Somewhere in the last two months, four of the biggest names in tech quietly agreed on something — without ever sitting in the same room.
Amazon put a billion dollars behind it. Microsoft put two and a half billion behind it, plus 6,000 people of its own. OpenAI and Anthropic had already made similar moves back in May. Different press releases, different dollar figures, same underlying bet: the model isn’t the hard part anymore. Getting it to work inside a real company is.
If you’ve been following AI mostly through the model releases — who’s got the best benchmark this month, who dropped a new version over the weekend — you’d be forgiven for missing this. It didn’t come with a flashy demo. It came dressed up as an org chart change. But I think it’s the more telling story of the summer, and it’s worth sitting with.
The move nobody’s naming yet
Start with Microsoft, since it’s the freshest and best documented. A few days ago, they announced “Microsoft Frontier Company” — not a product, not a model, but 6,000 engineers whose whole job is to sit inside client companies and build, run, and babysit AI systems for them. Early names attached to it: the London Stock Exchange Group, Unilever, Land O’Lakes, Novo Nordisk. This is what the industry calls “forward-deployed engineering” — a playbook Palantir has run under the radar for two decades, and one that, this year, every major AI player suddenly wants a piece of.
Here’s the part that made me pause. Microsoft’s pitch rests on a promise: you won’t be locked in. Run OpenAI, run Anthropic, run whatever model makes sense for you — the platform is designed to be model-agnostic. Fair enough, on paper.
But once Microsoft’s own engineers have spent months wiring your workflows, your data pipelines, and your approval chains into Azure’s plumbing, how much does “you can switch models” really mean? You can swap the engine. Nobody’s swapping the car it’s bolted into. That’s not a knock on Microsoft specifically — Amazon, OpenAI, and Anthropic are all building the same kind of dependency through the same kind of arrangement. The tension is built into this business model, not any one company’s intentions.
So the real story isn’t Microsoft
It’s that four competitors, who otherwise spend their marketing budgets trying to look nothing alike, landed on the same conclusion within about eight weeks of each other: selling a good model isn’t enough anymore. You have to move in with the customer.
That’s a quiet kind of admission. Nobody stood up and said “our AI doesn’t work without us in the building.” But that’s roughly what the org chart is saying.
Why this is the harder problem
For the last two years, the AI conversation has mostly been about intelligence — is this model smarter than that one, does it reason better, does it hallucinate less. Underneath all of that, a less glamorous truth has been building: most companies don’t have a model problem. They have an operating problem.
Their data is scattered across a decade of legacy systems. Nobody’s quite sure who’s allowed to approve what. There’s no clean way to measure whether the AI initiative is saving money or just moving the mess around. And even when all of that is solved on paper, someone still has to convince a room full of people to change how they’ve worked for fifteen years. No model update fixes that. You can hand a company the smartest system in the world, and it’ll still get stuck in the mud of permissions, data quality, and habit.
That’s the gap forward-deployed engineering is trying to fill — and it’s also, I suspect, why it won’t work as often as the press releases imply. A Gartner report focused specifically on these embedded-engineer arrangements projects that within two years, 70% of enterprises will be forced to abandon the agentic AI projects that come out of them, citing high vendor costs and a lack of internal skills to carry the work forward independently. Not because the AI was bad. Because deployment inside a real, messy, human organization is genuinely hard, and adding more engineers to a company doesn’t fix the underlying rot in its data or its decision-making. You can lead a horse to water, as they say — the rest is still up to the horse.
Agents raise the stakes, not lower them
This gets more pointed the moment you introduce agents — AI systems that don’t just answer questions but take actions on your behalf. The question used to be “is this answer good?” Now it’s “who approved this action, what data fed into it, and can we undo it if it’s wrong?” That’s not a model question. That’s a governance question, an audit-trail question, a “does anyone actually own this” question. And it’s exactly the kind of question a forward-deployed engineer can help you set up — or can just as easily paper over, if the underlying discipline isn’t there.
What I’d watch for
I don’t think the winners of the next stretch of this industry will be the ones with the loudest model. I think they’ll be the ones who figure out how to make AI boring — reliable enough, measurable enough, safe enough that a business can trust it every day without a team of embedded engineers permanently attached to keep it upright. That’s a far less exciting thing to put in a press release. But it’s what will actually separate the companies that make AI stick from the ones still explaining, two years from now, why the pilot never became the product.
