# FDE in Four Cities / 中国四城 FDE 行业观察 > Published 2026-07-09 · By lawted (https://x.com/lawted2) · Published on HA7CH (https://ha7ch.com) > Canonical: https://ha7ch.com/writing/four-cities-fde-report ## English From June 6 to July 4, 2026, HA7CH ran four closed-door FDE meetups in Shenzhen, Shanghai, Hangzhou, and Beijing, one Saturday afternoon each, 129 builders in total. This report is compiled from the live discussions at all four sessions. It is among the earliest firsthand testimony we know of on China's FDE ecosystem. Two things up front. First, attendance figures come from our registration system (31 in Shenzhen, 31 in Shanghai, 32 in Hangzhou, 35 in Beijing), and every price, revenue, and timeline figure quoted in this report comes verbatim from what was shared in the room, with no extrapolation. Second, everything has been anonymized: all views are attributed uniformly to ha7ch guild members, with only industry and background descriptions retained. No real names, company names, or school names appear, with the exception of public industry facts about OpenAI and Anthropic. --- 1. Why Now: Real Business, No Industry Yet FDE, Forward Deployed Engineer. A job title that until recently existed only on Silicon Valley careers pages was discussed, debated, and priced, intensively, across four afternoons in four Chinese cities. It still has no settled definition in China: there is no standard rate card, no established path into the work, not even a Chinese name everyone agrees on. But it already has real contracts: a 20,000-yuan all-inclusive textile outsourcing job, a roughly 200-million-yuan domestic-stack IT (xinchuang) deal, an ERP subscription taking a 5 percent cut pegged to headcount reduction, and e-commerce gigs starting at 500 yuan apiece. What this report wants to record is precisely this moment: real business, no industry yet. The backdrop is a set of mismatches. At the Shenzhen session, a guild member with a cloud-vendor background offered his read: since the new generation of models shipped late last year, coding agents' capabilities have exploded, but user-facing applications have not followed. Compute consumption is concentrated in a small number of users, product and engineering teams face layoffs precisely because of AI, and nearly every startup in Silicon Valley is crowding toward FDE. Capability in surplus, deployment in deficit, investment and returns badly mismatched. Standing in the gap of that mismatch is the FDE. A member at the Hangzhou session who watches the industry from a big-tech vantage point pushed the judgment further: the last generation of software was a tools business, while this generation of AI business delivers "revenue growth" or "cost cuts and layoffs" as the product itself. Big tech has already started cutting staff as coding gets penetrated, and more than 90 percent of industries have not been penetrated at all. His exact words were logged on the spot as a quotable line: "These two years are the biggest blue ocean there is. You have to grab it." What the four sessions produced is not secondhand trend commentary but firsthand testimony from the people doing the delivery, carrying payment receipts, failure postmortems, and layoff guilt. What follows is a city-by-city walk-through. --- 2. Shenzhen #001: On the Industrial Belt, Build First, Talk Money Later June 6, a Saturday afternoon, 31 people. Shenzhen was a classic industrial-belt scene, with the highest case density of the four cities, and every case came wrapped in words with body heat: on-site postings, commissions, headcount efficiency, layoffs. The opening introductions alone showed the spectrum: someone using large models to crack RGB-to-CMYK color conversion for the textile industry, someone doing marketing for a companion-robot company, someone doing RAG research, someone in cloud pre-sales, someone who had moved from sales into FDE. The ratio of traditional-industry people to AI practitioners was far more balanced than you would expect. Start with the small cases. A member with an indie-developer streak hand-built a photo auto-classification tool in a few days, solving a friend's real pain point: a traditional retail sales rep who had to manually sort two to three thousand photos a month. The tool pairs a vision model with local algorithms, and the friend just plugs in his own API key. A pricing debate broke out on the spot: what should this thing cost, priced by the sales rep's hourly wage, or by some other anchor entirely? No conclusion, but everyone in the room realized that pricing is this industry's biggest blank space right now. Another member had done a wilder job: automating invoice issuance for an accounting firm, using an agent to drive the browser through the entire flow, and the project is live. A third member, formerly in financial analysis, used AI to write code that made his own job so efficient he simply quit, then built a website and food-delivery system for relatives running restaurants overseas, solving unstable delivery fees with fixed delivery zones and prepaid pricing, and built a Telegram bot for staff to self-select shifts with automatic payroll settlement. Not one of these jobs appears on any job board, but they are real FDE business. Now the big cases. A member with deep roots in petrochemical IT shared three projects: using RAG to relieve in-house technical experts drowning in repeated Q&A, later extended to external pre-sales; a foreign-trade order-tracking system stalled at half-finished because email formats were too unstable; and AI-assisted decision-making layered onto an existing equipment management system, which landed smoothly. A member who spent years on core trading systems at a foreign bank drove the AI transformation of the trading stack, building a prediction system and a trade-flow agent. After launch it processed a heavy daily volume of trades, the bank subsequently cut staff, and he called himself "the guilty