Microsoft Sent 6,000 Staff Into Client Offices · Issue #197

Opening

On 7/2 (July 2), Microsoft said it would spend $2.5 billion to create a separate organization called “Frontier Company.” Its job is simple: put 6,000 employees inside customer companies.

2 days earlier, AWS committed $1 billion to do the same thing. Before that, in May (the 5th month), OpenAI and Anthropic each created similar organizations. Anthropic went even further, setting up a $1.5 billion joint venture with Blackstone and Goldman Sachs.

At a time when stories about people losing jobs to AI arrive every day, the companies selling AI are spending trillion-won-scale sums to hire people and seat them in other companies’ offices.

Let me start with the conclusion. I do not think we should read this as a heartwarming twist: “Even in the age of AI, humans are still needed.” This is the vendors’ admission that AI models are not finished products yet. And that admission has opened a window with an expiration date.

📊 FDE, the role whose hiring grew 10x in 12 months

The role is called FDE¹. In plain English, that means Forward Deployed Engineer, a term borrowed from the military. Palantir began using it when it sent its own engineers directly to U.S. military bases in Afghanistan. Judson Althoff, Microsoft’s commercial business CEO, has also acknowledged that Palantir popularized the title.

The core idea is this: do not sell the product and leave. The vendor’s engineer embeds with the customer’s team, studies the company’s workflow up close, and builds a system around it.

The numbers show how large this shift has become. According to aggregated hiring data, FDE job postings were up by more than 1,000% year over year as of early 2026. Compensation is aggressive, too. Senior FDE total compensation at frontier labs such as OpenAI and Anthropic sits in the $450,000-$550,000 range, while staff-level roles exceed $600,000. That is 2-3.5x the level of Palantir’s traditional FDSE roles.

One caveat is worth making. This compensation data comes from hiring platforms and community self-reports, so the top end is probably over-sampled. Even so, the direction is clear. The market is paying a premium for this role.

💰 Why would vendors pay to send their own people?

That raises an obvious question. If the product is good, should customers not adopt it on their own? Why spend your own money training engineers and embedding them, free of charge, inside someone else’s company?

The answer lies in failed adoption.

The most useful document for understanding this shift is a 2025 report from MIT Media Lab’s NANDA project. Its central claim is that about 95% of enterprise generative-AI pilots are failing to produce measurable profit-and-loss impact. For a while, that number was consumed in the press as proof that “AI is a scam.” But the original report is more nuanced.

The failure mode it identified was not model performance. It was the learning gap². The model does not understand the organization’s workflow. The organization does not know how to give the model context. And there is no one to bridge the gap. The company has bought a tool, then left it stranded without any understanding of what is actually happening inside the firm.

The methodology has limits, too. The study combines roughly 300 public cases, 52 executive interviews, and 153 survey responses; it is not a peer-reviewed paper. It would be hard to treat the “95%” figure as precise statistics. But vendor behavior supports the diagnosis. In an interview, Microsoft’s Althoff said customers are “all at very different places” and are still trying to figure out how to handle AI.

In other words, the FDE boom is not evidence that AI is succeeding. It is the invoice for failed AI adoption. Vendors have started paying that invoice themselves.

Microsoft also has a sense of urgency. Its stock is down 21% this year, the worst performance among megacap technology stocks. Copilot has not taken hold in the enterprise market as strongly as expected, and GitHub Copilot has ceded share to later entrants. The company can no longer explain away “the model is good, so why is no one using it?” as the customer’s fault.

🔑 But this job succeeds only by eliminating itself

Here is the real twist.

AWS stated a clear principle when describing its own FDE organization. Teams of 5-6 engineers go in for roughly 45-day cycles, with the goal of leaving the customer self-sufficient by the end.

Read that sentence again. FDE is a job whose success condition is making itself unnecessary.

That is the opposite of traditional consulting. Consulting turns relationships into retainers. FDEs go in so they can get out.

This is also how I read Microsoft’s decision to make this a separate profit-and-loss-bearing “Company,” rather than a consulting division. The point is not to grow services revenue. It is to raise platform adoption. For context, Microsoft’s existing enterprise and partner services revenue is about $2.1 billion per quarter, growing 2.5%. That means this is not where Microsoft is trying to make money.

