Microsoft launched Frontier Company on July 2: a new $2.5 billion business unit staffed by roughly 6,000 engineers whose job is to embed directly inside customer operations and co-design, deploy, and govern AI systems on-site, rather than sell a product and leave implementation to the customer (Microsoft Frontier Company: AI engineering that amplifies and protects your intelligence, The Official Microsoft Blog). Named clients include LSEG (London Stock Exchange Group), Land O’Lakes, Unilever, and Novo Nordisk, with Accenture, Capgemini, EY, KPMG, and PwC signed on as systems-integrator rollout partners. The unit is led by Rodrigo Kede Lima, a 30-year Microsoft veteran most recently president of Microsoft Asia, reporting to Judson Althoff, CEO of Microsoft’s commercial business.
The practice has a name, and it’s not new
What Microsoft is scaling here is known as forward-deployed engineering: sending your own technical staff to work inside a customer’s operations to build and operate systems on-site, instead of shipping a tool and providing support from a distance. It’s the model Palantir built much of its enterprise business around, and it’s notable that Microsoft, which has spent two decades selling software at arm’s length through licensing and partners, is now committing billions to doing the opposite for AI specifically. Microsoft is explicit that Frontier Company goes beyond typical forward-deployed engineering arrangements, describing it as intended to be the largest, most capable, outcome-driven engineering organization of its kind in the industry.
The number that explains the whole move
The actual justification lives in research that predates this announcement by almost a year: MIT Project NANDA’s “The GenAI Divide: State of AI in Business 2025” report, based on interviews with roughly 150 enterprise leaders, a 350-person employee survey, and analysis of 300 public AI deployments, found that 95% of enterprise generative AI pilots deliver no measurable return on investment, despite an estimated $30-40 billion in enterprise GenAI spending (Group Publications, MIT Media Lab NANDA). Critically, the report’s own conclusion is that the failure isn’t rooted in weak models, it’s what the researchers call the “learning gap”: most GenAI deployments don’t retain feedback, adapt to context, or integrate into a company’s actual workflows, structures, and culture.
That statistic is the actual problem Frontier Company exists to solve, not a lack of capable models. Microsoft, Copilot, and every other enterprise AI vendor have spent two years demonstrating that frontier models are impressively capable in isolation; the unsolved problem is turning that capability into a deployed system that changes how a specific company’s processes actually run, integrated with its existing data, security requirements, and workflows. That’s an integration and change-management problem as much as a model-capability problem, and it’s not one you solve by shipping a better model.
How Microsoft is positioning trust and lock-in
Two lines from Microsoft’s own announcement are worth quoting directly, because they preempt the two most obvious objections to a vendor embedding thousands of its own engineers inside a customer’s operations. On data usage: “a customer’s IQ is protected. Their data, their IP, their competitive advantage — none of it is used to train models in ways that commoditize [it].” On the lock-in question specifically: “Customers shouldn’t be locked into a single model any more than they should be locked into a single technology vendor.” Whether Frontier Company actually operates that way in practice is a separate question from what the announcement promises, but the fact that Microsoft addressed both concerns explicitly in the launch messaging suggests the company is aware those are the two biggest reservations enterprise buyers will have about the model.
What this signals for the rest of the industry
If a 95% pilot failure rate is roughly accurate across the industry, and it’s consistent with what a lot of enterprise IT leaders have said anecdotally over the past year, it suggests the next competitive battleground in enterprise AI isn’t primarily about who has the best underlying model. Anthropic, OpenAI, and Google are all racing on frontier capability; Microsoft’s bet with Frontier Company is that the harder, more defensible moat is actually deployment and integration expertise, the unglamorous work of getting a specific company’s specific systems to actually use AI in a way that shows up in a P&L statement. Watch whether AWS, Google Cloud, or a major systems integrator responds with a comparable forward-deployed offering over the next few quarters; if the failure-rate numbers hold up, this looks less like a one-off Microsoft initiative and more like where enterprise AI spending is about to concentrate.
The obvious tension in the model
Committing 6,000 of your own engineers to sit inside customer operations is expensive in a way that selling software licenses simply isn’t, and it doesn’t scale the same way. A software product can serve a million customers off the same codebase; forward-deployed engineering scales roughly linearly with headcount, more clients means proportionally more engineers embedded on-site. That’s exactly the model Palantir has run for years, and it’s part of why Palantir’s revenue per employee looks nothing like a typical SaaS company’s, the labor is the product, not just the delivery mechanism for it.
For Microsoft this is a deliberate trade: lower margins and slower scaling per customer, in exchange for deployments that actually succeed and, more strategically, deep operational lock-in once its own engineers have built a client’s AI workflows around Microsoft’s stack, a tension the company’s own “no vendor lock-in” messaging above doesn’t fully resolve. It only makes sense as a $2.5 billion bet if Microsoft believes the alternative, continuing to sell Copilot and Azure AI services at arm’s length while 95% of pilots quietly fail, is a bigger long-term threat to its AI business than the cost of the forward-deployed model. That’s a real bet on where the failure is actually happening in enterprise AI adoption, not just a marketing repositioning, and it’s the reason this is worth tracking rather than filing under routine product announcement.