one," half in jest. A member from cross-border e-commerce turned product-selection logic and customs paperwork into formulas plugged into an ERP: shipping workflows for over a hundred SKUs a week used to consume enormous labor, automation cut labor needs by about 60 percent, some staff were laid off, some moved to operating the system, and the transition took about two months. A tech-company CTO put it most bluntly: after going AI-native, they can finish in three or four days what used to take more than half a year, their internal workflows are automated, and efficiency gains and layoff anxiety are two sides of the same coin. One unresolved deal shows the boundary conditions of this market: a highly digitized manufacturing plant with in-house systems wants to bring in AI, an edge-deployment project with expected returns of 3 to over 5 million yuan. But the terms are harsh: no data goes to the cloud, deployment must be on-device, and the technology vendor must hand over the intellectual property. The original team passed; another member in the room said on the spot that he wanted it. Same clause, some see risk, others see a ticket in. What made Shenzhen distinctive was industry, academia, and research at the same table, a rare sight. The associate dean of an engineering school at a university in South China launched a project-based FDE training program on the spot: recruit students across majors, bring in FDEs with real delivery experience to teach, form teams to work on real corporate projects, students get jobs, companies get hands, and he solicited instructors then and there. The supply side was just as lively: a fresh master's graduate from a Singapore university, short on internships, built three B2B POCs on his own and brought them to the room job-hunting; big-tech people who had quit cold because of the agent explosion were looking for direction; a member born after 2000 with a game-engine background had already built an AI avatar of his mother for foreign-trade lead generation. A member doing engineering-survey digitization inside his company voiced the common plight of the internal reformer: he has built plenty of tools, but efficiency gains inside the company carry limited economic value and little bargaining power, and he is wavering on going independent. The room's advice was one line: get out and make money. Business models got their most direct and most unruly airing here: revenue shares pegged to layoffs or growth, technology for equity, AI roll-ups (raise capital to acquire traditional businesses, retrofit them with AI to lift margins, then keep rolling up more; startups in the US are already running this play), and splitting a project into training, coaching, and vision-selling as three separately billed phases. Pricing anxiety ran under the whole session, from "too embarrassed to charge" to "price by the client's anxiety level." The room also logged its expectations for the competition ahead: a price war may come, and the industry will move toward consolidation and standardization. --- 3. Shanghai #002: The Craft of Getting Paid June 13, a Saturday afternoon, 31 people. Shanghai was an intensely pragmatic business session, centered not on technology but on collections. The finance concentration was markedly higher than elsewhere: an asset-allocation product manager from a foreign bank, an undergraduate with nine finance internships, and students with brokerage backgrounds shared the room, putting constraints on the table you rarely hear elsewhere: heavy regulation, traceability and auditability, on-premise deployment, policy review for robo-advisory. Start with the definition. A member with an organizer's background drew a line for FDE: an FDE is not an outsourcer who standardizes data, builds databases, or deploys chatbots, but a role that shortens a company's decision loop. His example was an airline: can we fly in this rain, how do we handle an emergency. Once AI is deployed, what shrinks is the decision chain, not document-processing time. Another member's project made the definition concrete: he is building an agent for a professional racing team, handling PR and logistics, making advance decisions on crew lodging and cargo transport. His summary stuck with the room: AI efficiency does not necessarily mean making the car faster; it means faster decisions and better-rested drivers. The hardest-won material came from traditional-industry veterans. A member with over a decade in intelligent retrofits did a postmortem on smart-mine projects: hardware-inclusive projects carry heavy upfront work, errors and costs slip out of control easily, acceptance criteria are vague, and the client's leadership can seize on any problem as a reason to delay payment. His most extreme sample: a mining project in western China dragged on for years without completion, over client-side execution and budget problems, and the hardware warranty expired in the meantime. His countermeasure is not technical but relational: build communication with every level of the client organization, and screen out unsuitable industries and business types from the start. An operator at a state-owned IT company used a domestic-stack IT (xinchuang) deal worth hundreds of millions of yuan as his example and delivered a more structural judgment: FDE does not fit state-owned institutions. Pitching cost savings and efficiency to them goes nowhere; settlement requires layer upon layer of approval, and the logic never closes. State-owned enterprises (SOEs) answer upward, not downward, so the acceptance criterion is whether the people above are satisfied. If you must do it, get alignment with the very top first and push top-down. A member selling integrated solutions shared two ledgers, one about profit, one about cost. The profit ledger: an integrated solution bundling high-end life-support equipment with a prediction system, negotiated at 1 to 1.5 million yuan per case, with clear delivery