Microsoft launched Frontier Company on July 2. Two and a half billion dollars. About 6,000 engineers embedded inside enterprise customers, not delivering software to them but actually building and running AI systems on their behalf. Judson Althoff, Microsoft's chief commercial officer, announced it. The unit is not a separate legal entity — which is an important detail I'll come back to.

The explicit trigger is a stat I've been seeing cited independently for about a year. MIT's Project NANDA, drawing on 150 executive interviews, 350 employee surveys, and 300 public AI deployments, found that 95% of enterprise generative AI pilots yield zero measurable impact on P&L. Five percent hit their goals. The other 95 produce decks and demos and then quietly stop.

Here's what's important about that stat: the failure is almost never the model. The model is fine. What fails is everything else — data readiness, workflow integration, the absence of a defined business outcome before anyone starts building. These are fundamentally not software problems. They're people problems.

Microsoft is responding by selling people.

Source spread

• TechCrunch — Microsoft launches its own AI deployment company [builder] — the July 2 launch; $2.5B, 6,000 engineers, Althoff framing, early customer list.
• The Next Web — Microsoft Frontier Company, $2.5 billion AI deployment [skeptic] — raises the "services vs. products" structural question and the IBM Watson parallel.
• Fortune/MIT — 95% of GenAI pilots at companies are failing [academic] — the underlying Project NANDA research Microsoft is citing as its thesis.
• TechTimes — Frontier Company: 6,000 Engineers Target AI Pilot Failures [builder] — places the Microsoft launch alongside Amazon's $1B version and the May programs.

Pros & cons

What's real:

• The 95% stat is independently sourced and consistent with what every enterprise AI salesperson who will talk honestly will tell you. Pilots look promising, fail to scale, get blamed on the model. The model is usually not the problem. Microsoft building a unit that actually addresses the real bottlenecks — data, integration, defined outcomes — is the right diagnosis turned into a service.
• Embedding engineers at the customer site is structurally the correct intervention. The failure modes aren't fixable from the outside. You cannot ship a product that fixes bad data governance. Someone has to be inside the building.
• This is candid in a way enterprise AI marketing rarely is. Acknowledging publicly that Copilot for Microsoft 365 doesn't install itself is a meaningful shift from the "just turn it on" messaging of 2023 and 2024.

What deserves a side-eye:

• The margin profile of a services-heavy unit is fundamentally different from Microsoft's software licensing model. IBM ran this playbook with Watson Enterprise. It generated billable hours that scaled poorly, and clients couldn't operate the systems after IBM left. Microsoft needs to be designing for graduation, not dependency.
• Amazon committed $1B to an essentially identical initiative two days before Microsoft's announcement. OpenAI launched a comparable program in May. Anthropic launched one in May too. When every major lab is embedding engineers at enterprises, that's not a differentiation strategy anymore.
• "Not a separate legal entity" means there's no Frontier Company P&L for the market to audit. If this underperforms, Microsoft can quietly redirect without a press release. Convenient for Microsoft and not particularly reassuring for customers making multi-year commitments.

What builders need to know

• If you're selling AI into enterprises — as a product company, consultant, or indie builder — Microsoft is now competing with you in the integration layer, not just the product layer. Factor that into your positioning and your pricing.
• The 95% failure stat is not theoretical, and the causes are specific: data readiness, workflow integration, undefined business outcomes. If your product doesn't address those, you're building into the failing 95%. The differentiated position is being the vendor who helps enterprises avoid the stat, not just sells them capability.
• "Not a separate legal entity" is worth tracking. Microsoft can quietly wind this down without a press release if it's not working. If a customer is making a multi-year commitment to a Frontier Company engagement, ask what happens to the program if Microsoft deprioritizes it.
• Amazon, OpenAI, and Anthropic all have versions of this now. Table stakes for selling frontier AI into enterprises: don't just ship the model, offer to help install it.

Further reading

• TechCrunch — Microsoft launches its own AI deployment company with $2.5 billion commitment
• The Next Web — Microsoft Frontier Company: $2.5 billion AI deployment
• Fortune — MIT report: 95% of generative AI pilots at companies are failing
• TechTimes — Microsoft Frontier Company: $2.5B and 6,000 Engineers Target AI Pilot Failures