AI FDE in China

A fictional story about an embedded AI engineer in a Suzhou factory, and the contracts, shared teams, reusable assets, and exit tests that keep FDE work from becoming staff augmentation.

Jul 18, 2026

Vibe Coding Team

Linjiang Equipment, Zheng Yu, Tang Jing, and the operating numbers in this story are fictional. The role definitions, industry guidance, and regulatory requirements come from the public sources listed at the end. The illustrations were generated with AI.

At 6:47 on a wet Suzhou morning, Zheng Yu stood outside Gate Two of Linjiang Equipment. The security guard handed him a laminated badge. It said: Project Consultant.

He clipped it to his shirt and felt the first warning.

Twelve people were waiting upstairs. The head of IT asked about model parameters. Procurement asked how many engineers would be on site each day. Tang Jing, the production executive, waited until last.

"Will your team take all the requirements?" she asked. "We are short-handed."

The project had not started, and AI FDE was already standing at the door of outsourcing.

The visitor badge

A Forward Deployed Engineer enters a customer environment and turns model capability into a production system. OpenAI's current FDE role owns discovery, technical scoping, system design, building, and production rollout. Success is measured through production adoption, workflow impact, and evaluation feedback. Palantir's description starts with an open business question, then asks engineers to build a working answer beside the customer.

The role sits close to consulting and requires plenty of custom code. That makes the failure mode easy to recognize. Procurement buys people by the day. Business teams drop requests into a group chat. The FDE spends daylight in meetings and nights patching integrations. Three months later, the chat has two thousand messages. The customer still cannot explain the system, while the vendor has created a permanent on-site unit.

When AI FDE becomes ticket taking and integration patching, the project turns into a maze no one wants to touch

There is one useful test: after the FDE leaves, can the customer run, change, and extend the system?

If not, the job title changes nothing.

The first measure

Tang pushed the first request across the table. "Build an intelligent production scheduler. Two months to launch."

Zheng did not accept it. He asked to visit the line.

Beside the whiteboard on Line Three, he met Wu, a production planner. She barely used the scheduling system. A folded grid notebook sat beside her keyboard. When an urgent order arrived, she checked tooling time, called the warehouse, and asked quality control which batch could be released. The software held data. The operating rules lived in people, phone calls, and that notebook.

"You want a model to replace me?" Wu asked.

"Not yet," Zheng said. "First it should find every exception and put the options beside their evidence. You still make the call."

That conversation changed the project. The team stopped building a scheduling chatbot and started reducing exception-handling time. They assembled two hundred fictional historical cases with constraints, actions, and outcomes. Before launch, they agreed on concrete tests: reduce median handling time from a fictional baseline of 46 minutes to under 15; require planner approval for high-risk changes; trace every recommendation to orders, inventory, and equipment state.

Model quality mattered, but business change came first.

China's official guidance puts business first too. The Ministry of Industry and Information Technology's manufacturing AI application guide tells companies to assess their starting point, select high-value scenarios, set priorities, and define measurable indicators. Its examples include yield, false-negative rates, response time, and maintenance cost. They all live on the operating floor.

A practical deployment contract can fit on one page:

Run an eight-to-twelve-week deployment cycle instead of renewing headcount forever.
Test business outcomes, system behavior, and real adoption together.
Replay historical cases and conduct a live trial before production release.
Define the exit conditions, asset ownership, and internal owner on day one.

That thin page blocks a surprising amount of outsourced behavior.

Do not take the job

In week two, Tang assigned an internal owner named Chen Min. Chen managed MES integrations. He knew little about language models and disliked the word transformation.

Zheng was relieved.

They worked in one repository. Zheng wrote the first order connector; Chen wrote the next. Zheng assembled the evaluation harness; planners supplied failure cases. Every Friday, someone from Linjiang ran the demo. The author of the slide deck did not matter. The person who could handle Monday morning's alarm did.

