A few weeks ago, OpenAI launched something called the Deployment Company. Most of the headlines read it the same way: "OpenAI starts a consulting firm." A handful of people argued about whether it would hurt McKinsey.
I read it differently, because I work inside the exact problem it's trying to solve.
For the last few years, part of my job has been driving AI adoption inside a large energy company. The thing almost nobody outside enterprise understands is that the models are no longer the bottleneck. The tools available today can already do far more than most large organizations use them for. The gap is everything between a capable model and a workflow that real people use on a Tuesday afternoon under real constraints. I think of it as the deployment gap, and once you've seen it from the inside you can't unsee it.
What the gap looks like
It's a very good model sitting next to a twenty-year-old ERP system that nobody wants to touch. It's a workflow that could be automated, except the three people who understand why it works the way it does are busy, and the person excited about AI doesn't have the domain context to know what would break.
The missing piece is almost never the technology. It's someone who understands the business deeply enough to know what's safe to change, and understands the tools well enough to know what's now possible, sitting in the same chair at the same time. That combination is rare. When I've managed to be that person on a project, things move fast. When that person doesn't exist, the AI initiative becomes another slide deck.
That's the story of enterprise AI right now. Capability is cheap and translation isn't.
Why $14 billion says the same thing
When OpenAI put real money behind the Deployment Company, what struck me wasn't the consulting angle. It was the shape of it. They're putting engineers directly inside the customer, building on the inside rather than advising from outside. They even bought a firm to get a hundred and fifty of these forward-deployed engineers on day one. That's an admission that the value is no longer in the model alone. It's in the last mile, which is the hardest, least glamorous and most human part of the chain.
A few things clicked for me reading the partner list.
Why would McKinsey and Bain invest in something that could eat their business? Because the part that gets eaten is the generic strategy deck. The part that survives, C-suite trust, relationships, the muscle to change how an organization operates, is their real moat, and now it gets a frontier-model engine bolted on. They chose to own a piece of the wave.
Why is private equity all over this? This one took me a second, because PE already hires consultants. But PE brings something the labs can't buy: they own the companies. They can mandate adoption across fifty portfolio companies at once, with no sales cycle and a direct line from margin improvement to a bigger exit. The old model was a consultant writing a hundred-day plan. The new one fuses the advice and the execution into one embedded team.
Strip away the names and the dollar figures and the signal is simple. The market is reorganizing around the last mile, which is the thing I've been doing in my corner of the energy world for a couple of years.
What I think it means
Part of why I'm writing this is selfish. I'm heading into an MBA, thinking hard about where to point the next decade, and this reframed the question.
For a while I was asking the standard MBA question: consulting, tech or finance? I now think that's the wrong frame. The advice layer is the part getting automated. The deployment layer is the part growing. So the better question is which seat puts me closest to deploying AI, with real domain depth behind me.
Domain expertise is the moat nobody can fake. A model can generate a strategy. It cannot, yet, be the person who knows that this valve, this contract, this team's workflow is the one you don't touch. The people who win the next decade will be the ones who can walk into a messy business, make the technology work, and understand the business well enough to be trusted with it. That doesn't demo well. But a $14 billion company was just built to industrialize it.
I don't have this figured out. I'm working it out in real time, from inside the problem. If you're building a career in AI and staring at the same fork, one filter: don't chase the layer that's getting automated. Chase the last mile, and bring something to it that the model can't. For me that's energy. For you it might be healthcare, law, manufacturing, or whatever world you understand from the inside.
I write these as I think through them, issued for review, not for construction. If you're working in this gap too, I'd like to compare notes.