boundaries and fat margins. But at 60 to 70 percent completion, the project died quietly, tangled in the cadence of official red-header documents, school resources, and annual budgets. The cost ledger: he converted a data-annotation line to agents, and error handling genuinely got faster, but model consumption ran 30,000 dollars that month, more than the original small-scale cloud bill, and every workflow still needed human review at the end. Uncontrollable cost directly killed his ability to quote future deployments. Organizational problems were put squarely on the table here. A member automating the full sales pipeline said his system targets industrial firms' bidding and tender workflows and achieves essentially zero employee input, but the moment employees learned the agent would replace their work, friction appeared instantly. The room's solution was surprisingly mature: partner with firms specializing in labor arbitration and workforce transition, and settle compensation, redeployment, and non-compete questions with the CEO and HRBP during the organizational diagnosis phase; transfer whoever can move to other business units, retrain whoever can be retrained. Another member shared a lesson from logistics: from efficiency pitch to delivery, blocked at every turn, and the final landing was to extract the features as plugins grafted onto the client's existing ERP. Anything that touches core systems, like finance-business integration, do not touch if you can avoid it. On playbooks, an AI tooling company that got into FDE work early laid out its path: enter the enterprise with generic needs first, help the client bank quick wins and build trust, then capture high-margin work through deep needs. A member with a consulting background doing enterprise diagnosis added a feature of China's toB ecosystem: the product is sold to the decision-maker, not the user. Data security came up again and again: one team doing AI work for a leading consumer brand runs all client-document redaction on local models; another member mentioned a regional SOE's AI transformation project with budget in hand but little grasp of agent technology, so market education happens one training session at a time. Two details explain why non-technical factors matter. One member dissected why Microsoft Copilot sells well despite not being great: SLA breaches come with compensation, and the brand itself is a procurement rationale. You cannot build looking only at the technology; companies of different sizes are buying service and certainty. Another member cited a choice from the aerospace world: for the reliability and stability of long-cycle projects, they would rather stick with old software versions and forgo the latest technology. Put the two together and you have the real procurement logic of toB. Supply and demand closed the loop inside the room itself: an energy-storage materials company came looking for an algorithm partner, wanting to start at the mine with ore-sorting and process optimization; a doctoral student from a traditional engineering background, who had only touched vibe coding in April or May, was picked up on the spot by a team hiring interns. What lingered after the session was a genuine appetite for organization: one member proposed founding an FDE industry association to unify understanding and set standards. --- 4. Hangzhou #003: A Grassroots Blue Ocean, AI Landing in the Capillaries June 27, a Saturday afternoon, 32 people. Hangzhou was the most grassroots, most e-commerce-flavored session of the four, with the widest price spectrum. At one end, fast small jobs at 500 to 5,000 yuan apiece: a member who taught himself coding out of cross-border e-commerce operations now works as an FDE at an e-commerce company. His business lines are content replication (cloning viral product videos), RPA automation (simulating manual web operations), and data dashboards (as he put it, the thing scrappy small-town bosses welcome most). He prices by the client's budget, installation and training included, gets clients through short-video platforms and job boards, and screens them with almost disarming bluntness: use dressed-up open-source projects and demos to filter out the non-payers first. At the other end, a member with real nerve landed a CRM-plus-B2B-inquiry customer-profiling project through connections just two weeks after arriving in Hangzhou: 50,000 yuan for phase one, roughly 100,000 in total, a team of two or three. The cases in between were all over the map. A fresh graduate joined an enterprise-services firm as an FDE and was dispatched shortly after joining to a temple posting, teaching the monks to use AI and configuring systems. His field report: the masters are easygoing and highly receptive, but everything is hand-holding. A solutions engineer covering general industry did a postmortem on why a steel-mill project would not move: not because the technology fell short, but because the client had no confidence in its own data quality and the veteran operators feared being replaced by AI. The project died before the technology got a chance. An engineer doing high-performance computing at a major tech company had used agents to automate the long pipeline of algorithm deployment, but inside the big company he cannot use the latest models, budgets are tight, and there is little room to adjust internal processes. His self-assessment: close to the technology, far from the market, here to hear stories of FDEs dealing with clients on the front line. Career changers set the tone of this session. A financial accountant at a chain business, keeping books for hundreds of stores, first squeezed efficiency out of software, then used an AI coding assistant to automate 60 percent of his work, then built small websites and browser extensions. His colleagues' fear of AI is exactly what showed him the FDE direction; his worry is his non-technical background. A Java developer with five years' experience