Recently, a dramatic scene unfolded in the AI industry: OpenAI and Anthropic, long-time rivals, announced on the same day the formation of joint ventures for enterprise AI services, with an identical core strategy—sending Forward Deployed Engineers (FDE) into enterprises. OpenAI founded "The Deployment Company" with over $4 billion in initial investment, and immediately announced the acquisition of UK-based AI consultancy Tomoro, bringing roughly 150 senior FDEs directly under its wing and dispatching them to work on-site at client companies. Anthropic, meanwhile, partnered with Wall Street giants such as Blackstone and Goldman Sachs to form a $1.5 billion enterprise AI services joint venture, likewise adopting the FDE on-site model to drive deep deployment of the Claude model.
The world's two top AI companies have declared with billions of dollars: FDE is the role in urgent demand right now.
The data is even more striking. According to Indeed platform data, in April 2025 there were only 643 FDE-related job postings in the US market; by April 2026, that number had climbed to 5,330, a year-over-year increase of 729%. Recruiting firm Adecco notes that demand for FDEs is exploding at roughly 800% per year, while the talent supply is growing at only 50%, and the supply-demand gap keeps pushing compensation higher.
A 2026 survey of 1,500 FDEs by Perspective AI showed that the median total annual compensation for senior FDEs at frontier labs (OpenAI, Anthropic, etc.) reached $485,000, with senior staff level as high as $725,000, and total compensation at top labs falling in the range of $350,000 to $550,000.
The enormous compensation figures and explosive demand growth sketch the outline of a red-hot role. But to truly understand all of this, we need to return to the origin of the FDE role.
I. A Twenty-Year Evolution from US Military Intelligence Rooms to AI Companies
1.1 An Institutional Innovation Born of Necessity
The story of the FDE begins in 2003. That year, Peter Thiel, Alex Karp, and others founded the big data company Palantir Technologies. The company's first clients were US intelligence and military agencies, but they faced a seemingly unsolvable dilemma: clients could not possibly hand over their classified data, and the development team had no way of knowing what the software should look like.
The standard waterfall software development model failed. Requirements were written in classified documents, processes were hidden behind security clearances, and feedback could not be transmitted through normal channels. Palantir made a decision that seemed "clumsy" at the time: send engineers directly to the client's site to work side by side with users.
This is exactly where the military term "Forward Deployed" comes from—engineers were "deployed" to the front line. This model was initially a temporary tactical response, but its results far exceeded expectations. On-site engineers could not only reshape the software into what clients actually needed, but also discover, through interaction, needs that even the clients themselves had never clearly articulated.
1.2 From Temporary Arrangement to Institutionalization: The Birth of "Delta"
As the practice deepened, Palantir began to institutionalize this model. Around 2011, the company formally named these on-site engineers "Forward Deployed Engineers" (FDE) and gave them a resonant internal codename—"Delta" (Delta Force, the US military's special operations unit).
The choice of codename was highly apt. The Delta Force is renowned for accomplishing high-difficulty missions in extreme environments, and what FDEs must accomplish at client sites is likewise high-value problem-solving in complex, chaotic, information-incomplete environments.
Palantir also developed a classic "Echo-Delta" two-person collaboration model. The "Echo" is typically a domain expert from the client's industry, responsible for understanding the business and translating requirements; the "Delta" is an elite software engineer, responsible for rapidly turning requirements into working code. This design profoundly shaped the DNA of the FDE: he is neither a pure programmer nor a traditional consultant, but a fusion of the two.
1.3 Full Detonation in the AI Era
Palantir used this playbook to serve top clients such as the Department of Defense, the CIA, and JPMorgan Chase. After 2023, with breakthroughs in large language model capabilities, the FDE model was quickly discovered and adopted by a new generation of AI companies. The reason is straightforward: the situation AI companies face is almost a carbon copy of Palantir's twenty years ago—powerful technology in hand, but clients who don't know how to use it.
As Bob McGrew, former Palantir executive and former OpenAI Chief Research Officer, put it in retrospect: "You go to a new customer with a product, and the problem they want you to solve is one you've never solved before. The FDE takes the existing product and, with the help of the product team, figures out how to deliver that outcome, in a way that genuinely works for the customer."
