Why Microsoft Is Embedding 6,000 Engineers at Client Sites

Big Tech Is Stationing Its Own Engineers Inside Client Offices

On July 2, Microsoft announced it was spending $2.5 billion to set up a separate unit called “Frontier Company.” What this unit does is simple: it places 6,000 employees inside client companies.

Two days earlier, AWS committed $1 billion to the exact same play. Before that, in May, OpenAI and Anthropic each set up their own versions of the same organization. Anthropic went further and formed a $1.5 billion joint venture with Blackstone and Goldman Sachs.

While headlines about AI-driven layoffs pour out daily, the very companies selling AI are spending trillions of won to hire people — and then seating those people in someone else’s office.

I don’t think this should be read as a heartwarming twist where “humans turn out to matter after all, even in the AI era.” It looks much more like vendors admitting, in effect, that handing over a model alone doesn’t actually work at the client’s site. And this kind of demand for people is likely to last only as long as it takes for adoption to settle in.

📊 FDE job postings are surging

This role is called an FDE¹. Translated literally, it’s something like “Forward Deployed Engineer,” and the term comes from military jargon. Palantir started using it when it began sending its own engineers directly to US military bases in Afghanistan. Judson Althoff, Microsoft’s commercial business CEO, has himself credited Palantir with popularizing the title.

The core idea is: you don’t sell the product and walk away. The vendor’s engineers embed with the client’s team, take the company’s actual workflows apart from the inside, and build a system together that fits.

According to aggregated hiring data, FDE job postings had jumped more than 1,000% year-over-year as of early 2026. Compensation is just as aggressive. At frontier labs (OpenAI, Anthropic), a senior FDE’s total compensation runs $450,000–$550,000, and staff-level roles clear $600,000. That’s 2 to 3.5 times what Palantir pays its traditional FDSEs.

That said, this compensation data comes from hiring platforms and self-reported community figures, so the upper end is likely overrepresented. Still, the direction is unmistakable: the market is paying a premium for this role.

💰 Why Would a Vendor Spend Its Own Money to Place People?

If the product is good, shouldn’t customers just use it on their own? Why would a vendor build engineers at its own expense and embed them for free inside someone else’s company?

The answer lies in the failure rate of adoption.

A 2025 report from MIT Media Lab’s NANDA project is the most useful thing I’ve found for understanding this trend. Its headline finding: roughly 95% of enterprise generative AI pilots fail to show up as measurable results on the bottom line. For a while, this number got consumed by the press as proof that “AI is a scam” — but the original report reads quite differently.

The failure cause the report actually identifies isn’t model performance. It’s the learning gap². The model doesn’t understand the organization’s workflows, the organization doesn’t know how to give the model context, and there’s no one bridging the two. It’s as if a company bought a tool and then left it sitting there, with no idea what’s actually happening inside the business.

The methodology has real limits, too. This study combined roughly 300 public case studies, 52 executive interviews, and 153 survey responses — it’s not a peer-reviewed paper. It’s hard to treat the “95%” figure as precise statistics. Still, vendor behavior backs up this diagnosis. Microsoft’s Judson Althoff said in an interview, “Customers are all standing at completely different points, still trying to figure out how to work with AI.”

In other words, the FDE boom isn’t a sign that AI is working well — it’s a sign that adoption isn’t going well inside companies. Vendors have started shouldering, at their own expense, the adoption work and cost that customers were originally supposed to bear.

For Microsoft, there’s also a sense of urgency. Its stock is down 21% this year, the worst performance among megacap tech stocks. Copilot hasn’t gained the enterprise foothold it was expected to, and GitHub Copilot has been losing share to later entrants. This is a situation where “the model is good, so why isn’t it being used” can no longer be blamed on the customer.

🔑 A Job Whose Success Means Its Own Extinction

AWS has stated a specific principle in describing its FDE organization: a team of 5-6 engineers runs on a roughly 45-day cycle, and the goal is to leave the customer self-sufficient by the end of it.

In short, FDE is a role whose success metric is making the client capable of running the system without further help.

That’s the opposite direction from traditional consulting. Consulting’s revenue model depends on keeping contracts as long-running retainers, but FDEs are deployed on the premise that they’ll finish the handoff within a fixed period and then leave.

Microsoft’s decision to structure this not as a “consulting division” but as a separate “Company” with its own P&L reads the same way. It’s not a device for growing services revenue — it’s a device for pushing up platform adoption. For context, Microsoft’s existing enterprise-and-partner services revenue runs around $2.1 billion a quarter, growing at 2.5%. That tells you they’re not trying to make money here.

So it’s more accurate to see this hiring boom not as a durable category of jobs, but as headcount needed only during the adoption phase. Once a client becomes self-sufficient, that talent has to move on — to another client, or another role.

