The Rise of Onsite FDEs: Closing AI's Last Mile
When a powerful model dazzles in a demo but stumbles on real business data, the industry has found a very human answer: embed engineers directly at the customer site, writing production code against real data from day one, and stay until the system is actually used. That role is the Forward Deployed Engineer, or FDE — and in 2026 it has gone from a niche title to one of the most sought-after, best-paid jobs in AI.
Why FDE is taking off now
The concept is not new. Palantir has treated it as a core method since 2005: engineers work on the problems customers are actually facing, then feed field discoveries back to the core product team — a process the company calls "human backpropagation" in software development. The job description sounds familiar — go on-site, understand requirements, connect data, modify systems, stay until the project runs. Consultants and on-site developers have done similar work for years.
What changed in 2026 is that the bottleneck in enterprise AI has shifted from model capability to deployment. A model can ace benchmarks and still fail against messy real-world data, legacy systems and business edge cases. Companies bet on FDEs to do one thing: turn delivery that depends on individual experience into reusable software capability.
The hiring boom: 800% to 1000% growth
The numbers tell the story. Indeed reported that FDE-related job postings grew more than 800% year-over-year in the first nine months of 2025. By 2026, Perspective AI's analysis of roughly 1,000 live postings showed FDE hiring up more than 1,000% year-over-year, with compensation clustered at $300K–$550K total comp and senior roles at frontier labs clearing $1M+.
The giants are voting with real money. In May 2026, OpenAI announced a company dedicated to helping enterprises deploy AI and planned to acquire the applied-AI consultancy Tomoro, bringing in about 150 engineers and deployment specialists. The new company received over $4 billion in initial investment, with McKinsey, Bain and Capgemini among the participants. Anthropic, Google Cloud and Cohere are hiring at scale too. China is equally active: Minglue Technology has converted about 30% of its delivery staff into FDEs, COSCO Shipping is building its own FDE teams, and Softton, Kingdee and Seeyon have written FDE into their delivery systems.
A classic case: General Mills' supply chain
Palantir's supply-chain project with General Mills offers a representative sample. The food company connects 4,000 suppliers and more than 200 plants in North America, handling about 1.2 million orders a year. The two sides integrated 200 master and operational data tables to build the ELF supply-chain decision system. It checks about 3,000 orders daily and provides around 400 adjustment suggestions, more than 70% of which are accepted by staff; General Mills says the system saves about $40,000 a day.
The point is not the algorithm but the FDE's scope of responsibility: the system must understand orders, capacity, cost and business constraints, and know which suggestions can be executed and which decisions must stay with humans. Consultants can propose directions; traditional implementation engineers can connect databases. The FDE turns a fuzzy problem into a production system and judges success by adoption rate and business metrics.
China's playbook: turning experience into Skills and Agents
Chinese players take different approaches. Minglue Technology is converting its delivery teams — about 30% in the first half of the year — into FDEs, pushing its Pulian Software team toward state-owned enterprises. Where project experience used to live in consultants' heads, the company now tries to encode it into Skills and Agents so the next delivery depends less on individuals. Softton promotes an "AI Factory + industry experience + FDE" model covering consulting, model deployment and agent development. In August 2026, Sunshine Smart Operation and Kingdee launched a unified intelligent platform project, starting with contract, invoice and bid-document review, building an "AI first pass, human review" workflow.
The debate: is FDE just outsourcing with a new name?
As FDE goes mainstream, skepticism follows. Three questions separate a real FDE from a rebranded consultant: Does the engineer write production code? Is the engineer accountable for actual adoption? Does experience from one customer feed back into the standard product? Without the last one, a fancy title cannot change the nature of a project-based business.
Trust and permissions are unavoidable. External teams should receive the minimum information needed for a specific task: dev and production environments isolated, sensitive fields graded, accounts authorized per project, code merges approved internally, operations logged, and access revoked when the project ends. Mature enterprises must keep internal data owners, business owners and security teams. External FDEs can enter authorized rooms, but who opens which door, how business rules are explained and whether a system goes live remains an internal decision.
The best FDE project lets the FDE leave
The industry is converging on a view: external FDEs will exist for a long time because models and tools keep changing, but in any given business, an FDE should not be a permanent role. Once a project matures, daily operations return to internal teams, common needs enter the standard product, and low-level data transformation and troubleshooting get absorbed by AI — Palantir already offers an "AI FDE" that operates Foundry, manages codebases and edits data transformations.
A project is truly complete when three things hold: the system keeps running after external engineers leave, the internal team can modify it independently, and the vendor can serve the next customer without starting over. If on-site teams keep growing, contracts keep lengthening, and customers cannot run without those people, then the AI industry has simply given traditional outsourcing a more expensive name.
What it means for practitioners
The FDE wave is rewriting the AI practitioner's skill set. Perspective AI found the most-listed skills in FDE postings are no longer Python and SQL but customer discovery, problem decomposition and AI product judgment; Stack Overflow's 2026 survey shows about 41% of AI engineers now spend more than 30% of their time facing customers. When "getting a model to run" stops being a scarce skill, "getting a model used" becomes the new career moat — perhaps the most lasting impact of the FDE wave on the whole industry.