First in a series on building a virtual company.
I wanted help. Most of us do. There is always more work than there are hours, and the tools had finally gotten good enough that handing some of it off felt real. So I started handing things to an AI. Write this, look into that, pull together the other thing. Nothing grand. I just wanted to be a little less behind.
And it worked, the way a new tool works. I would give it something and it would give something back, and for a while that was enough to feel like I had found a second pair of hands.
It was best at the simple things, the self-contained jobs you can hand over in a sentence. Those came back clean and quick. But the bigger the work got, the more it had parts that leaned on each other, the more of it I was holding in my head at once and trying to keep in step, the more things began to slip. And the worst of it was the slips I couldn’t see. The work would come back looking whole and finished. And whatever was off sat tucked somewhere I wouldn’t notice, not until much later, not until it cost far more to find.
The more I leaned on it the more I ran into the same wall, and it took me a while to place where I knew the wall from. The work was only ever as good as what I handed over, and that was a harder bar than it sounds. Not because I was vague. Because you cannot say everything. The more there is to a thing, the more of it goes unspoken: the details you assume, the ones you gloss over because they seem obvious, the gaps you don’t even register as gaps. And every gap is a small decision passed to whoever is on the other end, and they fill it. Sometimes the way I would have, sometimes not. That is where the work came back sure of itself and wrong. Not stupid wrong. Confident wrong. The wrong of someone who heard you, filled in the gaps with their own idea of you, and ran with it.
Not stupid wrong. Confident wrong.
That is when I placed the wall. I had hit it a hundred times before, every time I handed something to a person and got back a version of what I said instead of what I meant. It was the space between us. The handoff. What I had failed to say, or said badly, or assumed they already carried.
In 1990, two economists, Oliver Hart and John Moore, set out to explain why contracts are always, in a way, unfinished. (Hart would later share a Nobel for this line of work.) What they showed, roughly, is that you can never write a contract that says what someone should do in every situation, because you can never see every situation coming. There are always gaps. And the work runs on those gaps, on the judgment of whoever ends up filling them, as much as it runs on anything you actually managed to say out loud. They were describing companies, and the contracts between them. They were also, without knowing it, describing the thing on my screen.
I had been treating the intelligence as the hard part. The intelligence was the easy part. The hard part was the same hard part it has always been.
So I stopped fiddling with prompts and started doing something that felt oddly familiar. I gave the system someone to sit between me and the work, to take what I meant and turn it into what got done. Then that someone needed people to hand things to, and I found I didn’t want them all to think the same way. The one asking whether a thing should ship shouldn’t be the same one asking whether it’s any good. In my own head those are different voices, and the arguing between them is where my better decisions come from. So I split them out.
I sat back and looked at what I had drawn, and it was not a tool I had configured. It was an org chart. People with roles. Things they owned. Someone they answered to.
There is a name for the person all of that answers to. I just hadn’t thought of myself that way since the last time I ran a team of actual people.
I had not set out to build anything. I only wanted to use the thing, the way you use a tool. But it turns out you cannot really use it, not well, until you put an organization around it. That was the piece that had been missing. Not a smarter model. The staff, the roles, the person they all answer to. A company, with me somehow running it, whether I had meant to take the job or not.
Once I saw it that way I couldn’t stop seeing it, and the strange relief was that none of my problems were new. Work coming back wrong was not a prompting bug, it was a bad brief, the oldest management failure there is. Two agents doing the same thing twice was not a glitch, it was two people nobody told who owned what (I’ve written about how their dysfunctions came to mirror ours too). The fix for confident and wrong was never a cleverer sentence. It was what you would do with a sharp new hire who went the wrong way with too little to go on. Sit closer. Say more. Build a better channel between you.
The fix for confident and wrong was never a cleverer sentence.
We have a hundred years of hard-won knowledge about exactly this. How to delegate without losing the thread. How to brief someone so they come back with what you meant. How to build a team whose judgment you trust enough to leave the room. We call it management, and most of us have spent a career either practicing it or quietly suffering under someone who couldn’t. None of it was written for this. All of it fits.
That, I think, is the part getting lost in all the noise about smarter and smarter models. The models are getting smarter, quickly. But I don’t think the people this rewards are the ones with the cleverest prompts. I think it rewards the ones who already know how to lead, because leading is quietly what the moment is asking for.
I came looking for help with my work. That is not what I found. I found a job I had done before, wearing a shape I didn’t recognize at first, and it asks more of me than the work ever did, not less.
Next in the series: The Average Isn’t You. On why one capable AI hands you the averaged answer, and what it takes to get the real choice back.