同界数科为离散制造企业交付岗位级数字员工——不是又一个问答助手，而是能接管报价、订单录入、成本稽核与经营分析四个岗位，对每一张报价单的毛利负责的「同事」。AgentX delivers position-level digital employees to discrete manufacturers. Not another chat assistant — a colleague that owns quoting, order entry, cost audit and management reporting, and stands behind the margin on every quote it issues.
不是因为没有系统，而是因为报价发生在系统之外——发生在老师傅的经验里、Excel 里、微信里。等到财务月结，亏损已经发生，且无法归因到某一张单。Not for lack of systems — because quoting happens outside them: in a veteran estimator's head, in spreadsheets, in chat threads. By month-end close the loss has already happened, and nobody can trace it to a specific order.
定价逻辑装在一两个人的脑子里，没有写下来，也没有校验。他们判断准确，但没人知道为什么准；他们一旦离开，企业的定价能力归零。Pricing logic lives in one or two heads — never written down, never checked. They are accurate, but nobody knows why. When they leave, pricing capability resets to zero.
物料主数据一物多码、BOM 版本与现场不一致、工艺路线仍在纸上、采购价散落在多个台账。想算一张单的真实成本，先要打三通电话。Duplicate item codes, BOMs out of sync with the shop floor, process routes still on paper, purchase prices scattered across ledgers. Costing a single order starts with three phone calls.
没有逐单核算，就没有反馈闭环。负毛利订单进了系统，被总量毛利掩盖，明年客户续单，同样的价格再来一次。No per-order costing means no feedback loop. Negative-margin orders slip in, get masked by blended margin, and the customer reorders at the same price next year.
行业研究反复指出，整厂级「AI 工厂大脑」类项目的落地失败率居高不下——范围太大、边界太模糊、责任无法认定。我们把交付单位缩小到一个可被考核的岗位：数字员工对这个岗位的输入、流程、产出与质量负责。Industry research keeps finding the same thing: whole-plant "AI factory brain" programs fail at a stubbornly high rate — scope too wide, boundaries too vague, accountability impossible to assign. We shrink the unit of delivery to a single measurable role, and make the digital employee own its inputs, workflow, output and quality.
人发起、人拼接、人兜底。AI 只是流程里的一次调用，出了错也说不清是谁的责任。A human starts it, stitches it and backstops it. The AI is one call in the flow — and when it's wrong, nobody owns it.
以岗位说明书为边界，从接收任务到交付结果自主闭环，并对结果的毛利与合规负责。Bounded by a job description, it closes the loop end to end — and answers for the margin and compliance of what it delivers.
企业信息化有两层是有人管的：个人效率层与系统管理层。中间那一层——数据在哪里发生、由谁负责、算得对不对——长期无人认领。那正是同界数科的位置。Two layers of enterprise IT already have owners: personal productivity and system-of-record management. The layer in between — where data originates, who is accountable for it, whether the number is right — has never had one. That's where we sit.
每一名数字员工都有一份岗位说明书：职责边界、接管的动作、交付物、考核指标。你可以像招人一样评估他们，也可以像管人一样管理他们。Each comes with a job description: scope, the actions it takes over, its deliverables and its KPIs. Evaluate them like candidates. Manage them like staff.
这不是宣传片，而是产品操作回放——左侧下达任务，右侧工作区逐步执行、逐步留痕。点击播放，或切换岗位。Not a promo film — an actual product replay. Assign the task on the left; watch the workspace execute and leave a trace on the right. Press play, or switch roles.
选择一个岗位，点击「派发任务」，观察从感知、规划、执行、复核到交付的完整闭环。Pick a role, hit dispatch, and watch the loop close: perceive, plan, execute, review, deliver.
制造业最贵的资产往往不在报表上——它在几位老师傅几十年积累的判断力里。我们不是把他们替换掉，而是把他们的判断显性化、可复用、可传承，让他们从「报价机器」变成「规则的定义者」。A manufacturer's most valuable asset rarely appears on the balance sheet — it sits in a few veterans' decades of judgment. We don't replace them. We make their judgment explicit, reusable and inheritable, so they move from being the quoting machine to being the author of the rules.
结构化访谈 + 真实单据复盘，记录老师傅在报价时实际看了什么、算了什么。Structured interviews and real-order walkthroughs capture what they actually look at and compute.
把一次报价拆成 20–40 个判断点：哪一步靠数据，哪一步靠经验，哪一步靠客户关系。Break one quote into 20–40 judgment points: data-driven, experience-driven, relationship-driven.