This capability is precisely the scarcest link in enterprise AI adoption. Between 2023 and 2025, according to LinkedIn data, 8,500 new FDE positions were created in the US, and the FDE spread from a "Palantir-exclusive practice" to a collective bet across the entire industry.
II. A Cold Look: Do Enterprises Really Need FDEs?
Despite unprecedented market enthusiasm, behind the FDE boom lies a fundamental question that cannot be avoided: is this a genuine enterprise need, or supply-side creation driven by the wishful thinking of AI companies? If we examine it from the standpoint of a corporate CFO or CEO with the most dispassionate eye, the answer is far more complicated than it appears on the surface.
2.1 The Math Doesn't Add Up: An Adventure Without ROI
Corporate investments require measurable returns, but the benefits AI brings are highly nonlinear and unstructured. AI may significantly improve customer satisfaction, yet the financial model linking satisfaction gains to profit growth is murky. By contrast, the direct costs of adopting AI—API call fees, compute resources, the high personnel costs of FDEs—are clear, immediate, and heavy. When a CFO faces a proposal of "spending millions more each year with effects that cannot be accurately measured," rejection is a highly rational choice.
2.2 Organizational Rejection: AI Is a Transplanted Organ, Not an External Device
Embedding AI deeply into core business processes has an impact far beyond the technology department itself. It will fundamentally change employees' workflows, performance evaluation methods, and even power structures. An AI-powered intelligent scheduling system could, overnight, make a production manager with twenty years of experience—who schedules production based on intuition and authority—feel that their core value has been dissolved. This is not a technical compatibility problem, but an organizational transformation, and the difficulty and risk of driving it should not be underestimated.
2.3 Legacy Baggage: Silent Existing Investments Are Formidable Barriers
Over the past ten or even twenty years, enterprises may have invested millions in certain core systems. These systems carry all the core data and business processes—perhaps outdated, but stable. Migrating them to an AI platform means not only that previous investments go down the drain, but also a prolonged and painful migration period and unpredictable business disruption risks.
2.4 Trust Deficit: 0.1% Hallucinations, 100% Disasters
Enterprise core production environments demand deterministic results, while the nature of large models is probabilistic. Even if they perform well 99.9% of the time, that 0.1% of "hallucinations" can cause catastrophic consequences in scenarios like finance, healthcare, and law. Gartner predicts that by the end of 2026, organizations will abandon 60% of AI projects, with one of the main reasons being insufficient data quality and reliability. No matter how strong an FDE is, they cannot fundamentally turn a "black box" into a "white box."
2.5 Strategic Confusion: Without Knowing Where the Disease Is, You Can't Prescribe the Cure
This "special forces" unit of FDEs can only deliver maximum value when a company's leadership has a clear directive on "which hill we need to take." Yet most enterprises are still in a vague state of "knowing they should transform, but unable to articulate where the core pain points lie." If a company doesn't know what problem it most needs to solve, the FDE's problem-solving firepower has nowhere to be unleashed, and the hefty investment becomes nothing more than an expensive experiment.
A more cautionary data point comes from Gartner: Gartner senior analysts predict that by 2028, "70% of enterprises will be forced to abandon AI agent solutions led by FDEs, because vendor costs are too high and enterprises lack the internal skills to independently evolve these systems." In other words, even if a company accepts the FDE model, if it fails to digest and transfer the capabilities during the process, it may ultimately be left high and dry once the FDEs leave.
III. The Difficulty of Execution: Even If You Truly Need FDEs, Implementation Is Never Easy
Suppose a company has navigated all the fog above and made a clear decision—it needs FDEs and is willing to pay for them. Even then, the execution process is far from smooth sailing.
3.1 On-Site Immersion and Diagnosis: Cracking the Code of "Unspeakable" Knowledge
The first challenge FDEs face upon entering a company is not a technical one, but acquiring "tacit knowledge." A company's most core operational logic often resides not in documents, but in the minds of veteran employees—a kind of "feel" that cannot be captured in writing. For example, a senior auditor can "intuitively" judge that a certain transaction is risky, but cannot explain the basis for that judgment. FDEs need to act like anthropologists—through long-term observation, building trust, and even personally operating the business—before they can translate this tacit knowledge into quantifiable logic that AI systems can understand.
