On every subject I am genuinely good at, AI never gets past 80 percent. The last 20 percent is always me. On subjects I do not know well it works out exactly the same way, and that is the uncomfortable half of the observation.
80 percent is the ceiling
The models keep getting better, and the current one is genuinely very good. None of that has moved the number so far.
On every subject where I actually know what I am doing, I arrive at the same place. The machine does 80 percent and I do the last 20. That is not a complaint about the technology. I now plan for it as a division of labour, rather than being surprised by it on every project.
So I make sure I keep the upper hand, and I deliberately pull very few decisions out of the machine. I never really rely on it, and that holds on good days and bad ones alike. Anyone who does not plan for that last 20 percent is planning a project that gets 80 percent finished and then stops there.
The ceiling does not move with your knowledge
Here is where it gets uncomfortable, because on subjects I do not know well it works out exactly the same way. There too, AI does not get past 80 percent.
The only difference is that I cannot see it. On my own subjects I spot immediately which 20 percent is missing, and I know how to supply it. On unfamiliar subjects the same 20 percent is missing, and the output still looks complete to me.
That is precisely why I never rely on it. The results are not bad at all. My ability to judge them varies with the subject, and the quality of the output does not vary with it. Hand an unfamiliar field to AI and you get 80 percent, with nobody in the room able to name the missing 20.
The people deepest in it trust it least
This is the part that surprised me most. When I talk to other AI agency owners, they say the same thing. My own two developers say it too: I do not trust AI.
The deeper someone is in a subject, the less they trust it. That is the opposite of what you would expect. You would assume experience makes delegation easier, because you can judge better what you are handing over.
What actually happens is the reverse, because experience is what makes the missing 20 percent visible in the first place. Anyone who cannot see it treats the output as complete and hands over more of the work. The greatest confidence in AI sits where there is the least expertise in the room, and that is worth sitting with for a moment.
AI is the builder, you are the architect
A lot of very successful people use AI for everything, and let it make the decisions too. That is the point where it tips over.
I always use the same simple picture for this. AI is the construction worker and you are the architect. The construction worker is fast, reliable and better than you at plenty of things. He still does not decide where the house goes or what it looks like.
In practice that means you hand over the execution and you keep the decisions. It sounds obvious, and in daily work it blurs very quickly. A machine that makes good suggestions feels like a machine that makes good decisions. Those are two different things, and only one of them is yours to hand over.
So how much can you actually hand over
About 80 percent, of everything. The rest stays with you whether you plan for it or not, and the only real choice is whether you budget for it in advance.
If you want to check where you currently stand, take a piece of output from a field you know well. Go looking for the missing 20 percent in it. You will find it. Then take a piece of output from a field you do not know at all, and assume the same amount is missing there too.
I work from Dubai with clients across Europe, and this arithmetic has been the same everywhere I have applied it. The models will keep getting better, and the current one is already very good. So far the 20 percent has not gone anywhere.