用历史成交数据回测每条规则，把「感觉」变成有置信区间的可验证规则。Back-test each rule against historical wins, converting "feel" into verifiable rules with confidence bounds.
常规询价由数字员工独立完成，老师傅只处理规则外的例外与新品类。The agent handles routine RFQs; the veteran only touches exceptions and new categories.
中标与实际毛利回流为反馈，规则持续修正——经验不再随人流失，而是复利增长。Win rates and realized margin feed back; rules keep improving. Experience stops leaking and starts compounding.
标准化，是规范化的前提；规范化，是资本化的前提。
把定价能力从「某个人会」变成「这家公司会」，是企业估值的一部分。Standardization precedes governance; governance precedes capitalization.
Moving pricing capability from "one person can do it" to "this company can do it" is part of what the company is worth.
数字员工要在制造现场可靠工作，需要两层地基：一层告诉它「世界由什么构成」，另一层告诉它「智能体在这个世界里该如何行动」。前者是本体（Ontology），后者是我们自研并公开发表的 BUN / AIB 智能体行为框架。To work reliably on a shop floor, a digital employee needs two foundations: one that says what the world is made of, and one that says how an agent should act within it. The first is Ontology. The second is BUN / AIB — our own published agent-behavior framework.
物料、BOM、工艺路线、工序、设备、工单、客户、合同——建成对象与关系网络，而非一张张孤立的表。参考 ISA-95 层级模型。Items, BOMs, routings, operations, equipment, work orders, customers, contracts — modeled as an object graph rather than disconnected tables, aligned to the ISA-95 hierarchy.
为智能体定义信念、效用与规范边界，使其在不确定环境中的行动可预测、可解释、可约束——这是「岗位级」区别于「工具级」的技术分界线。Defines an agent's beliefs, utilities and normative bounds so its behavior under uncertainty stays predictable, explainable and constrained — the technical line between position-level and tool-level.
开源可自主可控的数据管道：采集、入湖、质量校验、编排与建模，全链路血缘可追溯，不锁定单一厂商。An open, self-controllable pipeline: ingestion, lakehouse, quality gates, orchestration and modeling — full lineage, no vendor lock-in.
标准化对接主流 ERP / MES / PLM / OA 与自研系统；二期通过工业协议接入设备层，打通计划与实际。Standard connectors to mainstream ERP/MES/PLM/OA and in-house systems; phase two reaches the equipment layer over industrial protocols.
全过程留痕、结果自动复核、关键动作强制人工确认、权限与数据边界隔离，满足审计与合规要求。Full traceability, automatic self-review, mandatory human sign-off on critical actions, and strict permission and data boundaries.
为每名数字员工建立评估基准与 KPI 看板，像考核员工一样考核 AI：达不到指标就是不合格，而不是「AI 就这样」。Every digital employee gets a benchmark and a KPI dashboard. Miss the target and it failed review — not "well, that's AI."
制造业不需要又一个「先签三年、慢慢见效」的信息化项目。我们从零定金的盲测开始：用你自己的历史数据跑一遍，结果对不上，不进入下一阶段。Manufacturers do not need another three-year IT program that pays off "eventually." We start with a zero-deposit blind test on your own historical data. If the numbers don't hold up, there is no next phase.
你提供一批已成交的历史询价单（隐去成交价），我们的报价数字员工独立测算。当场比对偏差率，你判断值不值得继续。零定金。You hand over closed historical RFQs with the final price masked. Our quoting agent computes independently. We compare deviation on the spot, and you decide whether to continue. Zero deposit.
只用已有的静态数据——销售订单、BOM、采购价、工艺路线——建立本体模型与数据管道，输出第一张全量毛利地图。不碰设备、不碰产线，因此风险低、见效快。Using only static data you already have — sales orders, BOM, purchase prices, routings — we build the ontology and pipeline and produce your first full margin map. No equipment, no line changes: low risk, fast proof.
完成老师傅经验萃取与规则回测，两名数字员工进入真实业务流并行运行；人机双轨对比一个周期后逐步移交。After know-how extraction and rule back-testing, two digital employees run in parallel with the human process for one cycle before handover.
成本稽核持续运行，毛利回收形成可量化收益；管理驾驶舱面向决策层，让经营状况从「月结才知道」变成「随时可见」。Cost audit runs continuously and margin recovery becomes a measurable return; the cockpit turns "we find out at month-end" into "we can see it now."
