Five Counter-Intuitive Steps That Separate FDEs from Traditional Outsourcing

Step 1: Learn to Say "No"

An MIT report states that 95% of global enterprise generative AI investments in 2024-2025 yielded no measurable return. Most projects die quietly in a "proof-of-concept graveyard" — demos impress, contracts sign, systems launch, but nine months later no department actually uses them. Therefore, an FDE's first job is not coding but screening projects through three gates: (1) Is it a real pain point? "Improve efficiency" is a direction; "a manager spends three hours every Monday manually compiling work orders" is a pain point. (2) What is it worth? The solution must quantify saved headcount, time, or errors. (3) How far can we go? Given the client's current data and constraints, what is achievable? The counter-intuitive insight: rejecting the wrong projects is itself the profit logic, because FDE costs are front-loaded; sinking months into a dead-end project loses far more than a single deal.

Step 2: Pick the Right Client, Not the Biggest

A Chinese AI startup had three leads: a top brokerage with budget but no clear business owner; a top-tier hospital with real pain but data cannot leave the premises, making delivery timelines unpredictable; a regional retail chain where the CFO personally drove a concrete pain point — 3,000+ stores drowning in operational reports that regional managers couldn't process. Sales instinct would chase the brokerage, but the team chose the retail chain because the pain was clear, a decision-maker existed, and data was usable. For large clients with vague needs, the article suggests selling a paid diagnostic (2-3 weeks, fixed fee, delivering a scenario-priority report). If the client has genuine transformation intent, the diagnostic converges on a first landing scenario; if they only want AI window-dressing for an annual report, they won't even pay for the diagnostic — acting as a filter.

Step 3: Shadow First, Code Later

For the first two weeks on site, the team wrote zero code. They followed regional managers on store visits using a "shadow work method" — observing daily work like a shadow. They discovered the data managers actually relied on was not in the official system but in handwritten Excel sheets filled by store managers daily; the official system was slow and error-prone, so the front line had long stopped trusting it. Had they integrated the official system immediately, the resulting tool would have been unused. The key lesson: real workflows often differ from process diagrams; users' stated needs often diverge from what they actually do — FDEs must trust the latter.

Step 4: Launch ≠ Activation

System launch only completes procurement and acceptance. The real test is whether anyone still uses it three months later. Deployment prioritization therefore flips normal product logic: instead of ranking by feature importance, rank by "what blocks users from normal use tomorrow." Example: a color detail that makes factory workers feel the system is designed for office workers, causing refusal, becomes top priority; a powerful predictive feature that isn't needed today gets deferred. A useful daily habit: before leaving, ask "Which user step did today's change make smoother for tomorrow?"

Step 5: Success Means the Client No Longer Needs You

Traditional on-site outsourcing thrives on dependency — the longer the stay, the more indispensable, the steadier the revenue. FDEs aim for the opposite: knowledge transfer enabling the client's team to iterate independently is the true finish line. The professional ethic includes "create no dependency, sell no anxiety." Two common pathologies: deliberately building black-box systems that lock clients in, and hyping "adopt AI or perish" fear to close deals. A telling detail: one engineering team found that if every action triggers an approval popup, ~93% of requests are blindly clicked "Agree." More popups breed numbness; "human-in-the-loop" becomes theater. Guardrail designers cannot assume users stay perpetually alert.

Three Barriers to FDE Adoption in China

How to charge. Chinese clients are used to "buyout plus acceptance" models; outcome-based pricing doesn't yet connect. A pragmatic hybrid: phased acceptance payments, gradually writing quantified value metrics into acceptance criteria.

Is there a platform foundation. FDE presupposes "platform base plus on-site customization." If every project starts from zero code, it's just outsourcing renamed.

What is a person worth. China has no shortage of hardworking engineers willing to embed on-site, but they are long priced by person-days. A 2025 comparison: Yonyou revenue per employee ~¥480k, Kingdee ~¥620k, Palantir ~$1.1M. The gap reflects two logics: selling time by headcount vs. selling value by outcome.

Closing Thought

The previous article likened FDE work to mathematical modeling. After reading the book, the author adds: the hardest part of modeling is often not the building, but the very first step — judging which problem is worth modeling and which should be refused. Reject a seemingly large problem, pick a small clear one, shadow to understand it, then act. This sequence applies to AI deployment as much as to teaching children problem-solving.