wants to move into FDE, gets leads by posting in content communities, and has already built a dashboard for a friend's factory; a retail stock trader also came to him asking for quantitative trading code. A member with over five years of AI algorithm deployment in chip manufacturing wants to run a one-person company and admitted he understands neither client acquisition nor business. A member on a large-model team at an SOE software unit, working on efficiency for new-energy-vehicle software development, is stuck on requirements comprehension and test quality evaluation, and walked away with a usable testing idea from the room: traffic replay plus verification. Everyone was asking the same question: I have this half of the skill set, how do I get the other half? The supply-demand mismatch was most visible in this session. A founder whose main business is cloud computing said his channels have accumulated a large client base, the AI deployment demand is real and well-budgeted, but his team is strong on business development and short on delivery confidence. A member selling inside the ecosystem of a leading foreign-trade B2B platform was the exact inverse: the platform is pushing AI products hard this year, he has started selling, his clients have a rough grasp of AI but no deployment guidance, and he has client resources but no technology, weighing whether to close the technical gap himself or find a technical co-founder. One side has work and no hands; the other has hands and no work. What separates them is the trust and pricing machinery this industry has not yet grown. Hangzhou contributed several judgments that reached room-wide consensus. First, engineering is no longer the moat: agents will readily help you, and engineers on generic stacks are easy to find, but people who understand a specific business domain are hard to find. The FDE's core moat is domain business knowledge. Second, business data matters more than business logic: most corporate data cannot be fed to agents as-is, and data infrastructure is itself the opportunity. Third, clients pay for exactly two things: ROI they can compute, or making the boss happy. SOE deals that come with detailed PRDs are easy to execute and easy money, but they barely differ from traditional outsourcing and command no premium. Fourth, AI transformation must be driven by the boss in the top seat; bottom-up pushes rarely move, and before signing, check your counterpart's authority. The acquisition and screening playbook was equally blunt: for clients with no budget and no understanding, charging a consulting fee first is the best filter. One overseas reference kept being cited: OpenAI and Anthropic have each formed joint ventures with PE firms to lift portfolio-company margins, in essence using capital relationships to remove transaction friction and accelerate AI penetration. Silicon Valley startups have the last generation's toB muscle and VC backing, which makes FDE comparatively easy; Chinese VCs care more about grand trends, the FDE ramp is long, and the toB narrative is systematically undervalued in China. On billing, a member who serves as COO of a Shenzhen hardware company while running an FDE practice offered his three-part kit: train management and staff separately, start billing the moment the diagnostic phase begins on-site, then switch to project-based fees. The team formula also converged in this session: a domain expert plus an engineer with project experience. --- 5. Beijing #004: Systems, Organizations, and Client Psychology July 4, a Saturday afternoon, roughly 30 people. Beijing had the highest resource density and the widest spectrum. The first speaker of the day was a data annotator: he automated his job of copy-pasting evaluation sets from spreadsheets into chat windows with a Python script, packaged it as an EXE, spotted the FDE direction and decided to pivot. He dressed up his resume, sent it out, and found real market demand, but every time someone asks "how do you charge," he has no answer. At the other end of the same afternoon: the founder of an AI brand-marketing company with tens of millions of yuan in annual revenue, there to recruit a technical co-founder; a medical information-services company with deep roots in hospital invoicing, looking to go deeper into hospital economic-operations management platforms; and a veteran with nearly 20 years in hotel franchising, carrying a product concept billed annually at 10,000 to 40,000 yuan per property, looking for a development team, with a goal stated at full volume: ride this to an IPO together. Two young people's stories deserve their own note. One fresh graduate rebuilt a content-platform business with AI from scratch at an agency-side marketing firm, one person going head-to-head against what used to be a 30-person team. Batch generation of posts and data dashboards is already iterating, some of the numbers hold up, and the next step is taking on FDE contracts. A graduate student built an AI agent that helps foreign tourists hail rides, find restaurants, and plan itineraries, and is exploring overseas client acquisition. What they share is that neither waited for anyone's permission: build it first, then come to the room looking for an amplifier. The discussion in this session leaned distinctly toward systems and organizations. A government-and-enterprise digitization vendor is building an FDE curriculum and co-developing standards and certification; they use telecom-carrier client relationships as a wedge to serve heavy-industry and energy companies in the north, and their stated ask was matching strong FDEs to large B-side orders. One member wants to build an FDE talent platform: sourcing, training, matching to enterprises, with personal-brand incubation on the side. A carmaker is pushing AI-native reform through org-structure changes, building a structured AI coding knowledge base. A large consumer-goods company is rolling out AI