older world, not as verdicts.
The short version
- You own the first ten percent and the last ten percent. AI owns the eighty in the middle.
- LLMs deliver average answers with total confidence, and confidence reads as correctness.
- Checking the output only works if you decided the outcome before you asked.
- These models learned from a pre-AI world, so treat their strategy advice as dated rather than wrong.
Questions this raises
What is the 10/80/10 rule for AI work?
It splits a project into three parts. You own the first ten percent, which is the idea and the outcome you want. AI does the middle eighty, the repetitive build. You own the last ten, reviewing the result and fixing what is wrong. The name comes from flying, where the pilot is fully engaged for takeoff and landing and autopilot covers the rest.
If AI output is average, is it still worth using?
Yes. Average across eighty percent of the work is an enormous gain, as long as you supply the two ends. The failure mode is not that AI produces average work. It is shipping that work without first deciding what good looks like and then editing toward it.
Why does it matter that AI was trained on pre-AI material?
Because advice about business, strategy and leverage was written when work cost far more time than it costs you now. The model will hand you a confident, well-structured answer built on those assumptions. Use it as a starting draft and adjust for what you can now do in an afternoon.