在管理层数据闭环稳定后，再通过工业协议接入设备与产线数据，把「计划成本」升级为「实际成本」，并解锁产能利用率相关场景。Once the management-layer loop is stable, connect equipment and line data over industrial protocols to move from planned to actual cost, unlocking utilization use cases.
关于 ERP 上线时点：On ERP go-live timing: 如果你正在或即将上线新的 ERP，主数据治理必须走在上线之前，而不是之后。上线后再补料号、补 BOM、补标准成本，成本是数倍的。这是我们在制造业项目里最常见、也最昂贵的时序错误。If you are rolling out a new ERP, master-data governance must come before go-live, not after. Backfilling item codes, BOMs and standard costs post-go-live costs several times more. It is the most common — and most expensive — sequencing mistake we see.
我们不做所有行业。岗位级数字员工在下面这类企业里价值最高——它们的共同点是：多品种、小批量、非标报价、经验驱动。We don't serve every industry. Position-level digital employees pay off most in companies with one shared shape: high-mix, low-volume, non-standard quoting, expertise-driven.
需要，而且顺序很重要。ERP 管的是「已经发生的事实怎么记录」，我们管的是「事实是怎么产生的、算得对不对」。ERP 的报表质量完全取决于录进去的数据质量，而录入端一直是人。更实际的一点：主数据治理如果不走在 ERP 上线之前，上线后补的成本是数倍。Yes — and the order matters. An ERP records facts that already happened; we govern how those facts are produced and whether they're right. ERP reporting quality is capped by input quality, and input has always been human. Practically: if master-data governance doesn't precede go-live, backfilling afterwards costs several times more.
差在交付单位。订阅制助手交付的是「个人效率」——帮某个人写得更快、查得更快，用不用、用得对不对都取决于这个人。我们交付的是「岗位产出」：接入你的 ERP 与主数据、承担这个岗位的 KPI、结果可审计、出错可追责。两者不冲突，也不互相替代——但只有后者能改变毛利。The unit of delivery. A subscription assistant delivers personal productivity — one person drafts and searches faster, and whether it's used well depends entirely on them. We deliver role output: connected to your ERP and master data, carrying the role's KPIs, auditable, and accountable when wrong. They don't compete — but only one of them moves gross margin.
数据乱是常态，不是例外——我们见过的每一家制造企业都是如此。Phase 1 的目标就是治理，不是假设你已经治理好了。真正的判断标准不是「干不干净」，而是「能不能重建」：只要历史销售订单、BOM 与采购记录还在，单件真实毛利就能被还原。盲测阶段我们就会给出明确的数据现状诊断，能做多少、做不了什么，说清楚再谈钱。Messy is the norm, not the exception — it's true of every manufacturer we've worked with. Phase 1 exists to fix it, not to assume it's already fixed. The real test isn't "is it clean" but "is it reconstructable": if historical sales orders, BOMs and purchase records exist, true unit margin can be rebuilt. The blind test produces an honest data diagnosis first — what's achievable and what isn't — before any money changes hands.
这是项目最真实的风险，我们不回避。做法上：数字员工承接的是重复性询价，老师傅转向规则定义、例外裁决与新品类定价——职责升级而非削减，且在组织内明确署名为规则作者。执行上，经验萃取必须由管理层背书、并与激励挂钩；如果这一条谈不拢，我们会建议先不要启动，因为强行推进的成功率很低。It's the most real risk in the project and we don't dodge it. Structurally: the agent takes routine RFQs while the veteran moves to defining rules, adjudicating exceptions and pricing new categories — a scope upgrade, with explicit credit as the author of the rules. Operationally, know-how extraction needs executive backing and an incentive tied to it. If that can't be agreed, we recommend not starting — forcing it rarely works.
主收益来自三处，都可量化：一是负毛利订单的事前拦截与调价回收，二是报价响应时长缩短带来的成交率提升，三是关键岗位对个人依赖的解除。Phase 1 结束时你会拿到全量毛利地图与负毛利清单——那一刻就能算出可回收的金额上限，再决定后面投多少。我们不用「效率提升 30%」这类无法验证的口径。Three quantifiable sources: blocking and repricing negative-margin orders, win-rate gains from faster quote turnaround, and removing single-person dependency on a critical role. At the end of Phase 1 you hold a full margin map and loss list — enough to size the recoverable upside before committing further. We don't quote unverifiable "30% efficiency gain" numbers.
零定金盲测：你出题，我们答；对不上，不继续。两周内给出偏差率报告与数据现状诊断。Zero-deposit blind test: you set the questions, we answer. If it doesn't hold up, we stop. Deviation report and data diagnosis within two weeks.