across every department, pulling staff from business lines while hiring outside. A business-analysis role at a big tech firm offered a quantified internal sample: over 50 percent efficiency gains across business-analysis workflows, though external tools were ultimately dropped over problems with the tools themselves. The shift in hiring criteria was named outright by a partner at a Series A AI hardware company: from engineer-leaning to product-manager-leaning, favoring young people with a hybrid background of big tech plus startup plus SE service experience. Another member in the room added a field observation: students from top universities really do learn faster and adapt across more surfaces. His scaling sample was an FDE company in a vertical hardware niche with roughly 20 million yuan in annual revenue, growing its client base through investor referrals; another acquisition reference was a content creator serving private equity and investment banking, where same-industry case studies carry far more trust than cross-industry ones. The dissection of client psychology was the sharpest of the four cities. A member running multiple AI products (including an AI academic-writing product with steady revenue in the 100,000-yuan range) called it: some companies hire an FDE not for ROI but because the boss wants a person sitting there for peace of mind. So step one is figuring out whether the boss wants ROI or reassurance. A member who moved from vertical agents to B-side work testified with his own detour: a contract-review project stalled, and only when he pivoted to financial and business data analysis did things click. His summary: the boss cares about exactly three things, results, cost, and whether it can be replicated across other lines of business, and cares not at all how you built it. A member with a legal background added the other side: clients' expectations about where AI's capabilities end are generally fuzzy, so boundary management is itself part of the delivery. Methodological convergence was clearest in this session. Step one of enterprise AI adoption is datafication: small companies start with a database or knowledge base; only at large companies do business-process and org redesign come into play. The knowledge base is the main event: heterogeneous data and expert knowledge must be consolidated into a form agents can call, and what bosses care about most is whether the knowledge base can evolve with the business. That currently has no general solution, and traditional RAG accuracy is degrading. One reusable sample was a knowledge base built for a professional racing team: tires, cars, and logistics all structured, wired into a collaboration platform for agentic search. The room's verdict: a knowledge base plus data skills can solve 80 percent of FDE project problems. Acquisition advice circled back to basics: run your first deal through channels where trust already exists, accumulate data and case studies, and spin up the flywheel. Picking the industry matters more than picking the client: sectors like home services, low-margin and thin on talent, were named as having limited ability to pay. And AI startups should head for incremental markets like overseas expansion and foreign trade, steering clear of the tangled interests that cost-cutting plays stir up in saturated markets. --- 6. Cross-City Comparison: Pricing, Acquisition, Pain Points, Talent Start by laying the four cities' price points side by side. This may be the densest public sample of Chinese FDE pricing to date: Shenzhen. Manufacturing plant edge-AI project: expected returns of 3 to over 5 million yuan, with harsh terms including no cloud for data and surrender of IP. Shenzhen. Cross-border e-commerce supply-chain automation: roughly 60 percent labor reduction, a transition of about two months, over a hundred SKUs a week. Shenzhen. AI content for lead generation: 50,000 views in 4 days, converting directly into business inquiries. Shanghai. Integrated high-end life-support equipment plus prediction system: 1 to 1.5 million yuan per case. Shanghai. Domestic-stack IT (xinchuang) deal: contract around 200 million yuan. Shanghai. Data annotation converted to agents: 30,000 dollars of model consumption in a single month, above the original cloud cost. Shanghai. Long-term advisory retainers: on the order of a million yuan per year depending on company size, contracts typically signed for three years. Hangzhou. Small e-commerce FDE jobs: 500 to 5,000 yuan apiece, installation and training included. Hangzhou. CRM plus customer-profiling project: 50,000 yuan for phase one, roughly 100,000 in total, a team of two to three. Beijing. Textile inkjet-positioning outsourcing: 20,000 yuan all-in, replicable and resellable to other textile mills. Beijing. Yiwu manufacturing AI plus ERP: a subscription taking a 5 percent cut pegged to headcount reduction, 50,000 yuan a year. Beijing. Hotel AI product concept: annual billing, 10,000 to 40,000 yuan per property. Beijing. FDE company in a vertical hardware niche: roughly 20 million yuan in annual revenue. Beijing. AI brand-marketing company: tens of millions of yuan in annual revenue. Beijing. Consumer AI academic-writing product: steady revenue in the 100,000-yuan range. On models, each city leaned differently, but together they form a complete map: project-based (person-days, feature points, all-inclusive), subscription (cost-savings cuts, annual billing), consulting-first (charge a consulting fee to screen clients, bill from the moment on-site diagnosis begins), lock-in (technology for equity, the overseas reference of PE joint ventures), and reuse (templatize solutions for resale, then charge for tokens once the work is banked). Two pricing anchors won recognition across cities: the digital employee's annual salary (proposed in Shanghai: price AI services at the salary of the role they replace, which bosses accept more readily), and back-calculating