In the first week of July 2026, Microsoft committed $2.5 billion and 6,000 industry and engineering experts to a new business whose entire job is getting AI deployed inside customer operations. Amazon committed $1 billion to its own forward-deployed engineering effort days earlier, and Anthropic and OpenAI launched comparable ventures earlier this year. When the largest cloud and model providers on earth independently conclude that the bottleneck is not the model but the deployment, IT leaders should take the signal seriously. The industry just admitted, with its balance sheet, that buying AI and getting value from AI are two different projects.
Forward-deployed engineering is a practice where a vendor sends its own technical staff to work inside a customer's operations, designing, building, deploying, and then continuously operating systems on site rather than handing over a license and a training deck. Microsoft's version, announced July 2, 2026, is called Microsoft Frontier Company. It is led by longtime enterprise executive Rodrigo Kede Lima, and the stated model is to embed industry and engineering experts with customers to co-design and continuously improve AI systems against measurable business outcomes. Early named customers include Unilever, Novo Nordisk, and the London Stock Exchange Group.
The reason this model is suddenly worth billions is unflattering to the last three years of enterprise AI. Research from MIT's Project NANDA found that roughly 95% of enterprise generative AI pilots delivered no measurable impact on profit and loss. Not a modest return. No measurable impact at all. The models were not the problem. The pilots were real, the demos worked, and almost none of it reached the operational core of the business where money is actually made or saved.
The AI bottleneck has moved from capability to deployment. Roughly 95% of enterprise AI pilots produce no measurable P&L impact, which is why the largest vendors are now spending billions to put engineers inside their customers. The lesson for everyone else is not to buy a bigger model. It is to fund the deployment work: integration, data access, governance, and continuous operation.
The deployment gap is widening at exactly the moment agents are becoming default infrastructure. Gartner projects that 40% of enterprise applications will ship with embedded agents by the end of 2026, up from under 5% in 2025. Microsoft reports active agents in the Microsoft 365 ecosystem growing roughly fifteen-fold year over year. In practical terms, agentless enterprise software is about to become the exception.
That creates a specific problem for IT Directors. Agents are entering your environment through software you already own, on the vendor's schedule, not yours. You are not deciding whether to adopt agents. You are deciding whether they arrive governed or ungoverned. Adoption is running ahead of control. Agents are reaching production faster than the frameworks meant to scope, approve, and audit them, and that gap is where the real risk lives. We covered the control side of it in our AI agent governance checklist.
When we look at AI efforts that never made it out of the demo, the failure is almost never the model's reasoning. It is the unglamorous work around it that no one budgeted:
Notice that four of those five are operational disciplines, not data science. That is precisely why the hyperscalers are shipping engineers instead of features, and it is the model behind our managed intelligence practice.
Microsoft Frontier is aimed at Unilever and the London Stock Exchange Group. If you run IT for a 200-person manufacturer, a regional healthcare group, or a private equity portfolio company, no one is embedding a forward-deployed team in your operations at that scale. The strategic question is not how to get on that list. It is how to get the same operating model at your size.
The good news is that the model is reproducible, because what actually creates the value is structural, not headcount. Forward-deployed engineering works because the people building the system sit close to the work, own outcomes rather than deliverables, and stay after go-live. A mid-market company can buy exactly that shape of engagement without buying 6,000 people:
The embedded-agent wave changes the build-versus-buy calculation in a useful way. If 40% of enterprise applications will have agents built in by year end, then a growing share of routine capability arrives with software you already license. Paying to custom-build what your vendor is about to ship is a poor use of capital.
What vendors will not do is the part specific to you: your process, your data model, your approval chains, your exceptions. That is where custom AI agents and workflow automation earn their keep, and it maps cleanly to a rule of thumb. Take the embedded agent for generic work inside a vendor's own product. Build for the workflow that is genuinely your business. Embed engineering capacity, internally or through a partner, for the integration, governance, and operations layer that makes either one real.
For a deeper walk through the economics of that decision, see our 2026 look at AI agent economics and our AI ROI guide.
The most important enterprise AI story of this summer was not a model release. It was the largest cloud and model providers independently pricing the deployment gap and deciding it was worth billions to close by hand. Read that as confirmation of what your own stalled pilots have been telling you: capability is no longer the constraint, and the return comes from integration, governance, and sustained operation. You do not need a 6,000-person embedded team to apply the lesson. You need one workflow with a real metric, funded deployment work, governance from day one, and someone who owns the system after it launches.
Infonaligy is an AI-native managed intelligence provider based in the Dallas–Fort Worth metro, delivering AI deployment, governance, and operations across Texas and remotely nationwide.
Book an assessment and we'll pick one workflow with a real metric, scope the integration and governance honestly, and stay through operation so it actually reaches production.
Infonaligy