Why FDE — not outsourcing, not consulting
Your AI project never lacked a model — it lacked someone accountable for the outcome. This page covers three things: why enterprise AI keeps failing; how FDE differs from staff augmentation, implementation and consulting; and the work we explicitly decline.
Start with the data: failure is rarely technical
Share of enterprise generative-AI spend with no measurable financial return (MIT NANDA lab, "The GenAI Divide", 2025; 52 organizations interviewed, 300+ public projects)
What kills projects: employees won't use it, quality doubts, poor experience, no executive support, change difficulty — not one of them is "the model isn't smart enough"
In the same report, projects run with an external team succeeded roughly 3x more often than internal builds — teams that cannot afford to lose define the problem more carefully
The five blockers named in the report (by frequency)
- Employees refuse to use new tools — while using AI privately every day
- Worry about output quality
- Poor user experience
- Lack of executive support
- Change-management difficulty
Caveat: the report defines "failure" as no measurable P&L impact within six months, and its methodology is debated (short observation window; value limited to profit, cost, revenue). RAND and S&P Global point the same direction. We take the direction, not the decimal.
Four roles — where the line actually is
| FDE (Forward Deployed Engineer) | Staff augmentation | Implementation | Consulting | |
|---|---|---|---|---|
| How you pay | Staged delivery, acceptance against results — you can stop at the end of any stage | By the hour or person-day; the longer they stay, the more they bill | By project milestones; the sign-off is the finish line | By project; the deliverable is advice and a report |
| What walks in | An engineering foundation, an evaluation method, and delivery discipline | Code written from scratch; method depends on who is sent | Standard product plus configuration; customization extra | A proposal and a deck |
| What stays behind | We leave when it is done — capability stays in the system and your team | The longer the better; when people leave, the system stops | Delivered and gone; whether it gets used is luck | Advice delivered; execution is not theirs |
| Accountable for | Your business outcome: is the system actually used, is the business better | Attendance and hours | Feature checklist boxes | Report quality |
The seed of this table comes from three distinctions published by Qimeng Technology, one of China's first FDE providers (staged delivery vs. hourly billing; an engineering foundation vs. code from scratch; capability that stays vs. a system that stops when people leave), rewritten and extended for this practice. Put more plainly: presales is accountable for the signature, implementation for the sign-off, customer success for renewal — an FDE is accountable for whether your business actually got better.
What I sell is not code — it is the three things AI does not know
Industry consensus is converging: reasoning can be delegated to models. A deployer's real value is earning your trust and bringing these three things back from your factory.
Objective
Why you want this now — surfaced in a conversation with the owner, and turned into the system's direction anchor. Without it, the AI does not know where to push.
Context
How the process really runs, where the data lives, who decides what at which step — collected on the floor. Without scenario knowledge, output looks right and works wrong.
Taste
Which results you will actually accept and use every day — learned through sustained collaboration. It decides whether the deliverable feels like your factory's own tool or a vendor demo.
One common question, answered: why am I more expensive than a regular engineer or an outsourcer?
Within the same working hours I carry three layers of uncertainty: model risk — an upgrade can mean rework; client risk — unclear requirements and changing decision-makers, with the cost of change absorbed by me; product risk — product manager, architect, and engineer in one person, with the solution grown on site. What you are buying is not code hours; it is a single carrier for three layers of risk. A historical mirror: at the 2006 peak of enterprise IT in China, consultants who could run SAP into a state-owned enterprise's business charged RMB 3,500 per day — companies were never buying hours; they were buying the person who could make a new system truly run inside their business.
Sourcing note: the "three things" framework follows a public paraphrase of MingLue Technology CEO Wu Minghui's June 2026 analyst briefing (single media source — framework adopted, no figures quoted); the pricing explanation follows compiled industry salary analysis. Actual prices are on the services page and in quotations.
How we work
Right problem before right solution
Three gates before the first line of code: is this a specific person’s specific pain (not a direction like "efficiency"); how many person-hours a week does it eat and what is that worth; how far can we get given your data reality. Fail a gate, and we say so.
Shadowing: a day beside your employee
Not interviews — sitting beside them: which systems they open, where they copy-paste between spreadsheets, where they frown, which "official processes" they route around. The real pain lives in the workarounds, not in the process document.
A two-week minimal validation: real data, hard stop
One narrow cut (say, returns tickets only) done end-to-end; acceptance metrics fixed before we start; your real data, never demo data. If targets are missed, we part ways cleanly instead of pushing.
Evaluations first: define "good" before chasing it
Together with your domain experts we turn "what a good answer looks like" into a scoreable standard, and keep re-running it after launch. AI often fails without erroring — it just quietly becomes unreliable, and by the time users notice, trust is already spent.
Work we explicitly decline
Declining the wrong project is part of serving you well — FDE costs are front-loaded, and a quagmire ties up the whole delivery cadence. In these four cases we will talk you out of it, or suggest something lighter:
- The scenario is unvalidated and you just want to "explore AI" — start with free tools and public material; no need to pay
- One spreadsheet export would do the job — a part-timer is the cheaper answer
- No named business owner, only the IT department as counterpart — nobody accountable for the outcome usually means a stalled project
- "Cover every scenario across the company" from meeting one — usually means not ready for any single scenario
"We would rather you start later than start at the wrong time." — from Qimeng Technology's decline mechanism. We do the same.
Trust mechanisms: safety first
Validate before scaling
If pilot targets are missed, we do not move to implementation — you keep the results-and-limitations report; discovery fees can be partially credited toward a later contract.
Humans on key actions
Outbound actions — sending emails, quotes, orders — keep human approval. AI is a copilot first; autonomy comes later.
Security written into the contract
Sensitive-field masking, least privilege, local/private deployment options — specific measures go into the SOW, and we verify data licensing before touching anything.
References: MIT NANDA lab, "The GenAI Divide" (2025); Qimeng Technology's FDE service page; Fan Bing, "Forward Deployed Engineers" (fde4.ai, August 2026, open edition). The views and working methods on this page are Cai Yulin's own practice.
A 3-minute self-check: is your scenario ready for AI?
Seven yes/no questions. More ticks, more worth discussing; too few, and I will suggest you keep the budget for now — a project that fails this list wastes your money even if I take it.
Reading the score: ✅ 5+ ticks — book a call and bring "the most recent incident" to the table; 3–4 ticks — worth a free call to work through the unchecked ones first; ≤2 ticks — bookmark the public material on this site and come back when conditions mature; spending now would most likely go to waste.
Sourcing note: adapted from common client-screening criteria in the FDE industry and this practice's "free boundary" discipline; for expectation management only, not a formal assessment.
Not sure your scenario fits?
A free 30-minute call: first a read on whether it is worth doing and where to start. If it is not a fit, I will say so.