Frontline Case Studies

From Palantir’s Kandahar origin and Chinese enterprise samples to first-hand failure logs: how the FDE playbook works in a real organization, and how it breaks.

FDE is not a paper concept. It grew out of a real war and has been re-tested in one enterprise deployment after another. Put the cases side by side and the same formula keeps appearing.

After reading, you should be able to answer:

• How Palantir turned “send engineers into the field” into a repeatable model
• Where Chinese enterprise samples differ from the overseas playbook, concretely
• What the success formula is that recurs across the cases — the ones that worked and the ones that collapsed

The origin: servers in Kandahar

Palantir was the first company to make this a model. In 2008–2009 the engineer Mark Scianna was sent to a US brigade headquarters in Kandahar, Afghanistan — not to demo anything, but to deploy servers on site, build the data integration, and train the intelligence analysts. The insight you get sitting in an operations room is one a Palo Alto conference room never produces.

It was later named. Inside the company, forward deployed engineers are Deltas (NATO phonetic, with the special-forces connotation intact), the people who own the mission and the adoption path are Echoes / Deployment Strategists, and the platform builders are Dev. The distinction fits in one line: Dev serves many customers with one capability; Delta takes on many problems inside one customer. By 2016 Palantir’s FDEs briefly outnumbered its traditional software engineers. This was not a fire brigade — for those years it was the main go-to-market engine.

The bootcamp is this model in productized form: the customer arrives with real data, Delta and Echo embed for a few days, a working prototype gets built on site, and executives click through it themselves. It replaces the most expensive stretch of conventional pre-sales — the feasibility study and the deck a16z-palantirization. The Trust Dividend covers why this manufactures trust in a single pass.

Local samples: what deployment looks like in China

Move the camera to China and the main battleground is different: buyers are mostly enterprises and government, deployment favors private installation and the domestic stack, and value moves from “pilot” to “core productivity” — those rules are laid out on their own in China’s Deployment Rules. The samples below all come from 2024–2025 public reporting, and they build the intuition for what an FDE looks like here.

iFlytek: the FDE playbook behind the “China’s Palantir” label

iFlytek writes delivery into its method: go deep on the customer site, define the problem together, and turn the experience into a reusable platform. Its medical assistant now runs routinely in 680+ districts across all 31 provinces — 930 million AI-assisted diagnoses and 360 million standardized medical records. The Kunlun model built with CNPC reaches 300 billion parameters, speeds up seismic wave-equation solving more than tenfold, and is live in 100 scenarios. The Qingtian model built with Hefei is China’s first model dedicated to public resource trading, covering 24 engineering procurement categories, and was picked by the NDRC as its only “AI+” demonstration.

State Grid: the Guangming power model inside core operations

China’s first hundred-billion-parameter multimodal model for the power industry. Its marketing and supply-plan agent automates the whole “intake to plan drafting” flow; its highway emergency-command agent, live on the Jingxiong expressway, raised warning accuracy above 95% and cut emergency response from hours to minutes.

Shandong Haihua: a salt-chemical plant with “zero manual”

Working with Inspur on the Haiyue foundation, the plant built a salt-chemical process-control model. Phase one cost just over RMB 32 million and returned more than RMB 23 million in the same year: unplanned chlorine-compressor downtime fell from four to zero, maintenance cost from RMB 800K to RMB 200K, 4.5 GWh of electricity was saved, process stability rose 55%, key automatic-control coverage rose 61%, manual operations fell 68%, and 1,803 process steps were reduced to 400.

China Energy: AI into procurement and trading

After the smart bid-evaluation tool joined the workflow, total benefit reached RMB 1.9 billion. Human monitoring of power trading went from 1.5 hours to 15 minutes, per-kWh revenue improved 5%–10%, and about RMB 4 million of new income followed.

What these projects share is that none of them is “selling software”; all of them are “doing delivery” — embed in the customer’s business, define the real pain precisely, solve it with technology, and turn what remains into reusable platform capability. That is the FDE core described earlier.

What actually fails on the ground

Beyond the successes, more projects fail. The reason to put failure logs on the table is not to scare you; it is to know which holes someone else has already fallen into before you start work.

Strategy misread: one brokerage’s hundred-billion-parameter project

Launched in 2024 to “replace 80% of research analysts with generative AI” on a hundred-billion-parameter model, with compute eating 40% of the total IT budget. Regulators require human review that cannot be removed, so the ROI never closes. The board cancelled it in March 2025, the A100/H800 cluster was resold at a discount, and the book loss was RMB 110 million. The error was treating model capability as business value.

Technology and business out of sync: RMB 2 million, zero go-lives

A panel plant in East China spent RMB 2 million on a quality-inspection model. The engineering team’s definition of smart inspection was pixel-level defect segmentation at mAP above 0.95; the production line needed a miss rate below 50ppm at under three seconds per panel. The two KPI sets never lined up, the project froze, and the vendor’s final payment went unpaid. Requirement translation loss is the most expensive hidden cost in delivery.

Organizational island: 38% attrition in one AI center

A consumer goods group created an “AI center of excellence” in 2024 and hired 30 algorithm engineers. The business units treated it as an outside vendor and queued requirements for three months; the algorithm team did not know the POS or DMS data structures and rebuilt features four times. Attrition hit 38% in Q1 2025 and delivery rate stayed under 20%. The AI department had become an island prospecting for pilots.

