On June 30, 2026, AWS announced a $1 billion investment in a new Forward Deployed Engineering organization that will "embed thousands of experts with customers" to build and deploy agentic AI. Microsoft's Frontier Company runs on the same model, embedding engineering teams in a customer's business "to turn AI pilots into measurable outcomes." The two biggest clouds made the same bet, and it was not a bet on a better model. It was a bet on people who show up and build.
The catch for most mid-market companies: these programs are built for very large accounts. The model itself is sound, and you can buy it below the enterprise line if you hold a partner to the same three properties.
What is a forward-deployed engineer?
A forward-deployed engineer (FDE) is a senior engineer who works inside the customer's team and environment, building production systems against the customer's own data and processes. AWS describes its FDEs building "production AI systems with their data, governance, and processes." The contrast is an advisory engagement, where a partner assesses, recommends, and leaves the build to you.
We have worked as that embedded team for mid-market companies for years, so we read both announcements with interest. They confirm something we argue often, and they leave a gap our clients live inside.
What the hyperscalers conceded
The FDE investments admit that model capability was not what blocked enterprise AI. If adoption were blocked on model quality, a better model would fix it. You do not spend a billion dollars embedding engineers with customers unless the real blocker sits between the model and production: the data, the integration, the governance, the verification, the people.
That gap between what the technology can do and what an organization actually gets out of it is the story of enterprise AI in 2026. The hyperscalers priced it in public. We see the same gap when we are asked to take over someone else's AI proof of concept and get it into production.
Who the hyperscaler programs are built for
AWS names its targets plainly: regulated industries, financial services, and government, with customers including the NFL, the NBA, and Southwest Airlines. That is not a criticism. Embedding a team of hyperscaler engineers is expensive, and the economics work only on accounts large enough to repay it.
It does draw a line, and most mid-market companies sit below it. The companies with the messiest older systems, the thinnest platform teams, and the most to gain from hands-on help are the ones no hyperscaler will build a team around.
What mid-market buyers should demand instead
Mid-market buyers should demand the three properties that make the FDE model work, from whatever partner they can actually hire. Strip away the scale and those properties match what we see succeed in delivery.
Embedded beats advisory. The engineers sit inside your team and ship against your systems. Assessments and roadmaps do not ship systems. Builders do.
Outcomes beat deliverables. The engagement is scoped to a business result, not a stack of documents. If nothing runs in production, nothing happened.
Self-sufficiency is the exit. AWS says its customers leave with "trained internal champions ready to operate independently." Hold any partner to that. If the partner has to stay forever, you bought a dependency, not a capability.
None of the three requires a hyperscaler. Your realistic options are a commodity AI shop that hands you a demo, a large integrator running the same big-account economics as the clouds, or a small senior team that works embedded. Agentic delivery changed the math on that last option. In our own work, a few senior engineers using coding agents well now cover ground that used to take a much larger bench, which is what makes the embedded model affordable below the enterprise line.
Whoever you pick, hold them to embedded, outcome-scoped, and designed to leave, and work through what to ask an AI consultant before you sign. If you want to know whether your organization is ready for that kind of engagement, our AI readiness assessment is where we usually start, and our AI implementation framework is what the embedded work runs on.