from cost savings (Shenzhen's layoff-ratio revenue share, Beijing's 5 percent subscription cut). The problems Shenzhen exposed (too embarrassed to charge, no standard for quotes) had structured answers by the time of Beijing, such as three-phase billing: corporate training, paid on-site diagnosis, then development, delivery, and maintenance. The first two phases can be bought separately, and for long-term maintenance the advice is to teach the client's internal IT to take over. But the question of how to convert workflow improvement into money went unanswered in all four cities. On acquisition, the four cities' playbooks overlap heavily: content-based acquisition (short video, content communities; Shenzhen had the sample of 50,000 views in 4 days converting straight into inquiries), first deals through trusted personal networks (Beijing framed it as the starting point of the flywheel), channel reuse by industry veterans (20 years of hotel contacts, industrial-belt resources in Foshan and Yiwu), on-site entry followed by lateral expansion, and referrals backed by same-industry case studies. The difference is in client segments: Shanghai stressed SOE relationship management and the top-down play, Beijing added the route of entering government-and-enterprise accounts through carrier relationships, and Hangzhou was bluntest: charge a consulting fee first and keep the budgetless out. The pain points sort into four layers, and every city named all four, just in different words. The demand layer: clients cannot articulate their pain, offering only broad notions like AI marketing or AI lead generation, with no acceptance criteria, so vendors dare not promise outcomes. The data layer: corporate data cannot be handed to agents as-is, sensitive data is hard to connect, and clients do not even trust their own data quality. The organizational layer: employee resistance, veteran obstruction, department walls; efficiency gains directly trigger layoff anxiety, and management and staff need two separate narratives. The technical-commercial layer: however high single-task success rates get, chaining tasks collapses the success rate of the full business scenario; model costs are uncontrollable; payment cycles slip; internal reformers have no bargaining power; and AI consulting resists scale. The talent and transition signals escalated city by city. Shenzhen produced university project-based FDE training and fresh graduates job-hunting with self-built POCs. Shanghai is testing an "expert plus FDE plus intern" placement model across industry and academia, and a doctoral student who patched up his algorithm skills with vibe coding got snapped up in the room. Hangzhou proved the non-technical route works: a financial accountant who automated 60 percent of his job, e-commerce operators, and Java developers each found their own conversion. Beijing supplied the demand-side shift in standards: hiring is moving from engineer-leaning to product-manager-leaning, senior FDEs can go straight to co-founder seats, and curriculum and certification co-building is underway. The consensus running through all four cities: initiative first, domain knowledge is the scarce input, and pure engineering skill is being flattened by agents. --- 7. Trend Calls: The Next 6 to 12 Months Call one: pricing will converge on anchors and then head into the eve of a price war. Shenzhen named the pricing problems born of being too embarrassed to charge, and the notes logged expectations of a price war and consolidation; Shanghai contributed the digital-employee-salary anchor and million-yuan-a-year retainers; Beijing produced the replicable 5-percent-of-layoffs subscription formula. Anchors will spread faster than practitioners expect, and the low end (the 500-to-5,000-yuan jobs) will be the first to turn cutthroat. Call two: data infrastructure will displace agent development as the FDE's main battlefield. Hangzhou judged business data more important than business logic; Beijing concluded that a knowledge base plus data skills can solve 80 percent of project problems and that datafication is step one of enterprise AI adoption; Shenzhen argued that proprietary data is the entry point. Three cities converging independently on the same conclusion means the first line item on an FDE quote over the next six months will most likely be data governance, not agents. Call three: client segmentation will become the line between life and death. Shanghai delivered the structural verdict that the cost-savings pitch goes nowhere in state-owned institutions; Hangzhou noted that SOE deals with detailed PRDs are easy money but barely differ from traditional outsourcing and command no premium; Beijing named low-margin sectors like home services as unable to afford FDEs and argued for incremental overseas markets. Teams that pick the wrong segment will be bled dry at collections (the mining project that dragged on for years without completion is the cautionary tale), and industry-screening ability will separate winners from losers before delivery ability does. Call four: the efficiency-layoff paradox will move from a moral topic to a step in the delivery process. Shenzhen's bank engineer jokingly called himself the guilty one; Hangzhou's steel-mill project died on veteran resistance; Shanghai has already produced the mature practice of partnering with labor-arbitration and workforce-transition firms and settling compensation and redeployment with the CEO and HRBP during organizational diagnosis. Workforce-transition plans will enter the FDE's standard list of deliverables. Call five: FDE talent supply will systematize within 6 to 12 months. Shenzhen's university project-based training, Shanghai's industry-academia placements, and Beijing's curriculum-and-certification co-building and talent-platform ventures are three stages of the same thing. Add the talent spillover from big-tech layoffs in coding roles and the demand-side shift from