The most expensive toy: RMB 200 million of idle GPUs

A company spent RMB 200 million on GPUs “for future demand”; six months later it was still running proofs of concept with different suppliers, with nothing to show. Industry retrospectives call this the “GenAI trap”: staying on a single use case so long that procurement becomes the goal. One domestic survey adds a colder average: RMB 23 million invested in a large-model project, and an average lifespan of 8.4 months. Numbers from anonymous samples are magnitudes, not conclusions.

Overseas data is the same thermometer: MIT’s NANDA study is the sharpest — 95% of generative AI pilots never break out of their original department into production. Fluid Labs grouped what the survivors do into five practices: map the process before choosing technology, treat data readiness as a precondition, release under supervision, design for adoption, and start from one narrow workflow.

None of this is “the technology does not work”. Each failure lands exactly where the FDE role was invented to stand: fill the process, data, adoption, and scope holes before writing the first line of code.

How the industry defines the role

Job descriptions vary a lot from company to company, but the people actually hiring, doing, and organizing this role define it in strikingly similar terms — it is not “a salesperson who can code”, it is an engineer who brings a product into the field. Bloomberry parsed 1,000 FDE postings, and the most telling number is this: the share carrying a sales quota is 0%. The role owns the result without owning a quota bloomberry-fde-jobs.

Anjor Kanekar, seven years a Palantir FDE

They worked on Airbus assembly lines and in air-gapped environments. “FDE is not only delivery — what you learn on site has to flow back into the product roadmap.” That is the root difference from outsourced implementation: implementers leave when the handover is done; an FDE turns what the customer site taught them into the product’s next feature. Takeaway: FDE is a product-discovery function, not on-site outsourcing.

Colin Jarvis, Head of FDE at OpenAI

They describe OpenAI FDE as three phases: Scoping → Validation → Delivery. The core judgment is “turn what you found at the customer site into repeatable method and long-term product capability, not a one-off fix”. The same logic has a customer-side counterpart: Morgan Stanley built its scoring set first and used retrieval hit-rate as the gate before rolling the assistant out to advisors openai-morgan-stanley. Takeaway: the FDE is the product scout closest to the real problem, and what scouting produces has to pass evaluation first.

Leo Mehr, head of FDE at Ramp

Ramp built an FDE pod of about 15 people, and the job post says outright “win deals with sales” and “drive the core product roadmap”. The hardest of its four principles: “find the highest-impact problem and do not stop until it lands.” Takeaway: the role carries revenue and feeds product; it is not a support function.

LayerX: a business model rather than a role

LayerX’s group engineering manager for its FDE organization makes a sharper call: “FDE is closer to a business model than to a role. The decisive difference from on-site contracting is that you bring a product — the problems you solve turn into product features that get reused at the next customer.” Takeaway: FDE is not staff augmentation; the difference is whether something survives in the product.

A first-hand Chinese big-tech posting (Volcano Engine)

The published requirements are concrete: make technical decisions on site without layer-by-layer approval from headquarters; leave behind at least one cross-customer reusable asset per project (code, skills, prompt templates, or eval sets); act as the field proxy for product R&D, feeding gaps in the foundation back in structured form. Takeaway: at Chinese big tech the FDE is isomorphic with Palantir’s Delta.

Accenture’s RDE and the ArchSynapse AI column

The official Accenture–Anthropic partnership announcement is blunt: roughly 30,000 consultants trained on Claude, including “reinvention deployed engineers” whose job is embedding Claude into client environments accenture-anthropic. The title changed name; the kernel is the same — people go on site and integrate against the customer’s systems. The Chinese column ArchSynapse AI (27 lessons) holds the line on sourcing: primary material first, every number attributed, cases traceable. Takeaway: FDE has moved past the AI labs and become an organizational design pattern across products and consulting.

Read these definitions together with One Person, Four Battlefields and How FDE Differs from Pre-Sales, Consulting, and Implementation: whatever the company — OpenAI, Palantir, Ramp, or a Chinese big tech firm — the underlying definition is the same. Write production code inside the customer’s systems, embed, own the result, and feed field experience back into the product. Put the difference to the test of three questions — who carries revenue, who writes production code, who stays to maintain — and FDE separates cleanly from consultant, pre-sales, and outsourced implementation.

The repeatable formula

Take the cases apart and the same formula recurs, mapping one to one onto method in other parts of this site:

• Connect real data and running systems, instead of wrapping a prompt — the agent goes straight to accounts, trading APIs, and internal knowledge bases; a wrapper around a prompt has to be rebuilt the next time the model changes.
• Pick a high-frequency, repeating, measurable workflow first — the math has to be clear enough that a CFO will defend it.
• Grow governance on the use case, not on the model — guardrails belong on accounts and workflows, not on the model’s good intentions.
• Treat adoption as infrastructure — training plus opt-in creates internal demand; a mandate from above does not.
• Move people up into judgment, do not remove them — that is where read the old system, write the new one and launch is not activation both land.

One last number worth remembering: in MIT’s NANDA statistics, vendor-led deployments succeed about 67% of the time, while purely internal builds manage roughly a third of that. The gap is not the model — it is whether someone actually walked into the field and filled the cracks. That is also the reason the role exists.