At the same time, FDEs face a dual psychological game from day one: management often holds the inflated expectation that "AI can do everything," while frontline employees harbor silent resistance fueled by "AI will take my job." Before diagnosing technical problems, FDEs must first become astute organizational communicators—pulling unrealistic expectations back to reality and converting resistance into collaboration.
3.2 Evaluation and Alignment: Finding Consensus Amid Uncertainty
The core question FDEs must answer is: What value can AI actually create for this company? But the awkward reality is that many of AI's benefits (such as improving decision quality or enhancing customer experience) simply cannot be precisely quantified at the proof-of-concept stage. Clients demand to see ROI before investing, yet true ROI only emerges after large-scale deployment—a classic chicken-and-egg dilemma.
Meanwhile, there are multiple decision-makers inside the company: the CEO looks at strategy, the CTO at architecture, the CFO at cost, and frontline managers at whether processes will descend into chaos. FDEs need to write four evaluation reports with different emphases, in four completely different languages, for these four roles, and drive a difficult consensus-building meeting. Industry data from 2025 confirms this predicament: 92% of enterprises have already deployed AI agents in core business operations, but only 23% have achieved scaled adoption—a large number of companies are stuck at the "proof of concept done, but can't push it forward" stage.
3.3 Deployment and Integration: Navigating the Minefield of Legacy Systems
The deployment phase is the "dirtiest" work in the entire FDE execution process. A company's IT environment is a "legacy system" built up in layers over decades—APIs lacking documentation, outdated database drivers, scripts written by some long-departed IT manager, and so on. A seemingly simple "data call" may hide compatibility issues that take weeks to resolve. According to reports, 52% of professionals cite data quality and availability as their biggest AI adoption challenge—a proportion exceeding the combined total of skills gaps, regulation, and change resistance.
Even after a system goes live successfully, the real battle has only just begun. AI models can continuously produce biased results without anyone noticing—by the time such "silent failures" accumulate to the point of discovery, substantive damage may already have been done. Therefore, the endpoint of deployment is not system launch, but establishing a mechanism that enables frontline employees to understand, trust, and effectively intervene in AI decisions. This requires building multi-layered safeguards at the prompt layer, the architecture layer, and the application layer to ensure AI outputs always remain within safe boundaries.
3.4 The Cost Code of FDEs: Why Scaling Is So Hard
Finally, there is a dimension that is often overlooked: the economic model of FDEs themselves. Palantir's practice shows that when entering a new client for project deployment, the initial phase may be loss-making; as the product becomes better suited to that client's needs through continuous on-site exploration, delivery costs gradually decline, and it takes about a year for margins to turn from negative to positive. This means the FDE model inherently resists "rapid scaling"—it requires time and patience, which is precisely the quality the capital markets lack most.
IV. How Should Companies Position Themselves, and How Can Individuals Grow?
On one side, AI companies are paying sky-high prices to snatch up talent, with job demand surging sevenfold in a year; on the other, companies harbor deep reservations and face challenge after challenge in execution. Facing this reality, how should different market players act?
4.1 For Companies: Strategic Clarity Matters More Than the Posture of "Embracing AI"
(1) Companies Suited to "Testing the Waters" Now: Those with Structural Advantages
Not all companies are suited to introducing FDEs at this stage. Those where AI deployment can deliver immediate results typically share the following structural characteristics:
- Natively digital processes: Such as cross-border e-commerce, SaaS companies, and digital marketing agencies—core business processes are naturally digital, data is clean, interfaces are standardized, and AI deployment encounters the fewest interface conflicts.
- High fault tolerance for tasks: Such as content creation, general customer service, and ad creative generation—AI output is highly subjective, a 1% error rate won't trigger immediate disaster, and manual review costs are manageable.
- Extremely short value chains: The link between AI-driven efficiency gains and revenue growth is direct—for example, AI generates 100 ad copies and click-through rates rise by 20%, making the ROI obvious at a glance.