Oswarld’s Lens

The people who survive this shift, I think, won’t be the ones who use AI tools skillfully — they’ll be the ones who can connect what AI can do to what an organization actually does.

The FDE job postings I looked at didn’t just ask for coding chops. They also demanded the ability to grasp a client’s business processes, figure out what data lives where and in what shape, and navigate internal politics well enough to actually push a deployment through. This is the human role that fills the “learning gap” I mentioned earlier.

Building Notion’s Korea community, integrating AI products at Kakao Brain, and designing go-to-market strategy at Gamma — across all three, I kept running into the exact same pattern. A product almost never sells because it’s good. It sells once someone builds the context connecting that product to the client organization. The bottleneck was never the technology — it was context. And context-production capacity is precisely what Big Tech is now buying up for trillions of won.

But there’s one more thing worth thinking through when you look at this hiring trend. Reading this hiring boom as reassurance that “at least there’s still a place for humans” gets you only halfway there. FDE is a role designed to work itself out of a job. As long as AWS is targeting client self-sufficiency within 45 days, this position is structurally temporary by design.

So the more important question isn’t “can I get into this field,” but “what should I be accumulating while this demand still exists?” It’s not enough to just learn how to use the tools. You need to build up experience actually understanding the context of a business domain and translating it into a working system — because that experience carries forward afterward. FDE roles are growing in Korea too. Some people lump FDE in with SI (system integration), treating it as just another form of outsourced development, but I think FDE is closer to people who redesign how work itself gets done. Among the FDE teams I’ve personally encountered in Korea, I’d point to Hyuntae Hwang at Space Y, and the in-house FDE teams built at Kakao Pay and Rebellions as genuinely meaningful examples. Worth keeping an eye on if this space interests you.

Closing

Let me sum this up in three points.

First, the FDE hiring boom isn’t a signal that AI is working well — it’s a signal that adoption isn’t happening. Vendors have started covering the cost of that adoption failure out of their own pockets.

Second, the bottleneck isn’t model performance — it’s organizational context. That’s why what’s being rewarded right now isn’t coding ability, but the ability to understand a company’s workflows and data conditions and connect them to AI.

Third, this role isn’t permanent. It’s a job whose very definition of success is making itself unnecessary. So it makes more sense to treat this period as an opportunity to build domain context and hands-on implementation experience, rather than as a chance to land a long-term position.

If there’s one thing I’d suggest trying this week, it’s writing a single page answering: “What would I need to explain in order to hand this off to AI?” — based on whatever you’re working on right now.

The senior FDE’s total compensation in the $450,000s that we saw earlier is exactly what the market pays for the ability to produce that explanation.

---

💬 If you’ve worked on an AI adoption project, was the sticking point model performance — or was it organization, data, and process?

Drop a quick comment on which one it was. Once enough cases come in, I’ll organize them by type in a future issue.

---

📨 If you know a colleague wrestling with AI adoption, please pass this along.

References & Further Reading

Primary sources

• Jordan Novet, “Microsoft commits $2.5 billion and 6,000 employees to new AI implementation unit”, CNBC, 2026-07-02. : This includes an interview with Althoff. His line that “customers are all standing at different points” is the most honest explanation of why this unit exists.
• Microsoft, “Microsoft Frontier Company: AI engineering that amplifies and protects your intelligence”, Microsoft Official Blog, 2026-07-02. : Worth reading to see how the vendor itself packages this unit — the tone is strikingly different from the CNBC piece.
• CNBC, “AWS puts $1 billion into new AI unit to embed engineers with customers”, CNBC, 2026-06-30. : A 45-day cycle, teams of 5-6, and customer self-sufficiency at the end — these three details are the most important evidence behind today’s piece.
• MIT Media Lab NANDA, The GenAI Divide: State of AI in Business 2025, 2025. : The concept of a “learning gap” matters far more than the “95% failure” headline. Read it keeping in mind that this isn’t a peer-reviewed paper.

Background

• Palantir Technologies, “Form S-1 (2020 IPO prospectus)”, SEC. : You can trace where the FDE role originated straight from a document Palantir wrote itself.
• INLEVEL9 Letter Issue 134 (the case study on Amazon’s FDE unit). : Today’s piece is the sequel — the same trend spreading to Microsoft.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

Footnotes

1. FDE (Forward Deployed Engineer): a role in which a vendor’s engineer is stationed inside the customer’s organization, directly fitting the product to that company’s work. Rather than handing over a product and leaving, the vendor builds alongside the customer’s team. ↩

2. Learning Gap: a state in which the AI tool fails to learn the organization’s context, and the organization fails to learn how to feed AI that context. The NANDA report points to this gap — not model performance — as the real reason enterprise AI pilots fail. ↩