So this hiring boom is not permanent job creation. It is transitional infrastructure. These people build bridges. Once the bridge is built, the bridge-builders have to move somewhere else.

Oswald’s Lens

My answer to Oswald’s question, “What kind of attitude and posture helps a person survive?” is this.

The person who survives is not the person who uses AI well. It is the person who can translate between AI and reality.

If you dissect FDE job postings, the required skill is not just coding. It is the ability to read a customer’s business processes, identify what data exists where and in what condition, and push through the organization’s politics until the system is actually deployed. It is the role of filling the learning gap I mentioned earlier through human work.

I saw exactly this pattern again and again while building the Notion Korea community, integrating AI products at Kakao Brain, and designing GTM strategy at Gamma. Products almost never sold simply because they were good. They sold only when someone created the context between the product and the customer’s organization. The bottleneck was not technology. It was context. What big tech is now buying with trillion-won-scale money is precisely that capacity to produce context.

But I would add one thing to Oswald’s original frame. If you read this hiring boom as reassurance that “there will still be jobs for people after all,” you are only half right. FDE is a job that designs its own disappearance. If AWS is targeting self-sufficiency after 45 days, this role is structurally temporary.

So the real question is not “Can I get into this job category?” It is “What will I accumulate while this window is open?” Knowing how to use tools is not accumulation. What accumulates is domain context, and the experience of translating that context into systems. You can carry that with you to the next layer.

Closing

To sum up, there are 3 points.

First, the FDE hiring boom is not a sign that AI is working well. It is a sign that adoption is not working. Vendors have started paying the cost of that failure themselves.

Second, the bottleneck is not model performance. It is organizational context. That is why the thing rising in value is not coding ability, but translation ability.

Third, this role is not permanent. Its success criterion is making itself unnecessary. So this period should be seen not as a “job opportunity,” but as an “accumulation opportunity.”

If you try just 1 thing this week, I recommend writing 1 page on “What would I need to explain before handing this work to AI?” That 1 page is what the market is now paying $450,000 to buy.

💬 If you have worked on an AI adoption project in practice, where did it get stuck: model performance, or the organization, data, and process? Leave a short comment with which side it was. If enough cases come in, I will organize them by type in a future issue.

📨 If you have a colleague thinking about AI adoption, please forward this piece.

(Web version) 📬 Subscribe · I write a weekly newsletter analyzing technology, economics, and the humanities together.

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 the Althoff interview. His remark that “customers are at very different places” is the most candid explanation for why this organization exists.
Microsoft, “Microsoft Frontier Company: AI engineering that amplifies and protects your intelligence”, Microsoft Official Blog, 2026-07-02. : Read the original if you want to see how the vendor itself frames this organization. The tone differs sharply from the CNBC article.
CNBC, “AWS puts $1 billion into new AI unit to embed engineers with customers”, CNBC, 2026-06-30. : The 45-day cycle, 5-6-person team, and customer self-sufficiency at the end are the 3 most important pieces of evidence for today’s essay.
MIT Media Lab NANDA, The GenAI Divide: State of AI in Business 2025, 2025. : The concept of the “learning gap” matters much more than the “95% failure” figure. Read it with the caveat that it is not a peer-reviewed paper.

Background

Palantir Technologies, “Form S-1 (2020 direct-listing prospectus)”, SEC. : This is Palantir’s own document, and it shows where the FDE role came from.
Oswald’s Knowledge Talking issue 134 (the Amazon FDE organization case). : Today’s piece is the sequel, after that trend spread to Microsoft.

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis, including business data management and business analytics, while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master's from Korea University's Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

FDE (Forward Deployed Engineer): A role in which a vendor’s engineer stays inside a customer company and connects the product directly to that company’s work. Instead of handing over the product and leaving, the engineer builds together with the customer organization. ↩
Learning gap: A state in which an AI tool has not learned the organization’s context, and the organization has not learned how to give context to AI. The NANDA report identified this gap, rather than model performance, as the real reason enterprise AI pilots fail. ↩