engineers to product managers, and the first cohort of formally trained FDEs will hit the market in the first half of next year. Call six: the fight over paths to scale will decide whether capital shows up. Beijing's VCs admitted most are still in the getting-acquainted phase with no clear process; the worry that FDE looks like outsourcing and resists scale was relayed by founders in the room as the industry's standing question about FDE. Three answers have already surfaced: bank solutions for reuse and then charge for tokens, technology for equity, and the reference model of overseas model companies forming joint ventures with PE. Whichever produces a replicable sample first will decide whether China's undervalued toB narrative can flip within a year. --- 8. Closing Four cities, 129 builders, four afternoons, assembling the earliest terrain map of China's FDE ecosystem: Shenzhen's industrial-belt street fight, Shanghai's collections sobriety, Hangzhou's grassroots blue ocean, Beijing's systemic ambition. The industry has no consensus even on its name, and meeting transcription tools cannot spell out the acronym, but it already has a real contract spectrum running from 500 yuan to 200 million. ha7ch guild's next move is to turn this terrain map into infrastructure: keep running periodic closed-door sessions city by city, so supply and demand keep closing the loop in the same room; make hackathons and sprints the channel through which young people enter real enterprise sites; and bank the pricing anchors, client-screening methods, and delivery lessons that recurred across the four cities, so the next person converting to FDE does not have to start over from "too embarrassed to charge." "These two years are the biggest blue ocean there is. You have to grab it." That line came from the Hangzhou session. We leave it here exactly as spoken, and we will come back in a year to check the answer. ## 中文 2026 年 6 月 6 日到 7 月 4 日，HA7CH 在深圳、上海、杭州、北京连办了四场闭门 FDE Meetup，每场一个周六下午，四场合计129 位 builder。这份报告基于四场的现场讨论整理而成，是我们能找到的、关于中国 FDE 生态最早的一批一手证词。 先说清楚两件事。第一，到场人数来自报名系统统计（深圳 31、上海 31、杭州 32、北京 35），报告里引用的价格、流水、周期等数字全部来自现场分享原话，未做外推。第二，全部内容已匿名化：观点归属统一表述为 ha7ch guild 成员，只保留行业与背景描述；OpenAI 与 Anthropic 这类公开行业事实除外。 --- 一、为什么是现在：有生意，没行业 FDE，Forward Deployed Engineer，前向部署工程师。这个此前只存在于硅谷招聘页上的岗位名称，在中国四座城市的四个下午被密集地讨论、争辩、定价。它在国内还没有统一定义，没有标准报价，没有成熟的入行通道，甚至没有一个所有人都认的中文名字。 但它已经有了真实的合同：2 万包圆的纺织外包单，约 2 亿的信创大单，按裁员比例抽 5% 的 ERP 订阅，单笔 500 元起步的电商小单。这份报告想记录的，正是这个「有生意、没行业」的时刻。 背景是一组错配。深圳场一位云厂商背景的 guild 成员给出了他的观察：自去年底新一代模型发布后，coding agent 的能力爆发，但用户端应用没有跟着爆发，算力消耗集中在少数用户手里，产研团队反而因为 AI 面临裁员，硅谷的 startup 几乎都在往 FDE 方向挤。能力过剩，落地不足。错配的缝隙里，站着的就是 FDE。 杭州场一位从大厂视角观察行业的成员把这个判断推得更远：上一代软件生意是卖工具，这一代 AI 生意的交付物是「增收入」或者「降本裁员」本身；大厂因为 coding 场景渗透已经开始裁员，而 90% 以上的行业还没有被渗透。他的原话被当场记进了金句：「这两年就是最大的蓝海市场，一定要抢。」 四场会呈现的不是二手转述，而是一线交付者带着回款单、失败复盘和裁员愧疚感的一手证词。下面按城市走一遍。 --- 二、深圳 #001：产业带现场，先干出来再谈收钱 6 月 6 日，周六下午，31 位。 深圳场是典型的产业带现场，案例密度四城最高，而且全是带体温的词：驻场、提成、人效、裁员。开场自我介绍就能看出光谱：有人用大模型解决纺织行业 RGB 转 CMYK 的难题，有人在陪伴机器人公司做营销，有人在云厂商做售前，有人从销售转做 FDE。 案例先说小的。一位独立开发者气质的成员，几天手搓出一个照片自动分类工具，解决的是朋友的真实痛点：一位传统零售业务员，每月要手动整理两三千张照片。工具用视觉模型加本地算法。现场随即展开定价讨论：这个东西该收多少钱，按业务员的时薪算，还是别的锚点。没有结论，但所有人都意识到，定价是这个行业眼下最大的空白。另一位成员的单子更「野」：帮财务公司做自动开发票，用 agent 操作浏览器跑通整个流程，已经上线。还有一位原本做财务分析的成员，用 AI 提效自己的工作之后干脆辞了职，帮在海外开餐馆的亲戚做了网站和外卖系统，用划定配送区域、预先收费解决外卖费用不稳定，又用 Telegram bot 做了员工自助选班和自动算工资。这类单子不会出现在任何招聘网站上，但它们是真实发生的 FDE 业务。 案例再说大的。一位深耕石油化工信息化的成员分享了三个项目：用 RAG 解决企业内部技术专家被频繁答疑的困扰，后来拓展到外部售前；外贸订单追踪系统因为邮件格式不稳定，停在半成品；在既有设备管理系统上加 AI 辅助决策，落地顺利。一位在某外资银行做核心交易系统多年的成员，推动交易系统 AI 化，上线后日处理大量交易，行内随后缩减人员，他自嘲是「罪人」。一位跨境电商出身的成员把选货单逻辑和出关单据整理成算式套进 ERP，每周上百个 SKU 的运输流程自动化后降低约 60% 人效，一部分员工被裁，一部分转去操控系统，转型周期约两个月。一位科技公司 CTO 说得更直接：AI 化之后能用三四天完成以前半年以上的开发任务，效率提升和裁员焦虑是同一枚硬币的两面。 一个悬而未决的单子能看出市场的边界条件：某数字化程度很高的制造工厂，自研系统，想接入 AI，端侧项目预期收益 300 至 500 多万。但条件苛刻：数据不上云、端侧部署、供应商交出知识产权。原团队没接，现场另有成员当场表示想接。同一个条款，有人看到的是风险，有人看到的是门票。 深圳场的独特性在于产学研罕见地同桌。华南某高校一位工科学院副院长现场发起 FDE 项目制培训：跨专业招学生，请实战 FDE 授课，组队做真实企业项目，当场征集授课者。供给侧同样活跃：从新加坡高校毕业的应届硕士自己搭了三个 B2B 的 POC 带着来求职，裸辞的大厂人在找方向，一位 00 后游戏引擎技术出身的成员，已经给母亲做了一个 AI 分身用于外贸获客。一位做工程勘察数字化的成员说出内部改造者的普遍处境：工具做了不少，但在公司内部提效议价权低，正在犹豫要不要出来单干。现场给他的建议只有一句：出来赚钱。 商业模式在这一场聊得最直接也最野：按裁员或增长比例分成，技术换股权，AI roll-up（融资收购传统企业、用 AI 改造提升利润率再滚动收购），以及把项目拆成培训、陪跑、画饼三段分开收费。定价焦虑是全场暗线，从「抹不开脸收钱」到「按客户焦虑程度报价」。现场也留下了对市场竞争的预期：价格战可能出现，行业会走向整合与规范化。 --- 三、上海 #002：把钱收回来的学问 6 月 13 日，周六下午，31 位。 上海场是一场极其务实的商务局，讨论重心不在技术，而在回款。金融浓度显著高于其他城市：外资银行的资产配置产品经理、做过 9 段金融实习的本科生、券商背景的学生同场，把强监管、可追溯可审计、私有化部署、智能投顾政策评估这些别处少见的约束条件摆上了桌面。 先从定义讲起。一位组织者背景的成员给 FDE 划了线：不是数据标准化、搭数据库、部署聊天机器人的外包，而是让企业决策流程变短的角色。他举的例子是航空公司：下雨能不能飞、紧急情况怎么处理，AI 缩短的是决策链路，不是文档处理时间。另一位成员的项目把定义具象化了：他为某职业车队做 agent，负责公关、后勤，提前决策人员住宿和货物运输。他的总结被现场记住：AI 提效不一定是让赛车开得更快，而是让决策更快、车手休息更好。 最硬的干货来自传统行业老兵。一位做了十多年智能化改造的成员复盘智慧矿山项目：含硬件的项目前置工作多，误差和成本容易失控，验收标准不明确，甲方领导随时可以拿问题当理由推迟付款。最极端的样本是西部某矿业项目，因甲方现场落实和预算问题拖了数年没有完成，硬件质保期都过了。他的对策不是技术，是客情：跟企业每个层级建立沟通，从一开始就筛掉不适合进的行业。一位国有信息化企业的操盘者用一个亿元级的信创大单做例子，给出更结构性的判断：FDE 在国有单位不适合落地，谈省钱提效结算要层层审批，逻辑走不通；国央企只对上负责，验收标准就是让上面的人满意，要做就先跟最高层达成共识，自上而下推。 一位做整体方案销售的成员算了两笔账。利润那笔：含高端生命支持设备与预测系统的整体方案，单案例谈到 100 至 150 万，交付边界明确，利润空间大，但项目做到 60% 至 70% 时，因为红头文件有周期、学校资源和年度预算的问题，无疾而终。成本那笔：一条数据标注业务线 agent 化后效率确实提高，但当月模型消耗花掉 3 万美元，比原来的小规模云成本还高，且流程跑完仍需人工审核。成本不可控，直接卡死了他后续报价的能力。 组织问题在这一场被摆到了台面上。一位做销售全流程自动化的成员说，他的系统针对工业企业的标书和招投标流程，基本做到零员工输入，但员工知道智能体要替代自己的工作后，沟通阻力立刻出现。现场给出的解法成熟得让人意外：与专业人员安置公司合作，在组织诊断阶段就跟 CEO 和 HRBP 谈清补偿、转岗和竞业；能转岗的转岗，能培训的培训。物流行业的教训类似：提效方案处处受阻，最后把功能抽成插件嫁接到企业原有 ERP 上；业财一体这种动核心系统的事，能不碰就不碰。 打法层面，一家较早做 FDE 业务的 AI 工具公司给出路径：先用通用需求进入企业、快速建立信任，再用深度需求获取高毛利。一位有咨询背景、在做企业诊断的成员补充了中国 toB 生态的特点：产品是卖给决策者的，不是卖给使用者的。数据安全被反复提及：有团队给头部消费品牌做 AI 项目，客户文档脱敏全部用本地模型处理；某地国企的 AI 转型项目有预算，但对 agent 技术了解不足，市场教育只能靠培训一点点做。 还有两个细节解释了为什么非技术因素重要：微软 Copilot 不好用还卖得好，因为 SLA 出问题有补偿，品牌本身就是采购理由；航天系统为了长周期项目的稳定性，宁可坚持用旧版本软件。把这两条放在一起，就是 toB 的真实采购逻辑。 供需闭环在房间内直接发生：一家储能新材料企业现场找算法合作方，想从矿端开始做选矿和工艺优化；一位传统工科背景的博士生，四五月份才接触 vibe coding，现场就被在招实习生的团队接住了。散场时留下的是对行业组织化的真实诉求：有成员提议成立 FDE 行业协会，统一认知、制定准则。 --- 四、杭州 #003：草根蓝海，毛细血管里的 AI 落地 6 月 27 日，周六下午，32 位。 杭州场是四城里最草根、最电商味的一场，价格光谱极宽。一头是单笔 500 到 5000 元的小单快跑：一位从跨境电商运营自学 coding 转型的成员，在电商公司做 FDE，业务是内容裂变（复刻爆款带货视频）、RPA 自动化（模拟人工操作网页）和数据仪表盘（用他的说法，土老板最欢迎这类东西），收费看客户预算，含软件安装和培训，获客靠短视频平台和招聘平台，筛客户的办法直白：用包装过的开源项目和 demo 先把不付费的滤掉。