There's a quiet assumption in most AI workflows that I've come to think is the whole problem: that the model's job is to give you the answer. Ask, receive, proceed. It's efficient, and on anything that matters, it's a trap — because what you get back isn't an answer, it's one perspective delivered in the register of certainty. The certainty is free. It's attached to every output regardless of whether the reasoning underneath deserves it.
So I've reorganized how I work with these tools around a different idea. The models don't decide anything. They form a panel that does the part of the work I'm genuinely worse at doing alone: looking at a problem from several angles at once, quickly, without getting attached to the first frame. The research, the stress-testing, the "here's the case against what you're about to do" — that's what I delegate. The decision stays where it belongs, with the person who has to live with it.
How it actually works
In practice that means a brief goes out to a few models at once, each holding a deliberately different lens — one thinking about whether anyone actually wants the thing, one about what it does to the structure underneath, one about what it really costs to build. They work independently, blind to each other, because the moment they can see one another's reasoning they start to converge, and convergence is the opposite of what I want from them. Then I read all of it side by side, and I pay closest attention to the places where they disagree. That's where the real decision lives — not in the consensus, which is usually low-information, but in the fault lines between three competent, conflicting reads.
What it does to my own thinking
What surprised me is how much this changes the quality of my own thinking, not just the output. When you sit one model down and iterate with it, every exchange tends to deepen the original framing — you and the model talk yourselves further into the first idea. When you instead read three independent arguments that don't agree, you're forced to do the thing that actually constitutes judgment: weigh, discard, integrate, decide. The panel doesn't remove that work. It makes it unavoidable, and it gives it better raw material.
The decision I didn't make
The clearest case I can give — abstracted, but real — was an idea I was genuinely excited about: an automated energy-arbitrage project I'd half-convinced myself to build. I brought it to the panel expecting encouragement and instead got it methodically taken apart, from three directions at once: the demand I'd assumed, the edge that didn't survive the real costs, the maintenance I'd waved away. Nobody told me no. They simply showed me the decision in full, and the six months I'd have spent learning the same lesson the expensive way stayed in my pocket. I've worked this way for about half a year now — often enough that it's no longer an experiment, just how I approach anything I can't afford to get wrong.
Preparing a decision vs making one
I want to be precise about the boundary, because it's the part that matters most. There is a real difference between a model preparing a decision and a model making one. The first is leverage. The second is abdication. Everything I've built is on the first side of that line, and deliberately so. Anyone who works in a domain where decisions carry consequences — where someone has to be accountable for the call, not just correct — will recognize this instinct immediately. The tool gathers, compares, and documents. It does not absolve. The human reads, decides, and owns the result.
That, I think, is the mature version of working with these models. Not "AI decides and I rubber-stamp," and not "I ignore the machine and trust my gut" — but a panel of fast, tireless, differently-angled advisors feeding a human who still does the deciding. The models are extraordinary at showing you more of the problem than you could hold in your own head. They are not, and should not be, the thing that chooses. Keep that line bright and the whole arrangement gets better the more you lean on it.
Facing a similar decision in your company? → AI consulting