In the span of six weeks: Microsoft stood up a two-and-a-half-billion-dollar subsidiary to embed six thousand engineers inside customer companies. OpenAI announced a four-billion-dollar deployment company. Anthropic launched a delivery venture with Goldman and Blackstone. Amazon committed a billion. Palantir, which invented this playbook twenty years ago, watched the industry adopt it at once.
Read plainly, these are admissions. The largest software companies on earth have priced the last mile of AI — the distance between a capable model and a running operation — and concluded it must be crossed by hand. When your solution to a software problem is six thousand engineers, you are conceding that your software cannot yet make the crossing alone.
Why the giants must do it this way
Embedding humans is the rational move for a company whose platform cannot govern autonomous work. A human engineer is the original governed agent: scoped, supervised, accountable, expensive. If your infrastructure cannot enforce policy, rehearse changes, and bound failures without a person in the loop, then the person is the safety system — and you must ship one with every deal.
Headcount is what the last mile costs when deployment is a service instead of a capability.
The consequence is a cost floor measured in salaries, and a delivery calendar measured in quarters. Neither can be engineered away from inside that model — the model is the constraint.
The tell
For anyone building in this era, the signal is unambiguous: the last mile is unsolved at exactly the moment the giants have committed billions to solving it with people. The counter-position is not better consultants. It is infrastructure in which agents do the embedding — under policy, through gates, onto a ledger — and a human holds authority instead of a wrench. Whoever makes deployment a capability rather than a service sets the cost curve for everyone else.