另一头是一位敢想敢做的成员，到杭州才两周就靠关系签下 CRM 加 B 端询盘客户画像的项目：一期 5 万，整体约 10 万，团队 2 到 3 人。 中间夹着的案例五花八门。刚毕业的成员入职一家企业服务商做 FDE，入职不久就被派到寺庙驻场，教师傅用 AI、配置系统，他的现场报告是：大师们好说话、接受度高，但要手把手教。一位泛工业方向解决方案工程师复盘了钢厂项目为什么推不动：不是技术不行，是客户对自己的数据质量没自信，老师傅担忧被 AI 取代，项目死在了技术之前。有大厂里做高性能计算的工程师，把算法部署的长链路用 agent 自动化了，但在大厂内部用不了最新模型、预算受限，内部流程调整空间也小，他的自我评价是：离技术近，离市场远，来这里想听 FDE 在一线跟客户打交道的故事。 转型者浓度是这一场的底色。一位连锁企业的财务会计，管着上百家门店的账，用 AI 编程助手自动化了 60% 的工作，同事对 AI 的恐惧反而让他看到了 FDE 这个方向，担心的是自己非技术出身。一位有 5 年经验的 Java 开发想转 FDE，靠在内容社区发帖获客，已经帮朋友的工厂搭了大屏，还有炒股散户找上门让他写量化交易代码。一位在芯片制造行业做了 5 年多 AI 算法落地的成员想做一人公司，坦承不懂获客和商业。一位在国企大模型小组为汽车软件开发提效的成员，卡在需求理解和测试质量评估上，现场直接收到可用的测试思路：流量回放加验证。各路人马问的是同一个问题：我这半边能力，怎么补上另外半边。 供给和需求的错位在这一场看得最清楚。一位主业做云计算的创业者说，他靠渠道积累了大量客户，AI 落地需求真实存在而且预算充足，但他的团队商务有余、交付信心不足。一位在头部外贸 B2B 平台生态里做销售的成员正好相反：平台今年力推 AI 产品，他已经开始售卖，客户对 AI 有初步认知但缺落地指导，他有客户资源缺技术，正在考虑补技术短板还是找技术合伙人。一边是有活没人干，一边是有人没活干，中间隔着的就是这个行业还没长出来的信任和定价机制。 杭州场贡献了几条全场共识级的判断。其一，工程不再是门槛：很容易有 agent 帮你，也很容易找到通用技术栈的工程师，但很难找到懂某个业务领域的人，FDE 的核心壁垒是领域业务知识。其二，业务数据比业务逻辑更重要：企业数据大多无法直接供 agent 使用，数据基建本身就是机会。其三，客户付费考量只有两类：按 ROI 算账的，和让老板开心的；国央企出详细 PRD 的单好做、易赚钱，但和传统外包流程差别不大，难收溢价。其四，AI 转型必须老板一号位推动，从下往上基本推不动；商务签单前先看对接人的权限。获客与筛客的打法也直白到底：对没预算没认知的客户，先收咨询费是最好的筛选器。 还有一条海外参照被现场反复引用：OpenAI 与 Anthropic 分别与 PE 成立合资公司，改造投资组合公司的利润率，本质是用资本关系解决交易摩擦、加速 AI 渗透。硅谷的 startup 有上一代 toB 的积累和 VC 支持，做 FDE 相对容易；而国内 VC 更关注宏大趋势，FDE 爬坡周期长，toB 叙事在中国被系统性低估。计费方法上，一位同时做 FDE 业务的硬件公司 COO 给出三件套：管理层与员工分开培训，诊断层入驻即计费，之后走项目费用制。团队配置公式也在这一场收敛：行业专家，加上有项目经验的工程师。 --- 五、北京 #004：体系、组织与甲方心理学 7 月 4 日，周六下午，35 位。 北京场资源密度最高，光谱最宽。开场第一位发言者是一位数据标注员：他把评测集从表格复制粘贴到对话框的工作用 Python 脚本自动化，打包成了 EXE，看到 FDE 方向后想转型，包装简历投出去发现市场需求真不少，但每次被问到「你怎么收费」就答不上来。同一个下午的另一端，是一家年营收数千万的 AI 品牌营销公司创始人，来招技术型联合创始人；一家深耕医院发票服务的医疗信息服务公司，想深入做医院经济运行管理平台；还有一位近 20 年酒店招商加盟经验的老兵，拿着按年付费、单店 1 万到 4 万的产品设想找研发团队，目标说得很大：一起冲击上市。 两个年轻人的故事值得单独记下。一位刚毕业的成员在乙方广告营销公司从零到一用 AI 重构内容平台业务，一个人对着原来 30 人的团队 PK，批量生成图文和数据看板已经迭代出来，下一步想接 FDE 商单。一位在读研究生做了帮外国游客打车、找餐厅、安排行程的 AI agent，正在探索出海获客。他们的共同点是没有等任何人给许可：先做出来，再来现场找放大器。 这一场的讨论明显偏体系和组织。一家政企数字化服务商在打造 FDE 课程体系、参与标准和认证共建，以运营商客情为切口服务北方重工业和能源企业，现场发布的需求是找优秀 FDE 匹配大 B 订单。有成员想做 FDE 人才平台：找人、培训、对接企业，附带个人 IP 孵化。某车企在通过组织架构变革推动 AI native 改革，构建结构化的 AI coding 知识库。一家大型消费品企业全部门推 AI 化，从业务线抽调员工，同时对外招人。某大厂的经营分析岗给出了内部提效的量化样本：商业分析各环节提效超过 50%，外部工具则因工具本身的问题用了又停。招聘标准的转向由一位 A 轮 AI 硬件公司合伙人说破：从偏工程师转向偏产品经理，偏好年轻、有大厂加创业公司加 SE 服务经验的复合背景；现场另有成员补充观察：顶尖高校的学生学习能力和多端适应力确实强。他分享的规模化样本是一家垂直硬件领域的 FDE 公司，年营收约 2000 万，靠投资人关系转介绍扩客；另一个获客参照是一位服务私募投行的自媒体博主，同业案例的信任背书比跨行案例强得多。 对甲方心理的拆解是四城最犀利的。一位运营多个 AI 产品的成员（其中 AI 论文产品稳定有 10 万级流水）点破：有的公司招 FDE 不是要 ROI，是老板想要一个人放在那里心安。所以第一步要分清楚，老板要的是 ROI 还是心理安抚。一位从垂类智能体转 B 端的成员用自己的弯路作证：合同审核项目遇阻，转做财务和业务数据分析才跑通，他的总结是老板只关注三件事，效果、成本、能不能复制到其他业务，完全不在乎你怎么实现。一位法律背景的成员补充了另一面：客户对 AI 能力边界的期待普遍不明确，边界管理本身就是交付的一部分。 方法论层面的收敛在这一场最清晰。企业 AI 化的第一步是数据化：小公司先做数据库或知识库，大公司才谈业务流程和组织架构重构。知识库是重头戏：要把异构数据和专家知识整合成 agent 能调用的形态，老板最关心的是知识库能不能随业务自进化，这一点目前没有通解，传统 RAG 的准确度在下降。一个可复用的样本是给某职业车队搭的知识库：轮胎、赛车、后勤信息全部结构化，接上协同办公平台做 agentic search，现场的判断是知识库加 data skills 能解决 80% 的 FDE 项目问题。获客建议回到朴素：从身边有信任关系的渠道跑出第一单，形成正向飞轮；选行业比选客户重要，家政这类低毛利、低人才密度的行业被点名支付能力有限；AI 创业应该走出海、外贸这样的增量市场，避开存量市场里降本增效的利益纠葛。 --- 六、横向对比：定价、获客、痛点、人才 先把四城的价格数据点摆在一起。这可能是目前关于中国 FDE 定价最密集的一份公开样本： 深圳 · 制造工厂端侧 AI 项目：预期收益 300 至 500 多万，附数据不上云、交出知识产权等苛刻条件。 深圳 · 跨境电商供应链自动化：降约 60% 人效，转型周期约两个月，每周上百个 SKU。 深圳 · AI 内容获客：4 天 5 万播放量，直接带来企业询单。 上海 · 高端生命支持设备加预测系统整体方案：单案例 100 至 150 万。 上海 · 信创大单：合同约 2 亿。 上海 · 数据标注 agent 化：单月模型消耗 3 万美元，高于原云成本。 上海 · 长期陪跑咨询：按企业规模每年百万量级，合同多签三年。 杭州 · 电商 FDE 小单：单笔 500 至 5000 元，含安装加培训。 杭州 · CRM 加客户画像项目：一期 5 万，整体约 10 万，2 至 3 人团队。 北京 · 纺织喷墨定位外包：2 万包圆，可复制转卖给其他纺织厂。 北京 · 义乌制造业 AI 加 ERP：按裁员比例抽 5% 订阅费，一年 5 万。 北京 · 酒店 AI 产品设想：按年付费，单店 1 万至 4 万。 北京 · 垂直硬件领域 FDE 公司：年营收约 2000 万。 北京 · AI 品牌营销公司：年营收数千万。 北京 · C 端 AI 论文产品：稳定 10 万级流水。 模式上，四城各有偏重，但拼起来是一张完整图谱：项目制（人天、功能点、包圆），订阅制（降本抽成、按年付费），咨询先行（先收咨询费筛客、诊断入驻即计费），绑定制（技术换股权、与 PE 合资的海外参照），复用制（方案模板化转卖、沉淀后收 token 钱）。定价锚点有两个被多城认可：数字员工年薪（按替代岗位的年薪给 AI 服务定价，老板更容易接受），和按降本增效倒算（深圳的按裁员比例分成，北京的抽 5% 订阅）。深圳暴露的问题（抹不开脸收钱、报价没标准），到北京已经有了三阶段计费这类结构化答案：企业内训、驻场诊断收费、开发交付加运维，前两段可以单独买，长期运维建议教会企业内部 IT 自己接手。但「工作流程改善的价值怎么折算成钱」这个问题，在四城都没有解。 获客上，四城打法高度重叠：内容获客（短视频、内容社区，深圳有 4 天 5 万播放带来询单的样本），熟人信任关系跑首单（北京总结为正向飞轮的起点），行业老兵渠道复用（酒店 20 年人脉、佛山义乌产业带资源），驻场切入后横向扩展，转介绍加同业案例背书。差异在客群：上海强调国央企的客情关系和自上而下打法，北京补充了以运营商客情切入政企的路径，杭州最直白，先收咨询费把没预算的挡在门外。 痛点可以归成四层，四城都在说，只是说法不同。需求层：客户说不清痛点，只提 AI 营销、AI 获客等宽泛概念，没有验收标准，乙方不敢承诺效果。数据层：企业数据无法直接给 agent 用，敏感数据难打通，客户甚至对自己的数据质量没自信。组织层：员工抵制、老师傅阻挠、部门墙，提效直接触发裁员焦虑，管理层与员工需要两套叙事。技术与商业层：单个任务成功率再高，多任务串联后整个业务场景的成功率会大幅下降；模型消耗成本不可控；回款周期失控；企业内部改造者议价权低；AI 咨询难规模化。 人才与转型路径的信号在四城逐场升级。深圳出现高校 FDE 项目制培训和应届生自建 POC 求职；上海在摸索「专家加 FDE 加实习生」的产学研派驻模式，博士生靠 vibe coding 补齐算法能力后现场被抢；杭州证明非技术出身可行，财务会计自动化 60% 工作、电商运营、Java 开发各有转法；北京给出了需求侧的标准变化，招人从偏工程师转向偏产品经理，高阶 FDE 可以直通联合创始人席位，课程体系与认证共建已经启动。贯穿四城的共识：主观能动性优先，领域知识是稀缺项，纯工程能力正在被 agent 拉平。 --- 七、趋势判断：未来 6 至 12 个月 判断一：定价将从锚点收敛走向价格战前夜。深圳场已点名「抹不开脸收钱」带来的定价问题，纪要同时记下了价格战与行业整合的预期；上海给出数字员工年薪和陪跑年费百万量级的锚点；北京出现按裁员比例抽 5% 的可复制订阅公式。锚点扩散的速度会快于从业者预期，低端市场（500 至 5000 元的小单）最先卷。 判断二：数据基建将取代 agent 开发成为 FDE 的主战场。杭州场判断业务数据比业务逻辑更重要，北京场总结知识库加 data skills 可解决 80% 的项目问题、企业 AI 化第一步是数据化，深圳场提出独家数据就是切入位点。三城独立收敛到同一结论，意味着未来半年 FDE 项目的第一张报价单大概率是数据治理，而不是 agent。 判断三：客户分层将成为生死线。上海场对国有单位给出「省钱提效叙事走不通」的结构性否定，杭州场指出国央企出详细 PRD 的单好做易赚钱、但和传统外包流程差别不大且难收溢价，北京场点名家政这类低毛利行业付不起 FDE，并主张走出海增量市场。选错客群的团队会在回款环节被拖死（拖了数年没有完成的矿业项目是前车之鉴），行业筛选能力将比交付能力更早分出胜负。 判断四：提效裁员悖论将从道德话题变成交付流程的一部分。深圳的银行工程师自嘲罪人，杭州的钢厂项目死于老师傅抵触，上海已经出现与安置公司合作、在组织诊断阶段谈清补偿与转岗的成熟做法。人员安置方案会进入 FDE 的标准交付物清单。 判断五：FDE 人才供给将在 6 至 12 个月内体系化。深圳的高校项目制培训、上海的产学研派驻、北京的课程体系与认证共建和人才平台创业方向，是同一件事的三个阶段。叠加大厂裁员带来的人才溢出和招人标准转向产品经理的变化，第一批「科班 FDE」会在明年上半年进入市场。 判断六：规模化路径之争将决定资本是否入场。北京场的 VC 坦言大多还在接触阶段、流程尚不明确；「像外包、难规模化」则被现场创业者转述为业界的普遍顾虑。现场已出现三条回应路径：方案沉淀复用后收 token 钱，技术换股权，以及海外模型公司与 PE 合资的参照模式。哪条先跑出可复制样本，决定国内 toB 叙事被低估的局面能否在一年内翻转。 --- 八、结语 四座城市，129 位 builder，四个下午，拼出的是中国 FDE 生态最早的一张地形图：深圳的产业带肉搏，上海的回款清醒，杭州的草根蓝海，北京的体系野心。这个行业还没有名字上的共识，连会议转写工具都拼不对它的全称，但已经有了从 500 元到 2 亿的真实合同谱系。 ha7ch guild 的下一步方向，是把这张地形图变成基础设施：按城市持续办周期性闭门会，让供需两端继续在同一个房间完成闭环；把黑客松和 Sprint 做成年轻人进入真实企业现场的通道；沉淀四城反复出现的定价锚点、筛客方法和交付教训，让下一个转型 FDE 的人不必从「抹不开脸收钱」重新走一遍。 「这两年就是最大的蓝海市场，一定要抢。」这句话出自杭州场。我们把它原样放在这里，一年后回来对答案。