Outsourced delivery often leaves the customer in the meeting room and the vendor in the machine room. An FDE engagement has to remove that wall. Domain experts help define correct behavior. Internal engineers deploy, roll back, and inspect logs with their own hands.

This feels slower at first. Chen took three times as long as Zheng to change his first connector. In week six, the warehouse changed an interface. Chen did not open a vendor ticket. He updated the mapping, ran the regression set, and shipped it.

There was no launch ceremony. Zheng crossed "warehouse field mapping" off the list and wrote one note beside it: Linjiang owns this now.

Leave the toolbox

An FDE team should leave behind a scenario map, data connectors, an enterprise knowledge base, evaluation sets, observability, and permission and audit templates. Source code is only one item.

A successful AI FDE engagement hands data, knowledge, evaluation, and governance tools to the internal team

The MIIT guide describes enterprise knowledge in equally concrete terms: mechanism libraries, simulation libraries, experience libraries, data labeling, quality assessment, classification, and continuing improvement. It calls for live production trials and regular reviews of accuracy, latency, cost, safety, and business performance.

The product boundary follows from that work. Customer differences should live in configuration, data, and permissions where possible, rather than in a forked codebase for every account. When the same request appears again, it moves into the shared product backlog. Connectors get versions and tests. Prompt changes ship with evaluations. An incident becomes a guardrail for every deployment.

An outsourcing firm worries about the project ending. A product company wants repeated problems to disappear. FDE teams belong on the second side.

Their internal scorecards should change too. "We delivered 27 requests" is not enough. How much work entered the common platform? Did onboarding time fall? Can earlier customers handle routine failures without the original team? Did field evidence change the product roadmap?

China's extra gate

Enterprise AI in China also has a gate that cannot be bolted on at the end: data and regulatory design.

Linjiang initially wanted to send customer emails, equipment logs, and maintenance records through one model. Zheng asked the team to draw four lists first. Which features were internal and which faced the public? Who owned each category of data? Which models and environments could receive it? Which actions always required human approval?

This was not paperwork for the final week. Once the architecture hardens, separating data becomes expensive.

Article 2 of China's Interim Measures for Generative AI Services says the measures do not apply to internal enterprise research and use that is not offered to the domestic public. Internal systems still have to respect personal information, trade secrets, data controls, access permissions, and applicable sector rules. Public-facing services may also need to address lawful data sources, service agreements, security assessment, or algorithm filing, depending on the service.

Systems that generate public text, images, audio, or video must also consider the AI-generated content labeling measures, effective September 1, 2025. Visible labels, file metadata, and distribution responsibilities can all shape the product interface and content pipeline.

Chinese FDE teams should build model routing early. Routine knowledge work can use one path. Sensitive data can remain in a controlled environment. High-risk operations can produce recommendations without receiving execution rights. Models may change, but evaluation sets stay. Vendors may change, but business rules remain with the enterprise.

That is more work than a polished demo. Production systems are supposed to be more work than demos.

Exit day

In week ten, Linjiang changed a quality-release rule. Chen and the planners added cases, updated the evaluation set, ran shadow traffic, and released the new version. Zheng sat at the back and did not touch the keyboard.

"Will you stay for phase two?" Tang asked.

"First tell me which capability is missing," Zheng said. "Do not start with how many people are missing."

An engagement is ready to end when the internal team can deploy and roll back; domain experts and engineers can add new rules to the evaluation suite; routine failures no longer depend on the original FDE; model replacement does not erase the business layer; and the core metric holds for several weeks.

If those conditions are not met, keep building together and name the gap. If they can never be met, the engagement has become outsourcing.

Zheng returned the Project Consultant badge at the gate. Wu's grid notebook was still on her desk, but a screen now sat beside it. The screen had not replaced her. It had gathered the evidence that once required six phone calls.

Companies should buy a defined period of joint work, designed to end. When writing an AI FDE contract, put in the exit date before the start date.