For two years the AI industry sold a simple story: buy access to a powerful model and productivity would follow. The past week suggests the sellers no longer quite believe it. On Thursday Microsoft announced Microsoft Frontier Company, a new operating business backed by $2.5 billion and 6,000 engineering and industry experts, whose entire job is to sit inside customer organisations and make enterprise AI deployments actually deliver a return. Two days earlier, Amazon Web Services committed $1 billion to a near-identical unit. OpenAI and Anthropic have launched their own versions. The shape of the AI business is quietly changing from shipping software to showing up in person.
The industry even has a slightly awkward name for it: forward-deployed engineering, borrowed from Palantir, which has run this playbook for years. The model is old consulting logic in new clothes. Rather than hand a company an API and wish it luck, the vendor sends engineers to learn the customer's messy internal data, wire the models into existing systems, and stay until something works. Microsoft's commercial chief Judson Althoff bristled at the label, calling his effort "the largest, most capable, outcome-driven engineering organization in the industry," but the resemblance to AWS's freshly announced billion-dollar embedded-engineer unit is hard to miss.
Why now? Because the gap between what AI does in a demo and what it does inside a real company has become impossible to ignore. Microsoft's own telling is unusually candid. Althoff told Reuters that when the company built Copilot three years ago it "made a mistake by binding it to OpenAI models only," and that customers now want the freedom to swap models as the frontier moves and to fine-tune them on their own data. The lesson learned the hard way is that the model matters less than the plumbing around it, and the plumbing is where deployments stall.
There is a quieter, more defensive motive too. As one analyst put it, large firms increasingly suspect that leaning on Anthropic or OpenAI will eventually let those labs learn enough about their business to compete with them, in fields such as coding and law. A deployment arm that lets customers keep the results of the work, as Microsoft Frontier explicitly promises, is partly a reassurance against exactly that fear. It is also a land grab: Microsoft already has engineers embedded across much of the Fortune 500, a head start that AWS and the labs will now spend heavily to match.
The strategy is expensive, and it concedes something. Sending thousands of engineers to babysit installations is the opposite of the frictionless, self-serve scaling that made software such a wonderful business in the first place. It looks a lot more like the services model that AI was supposed to disrupt. The bet across the industry is that whoever gets closest to the customer's data now will own the relationship when the technology finally does become plug and play. If that day arrives, the armies can go home. Until then, the fastest-growing job in AI may be the person who flies out to make it work.