Microsoft announced on July 2 that it is standing up a new operating unit called Microsoft Frontier Company. Judson Althoff, who runs the Commercial Business, made the call. Rodrigo Kede Lima, formerly president of Microsoft Asia, runs the new unit. The commitment is $2.5 billion and roughly 6,000 engineers. The named launch customers are the London Stock Exchange Group, Unilever, Land O’Lakes, and Accenture. Althoff described the group as the largest outcome-driven engineering organization in the industry, which is corporate for “the pilots are stalling and we are sending our own people in to unstall them.”
The context is the tell. Two days before the Microsoft announcement, Amazon committed $1 billion to the same shape of business, embedding AWS engineers inside customer companies to get AI into production. Anthropic and OpenAI launched analogous programs earlier in the year. What used to be a distinctive Palantir move, the forward-deployed engineer sitting inside a customer’s operations center for months at a time, has become the default enterprise-AI playbook of 2026. Every major AI vendor is now, in the shape of one of its business units, a consulting firm.
The reason is not mysterious. The MIT State of AI in Business report earlier this year put roughly 95 percent of enterprise generative-AI projects at zero measurable return, and the finding held up under scrutiny. The gap is not model quality. Sonnet 5 and GPT-5 and Gemini 3.5 are all capable of running most of the workflows enterprise buyers want them to run. The gap is the last twenty feet of deployment, the piece where the vendor’s model has to talk to Workday and SAP and a data pipeline nobody has documented since 2019. That work is expensive, unglamorous, and does not scale like software. It scales like plumbing. Microsoft has decided plumbing is now core to its business.
The strategic frame worth naming: Frontier Company is a defensive move dressed as an offensive one. If Microsoft’s Azure AI customers keep failing to deploy, they eventually renegotiate or leave. Sending 6,000 Microsoft engineers into their buildings to get the deployments over the line is cheaper than losing those accounts, and it deepens Azure dependence in the process. The pitch is that customer data will not train Microsoft’s models and clients can still run competing AI systems, both of which are true and both of which matter less over time. Every workflow the Frontier team builds is one more piece of Azure that the customer would have to rip out to switch. That is not an accident. That is the plan.