Such companies can adopt a "small incision, deep well-digging" strategy: identify a single business point with clear ROI and ready data, set explicit quantified goals and a 6-to-8-week sprint cycle with the FDE team, and quickly close the loop on value validation.
(2) Companies Suited to "Watching with Preparation"
Most companies fall into this camp—finance, manufacturing, healthcare, government, and so on. They are not rejecting AI, but rather have rigid requirements around reliability, compliance, and organizational stability. These companies should start three "preparatory actions" now:
- Launch a data asset inventory: Map out which systems core data is scattered across, what its quality is, and whether it can be accessed. This is infrastructure work that cannot be avoided whenever AI is eventually adopted.
- Make decision logic explicit: Require core business departments to document, as much as possible, the tacit rules behind key decisions, drawing a "mine-clearing map" for future FDEs.
- Conduct organizational immunity tests: Introduce AI tools on a small scale, with the primary purpose not being efficiency gains but observing the organization's real reactions—where does employees' fear come from? Which department shows the strongest resistance? This requires more advance preparation than the technology itself.
(3) A Common Warning for All Companies: Beware of "FDE Dependency"
The ultimate goal of FDEs is to make themselves no longer needed as quickly as possible. Companies must explicitly require FDE teams to carry out systematic knowledge transfer—handing over code, documentation, and decision logic in full and training internal teams—to ensure that after external FDEs leave, the company has the capability to independently operate and iterate on its AI systems. Otherwise, the FDE model will slide from "service-led growth" into a "service-dependency trap"—which is precisely the root of the 2028 crisis Gartner warns about.
4.2 For Individuals: How to Become an AI "Special Forces Soldier" Who Navigates the Fog
For individuals aspiring to a career in FDE, this is an exhilarating opportunity for career reinvention. But the combination of skills this role demands is far more complex than most people imagine.
Layer One: The "T-Shaped" Fusion of Technology and Business
An FDE must first be a qualified software engineer—proficient in Python, familiar with large model application frameworks (such as LangChain and RAG architectures), and well-versed in containerized deployment and API integration. But beyond technical depth, they also need to expand into business breadth: building a framework for quickly understanding business models, developing deep knowledge of one or two target industries, and cultivating the habit of asking "why build this feature" rather than merely "how to build this feature." A Perspective AI survey of 1,500 FDEs found that top FDEs spend 2.4 times more time on structured customer research than the bottom performers—a difference that drives a bigger performance gap than engineering experience or AI model proficiency.
Second layer: growing through the "friction" of the real battlefield
What truly sets an FDE apart from an ordinary engineer is that kind of "hard-to-articulate" on-the-ground experience. How to precisely translate a vague business complaint into a clear technical task, how to negotiate unreasonable deadlines in client meetings, how to explain to non-technical people why data quality affects model performance—these skills can only be accumulated through repeated real customer interactions. For engineers who want to enter this field, the most effective growth path is not taking more technical courses, but seeking out hands-on opportunities at customer sites as early as possible.
Third layer: from solving problems to defining problems
This is the watershed capability of an FDE. When a client says "I need a report," a junior engineer will simply build the report, while a mature FDE will ask: "What business question do you want this report to answer?" They may then discover that what the client really needs is not a report, but a real-time risk alerting system. This ability to probe the essence and redefine the problem marks the FDE's elevation from senior technician to strategic partner.
The changing demands on FDEs are a microcosm of AI's journey from the lab to industry. When technological complexity outpaces the market's ability to absorb it on its own, the "last mile" becomes the most critical link in the entire value chain. FDEs stand precisely at this historic juncture: the arms race in model capabilities is giving way to a positional war over deployment and integration capabilities.
For enterprises, this means abandoning the fantasy of "buy and it works" and preparing for a profound organizational evolution. For individuals, it means an exhilarating career reinvention—becoming the person who can master both code and human psychology, and bridge the gap between technology and reality. Whether the springtime of the FDE truly arrives depends not on how strong AI companies' hiring demand is, but on whether this role can genuinely help enterprises make the leap from "technically feasible" to "commercially successful."