Six months ago I'd have told you the companies winning with AI were the ones with the best models. I've spent the last two weeks on about 105 calls with founders, operators, and people inside mid-size and enterprise companies. The model turns out to be almost beside the point.
Almost everyone is paying for the same AI now, and most of them feel like it isn't doing much for them. Then you look at their engineering teams. Same company, same spend, and the engineers are getting two to three times more done than they were a year ago. Everyone else is roughly where they started. One group leapt. The rest are stuck, despite using the same model.
That gap kept bothering me, because the easy explanations don't survive contact with the calls. It isn't that engineers are smarter or more "AI-forward." It isn't necessarily that the technology works for code and falls apart everywhere else. The intelligence is the same intelligence. What differs is the kind of AI each group was handed.
Engineers got AI that works. Tools like Claude Code and Codex have been refined for the developer workflow. They take a goal, break it into steps, use other software, check their own output, and keep going until the thing is done and validated. The person points; the AI does the labor. That is a real change in who does the work and how productive engineers are.
Everyone else got a chat box. You type a question, you get an answer, and the whole job of turning that answer into something useful stays with you. The AI wrote a paragraph. Or refined an email. Or wrote a first pass at the blog post. You still had to prompt it, hand it context, and then put it somewhere. You have to send it, track the reply, and remember where you left off tomorrow. The AI helped for a few minutes. You did the other six hours.
A chat box is reactive. It does nothing until you show up and prompt it, and it hands every output right back to you. So the harder you push it toward real work, the more you become its operator, feeding it context, stitching its answers together, walking it through every step by hand. One marketing lead I talked to could name the exact five roles she would hand to AI if she could. She had no way to make them real, so she was back to pasting prompts into a chat tab between meetings. That is the pattern almost everywhere. People didn't fail at AI. They were handed a tool that makes them do all the assembly themselves.
Some of you are already pushing back: "I have Cowork, isn't that a coworker?" It can be. But I asked a dozen people who'd tried it to show me their actual setup, and almost all were doing the same thing: opening a task and chatting. No projects. No automated routines running in the background. Files never connected. Only a few folks had something set up that provided recurring output, and most of them had someone else help set it up. The tool can genuinely act like a teammate. But almost nobody has it set up that way, because getting there takes exactly the kind of configuration the rest of us don't have the time or the instinct for. So it quietly becomes another chat box with a nicer logo.
So the upgrade most people are waiting for is the wrong one. They keep holding out for a smarter model, the next release that will finally make this click. But engineers pulled ahead for one concrete reason: their AI could act on its own. Same model as everyone else, with something around it that let it carry out the work instead of just describing it.
What the other ninety percent are missing is already sitting in front of them. It is the same model the engineers use. The missing piece is whatever turns it from a tool you operate into something closer to a colleague. Three things separate the two. A colleague is proactive: it does not wait for you to hand over every task, it sees what is happening and moves. A colleague is persistent: it carries your context forward, so Monday morning does not start with re-explaining everything. And a colleague does the work, it takes something off your plate and brings it back finished, instead of returning a draft and a list of next steps for you to do.
This is human-agent collaboration in the plainest sense: a person and an AI sharing context and passing work back and forth the way two coworkers would. Engineers have had a rough version of it for a year. For everyone else it is only now starting to show up.
I think this is the real race, and most people are watching the wrong scoreboard. Every large company has access to roughly the same intelligence today, so "best model" quietly stopped being the game. The game is who gives the rest of the world, the people buried in work who never learned to drive a chat box the way an engineer does, an AI that works like a coworker. Whoever does that will do for the other ninety percent what the coding tools already did for engineering.
A week ago I wrote that almost nobody is actually ahead on AI, that the race has barely started. This is the other half of it. The reason it still doesn't feel like much for most of us comes down to one thing: we were handed a tool that waits. The moment your AI stops waiting, the whole thing changes.