The gap between buying AI and using it
For the past two years, the enterprise AI story has mostly been about access — which model is smartest, which vendor has the best chatbot, which company signed the biggest cloud deal. Frontier Company, unveiled by Microsoft Commercial Business CEO Judson Althoff and set to be led by longtime Microsoft executive Rodrigo Kede Lima, is built around a different premise: that access to a model was never the hard part.
Early customers named by Microsoft include Unilever, Novo Nordisk, Land O’Lakes, and the London Stock Exchange Group, whose engineers reportedly worked with Microsoft’s teams to weave AI search features into an existing trading and research product rather than bolt on a new one. That kind of work — embedding a feature into a system people already use, rather than shipping a standalone tool — is precisely where many companies get stuck. It’s one reason more than half of surveyed executives say their organizations haven’t yet seen meaningful cost savings or revenue gains from AI, despite having deployed the tools widely.
The reason isn’t usually that the model is weak. It’s that the data feeding the model is scattered across departments, locked behind mismatched permissions, or simply not structured in a way any system — human or algorithmic — can use reliably. A customer-service AI can be extremely capable and still produce mediocre answers if it can’t see the shipping records, the refund policy update from last month, and the support ticket history in one place. Enterprise AI, in other words, is only as good as the plumbing underneath it.
Three layers that often get treated as one
Part of what makes this problem hard to talk about is that "having your data ready for AI" is actually three separate jobs that get collapsed into one sentence. There’s data availability — does the information exist somewhere in digital form. There’s data governance — who is allowed to see it, use it, or export it, and what gets logged for audit purposes. And there’s data usability — whether it’s organized, current, and connected to the actual workflow where a decision gets made. A company can ace the first without touching the other two, which is often exactly what happens after a rushed AI pilot.
| Layer | What it actually means | Typical failure point |
|---|---|---|
| Model/API access | Buying or licensing an AI system | Assuming access alone creates value |
| Data integration | Connecting systems so information flows between them | Data trapped in silos or legacy formats |
| Permissions & governance | Defining who can see, use, or audit data | Inconsistent rules across departments |
| Workflow redesign | Rebuilding how a task actually gets done | Old process kept, AI just bolted on |
| Outcome measurement | Tying AI use to a business result | No clear metric beyond "usage" |
Microsoft’s pitch is essentially that Frontier Company exists to walk customers through all five rows, not just sell them the first one. Whether that reflects a genuinely new capability or a repackaging of consulting and deployment work Microsoft was already doing — through programs like Industry Solutions Delivery and its FastTrack initiative — is a fair question, and one Microsoft hasn’t fully answered; most of the 6,000 people are existing employees reorganized under a new banner.
From fragmented data to something a business can act on
The logic behind embedding engineers becomes clearer if you trace the path data actually has to travel before an AI feature changes anything about how a company runs.
flowchart LR A[Fragmented internal data] --> B[Governed, connected systems] B --> C[Embedded engineering support] C --> D[AI features built into workflow] D --> E[Measured business outcome]
Each arrow in that chain is a place where projects commonly stall. Data can be connected but ungoverned, which creates compliance risk rather than value. Governance can be airtight but disconnected from daily work, so nobody actually uses the system. Even a well-integrated tool can sit unused if no one measures whether it changed a cost or a decision. Embedding engineers is one attempt to keep the chain intact from end to end — though it’s worth noting that this is one operating model among several, not a universally proven fix, and it works best for organizations large enough to absorb a long, service-heavy engagement.
Why customer data promises deserve a careful read
Microsoft has made two specific commitments to prospective Frontier Company customers: that client data won’t be used to train its models, and that customers can choose whichever AI model — from OpenAI, Anthropic, Microsoft, or open-source providers — suits a given task. Both promises target real anxieties. Companies worry that heavy reliance on a single model provider could let that provider absorb their industry expertise over time, a concern analyst Patrick Moorhead has raised specifically around fields like coding and law. Althoff has also pointed to Microsoft’s own experience: tying Copilot exclusively to OpenAI’s models early on limited flexibility as rivals improved.
These are useful design choices, but they remain vendor statements rather than independently verified contractual guarantees, and it’s reasonable to read them that way until confirmed in official product terms. There’s also a structural tension worth sitting with: even a genuinely open menu of models still runs on Microsoft’s cloud, security tooling, and engineering relationship. Model choice reduces one kind of lock-in while leaving the deeper infrastructure dependency largely intact.
A crowded field chasing the same insight
Microsoft isn’t alone in reaching this conclusion. Amazon Web Services launched a comparable $1 billion "forward deployed engineering" unit just days before Microsoft’s announcement, and OpenAI and Anthropic have each backed similar embedded-deployment ventures with outside financial partners. That four major AI players converged on nearly identical models within weeks of each other is itself telling: it suggests the industry now sees implementation, not model quality, as the real competitive battleground for 2026.
The takeaway beneath the headline number
None of this guarantees Frontier Company will lift Microsoft’s enterprise AI returns, and Microsoft has declined to say whether the $2.5 billion is new money or reshuffled budget — a detail that matters for judging how serious a commitment it really is. What the announcement does confirm is a shift already underway across the industry: the scarce resource in enterprise AI is no longer clever models but the patient, unglamorous work of connecting them to how a business actually runs — its data, its permissions, its habits. Whoever gets good at that layer, rather than whoever ships the flashiest demo, is likely to be the one companies keep paying for after the pilot ends.
