VWD AI Guide v3.9

IMP 库存优化:给采购与排产装一个"决策大脑" IMP Inventory Optimization: A Decision Brain for Purchasing & Planning

内部分享 · 排产 & 采购部 · 跨月推进中 Internal Share · Planning & Procurement · Ongoing

J

James Ye

CEO · 项目发起人 / 领导CEO · Project Sponsor / Lead

K

Kai

生产计划经理 · 项目双负责人Production Planning Manager · Co-Lead · kai@valuewds.com

M

Martin

采购经理 · 项目双负责人Purchase Manager · Co-Lead

系统能算出"哪天会缺料",比我们凭经验拍脑袋早得多。我们不写代码——用中文把业务规则讲清楚,Claude 把它变成每天自动跑的工具。 The system can tell us which material will run short, and when — far earlier than gut feel ever could. We don't write code: we explain the business rules in plain language, and Claude turns them into a tool that runs itself every day. — Kai & Martin
📌 一句话看懂这个项目The Project in One Line

不替换现有的 NetSuite 和 Cantor,而是在它们之上加一层 "决策智能层":把分散在两个系统里的需求、库存、在途采购、真实交期拼到一起,每天自动算出"什么料、大概什么时候、该下单还是该催货还是该空运",把判断结果只推给排产和采购。

It doesn't replace NetSuite or Cantor — it adds a decision-intelligence layer on top. It stitches together demand, inventory, in-transit purchasing, and real lead times that were scattered across two systems, and each day it works out which material is heading for a shortage, roughly when, and whether to reorder, expedite, or air-freight — routing those calls only to planning and procurement.

1. 我们遇到了什么问题The Problem We Were Sitting On

门窗生产高度依赖进口铝型材和配套材料。这些材料的真实采购周期比系统里设定的补货触发点长得多——也就是说,等系统提示"该补货"时,其实早就晚了。叠加单一来源、供应商质量波动、个别配件供应链特别长等情况,缺料停产、客户交期被迫推迟成了反复发生的痛点。

Our production depends heavily on imported aluminum profiles and related materials. Their real procurement cycle is far longer than the reorder trigger configured in the system assumed — so by the time the system said "reorder," it was already too late. Combined with single-source supply, supplier quality swings, and a few components with especially long supply chains, material shortages and slipped customer dates became a recurring pain.

⚠️ 以前Before

  • 需求在 Cantor、采购与库存在 NetSuite,两边对不上
  • Demand lived in Cantor, purchasing & inventory in NetSuite — never lined up
  • 补货触发点远短于真实交期 → 系统性偏晚
  • Reorder trigger far shorter than real lead time → systematically late
  • 缺料靠经验和救火,等发现往往来不及
  • Shortages caught by gut feel and firefighting — usually too late
  • 两个仓库的库存要人工拼,容易误判"有货/没货"
  • Stock across two warehouses pieced together by hand — easy to misjudge
  • 缺料停产 + 客户交期被迫推迟
  • Stoppages from shortages + slipped customer dates

✅ 现在Now

  • 两个系统的数据每天自动拼到一张图上
  • Both systems' data auto-stitched into one picture daily
  • 提前算出"哪天会断料",留出下单/催货时间
  • Forecasts the day each material runs short, with lead time to act
  • 自动给出建议动作:下单 / 催货 / 空运
  • Suggests the action automatically: reorder / expedite / air-freight
  • 合并两仓库存,消除"假缺料 / 假有货"
  • Merges both warehouses, killing false shortages / false stock
  • 从"事后救火" → "提前预警"
  • From firefighting → early warning

2. 我们搭了一套什么系统What We Built

核心思路是"不动现有系统、只加一层判断"。Cantor 继续负责销售订单、工单、产能和交期;NetSuite 继续负责采购、收货、库存。新增的智能层只做一件事:把这些数据读过来,按业务逻辑算出材料风险,再把结论交给人去决策。整个判断拆成 5 层逻辑:

The core idea is "don't touch the existing systems — just add a layer of judgment." Cantor still owns sales orders, work orders, capacity, and delivery dates; NetSuite still owns purchasing, receiving, and inventory. The new layer does just one thing: read that data, reason about material risk by business rules, and hand the conclusion to a human to decide. The reasoning is broken into five logical layers:

1

需求Demand

从已签订单、排产计划和报价推算未来要用掉多少材料。

Derives how much material upcoming orders, the production plan, and quotes will consume.

2

供应Supply

合并多个仓库的真实库存 + 在途采购,得到"手上加在路上"的总量。

Merges real on-hand stock across warehouses with in-transit purchasing for a true "on-hand + inbound" total.

3

净缺口Net Gap

需求减供应,沿时间轴画出每种材料的"净位置曲线",看它哪天跌破零。

Demand minus supply, plotted as a net-position curve over time to see the day each material dips below zero.

4

真实交期Real Lead Time

不按单个料号、而是按"采购路径"把材料分桶,用每桶的实际历史交期,而不是理论值。

Groups materials into buckets by procurement path (not per part number), using each bucket's actual historical lead time rather than a theoretical one.

5

跟单Follow-Up

跟踪每张采购订单的全生命周期,卡在某个阶段太久就自动报警。

Tracks each purchase order through its full lifecycle and flags any that stalls too long at a stage.

3. 每天早上自动跑一遍It Runs Itself Every Morning

整套流程设成每个工作日早上自动刷新:从两个系统拉最新数据 → 合并库存 → 算净位置 → 生成预警 → 更新看板。团队上班打开看板就是最新结果,没人需要手动操作。

The whole pipeline is set to auto-refresh every workday morning: pull the latest data from both systems → merge inventory → compute net position → generate alerts → update the dashboard. The team opens the dashboard to fresh results — no one has to run anything by hand.

⬇️拉数据Pull Data
从 NetSuite / Cantor 取库存与在途Inventory & inbound from NetSuite / Cantor
→
🧮合并 & 计算Merge & Compute
合并多仓库存,算净位置曲线Merge stock, build net-position curves
→
🚨生成预警Generate Alerts
标出建议动作:下单/催货/空运Tag the action: reorder / expedite / air
→
📊更新看板Refresh Dashboard
多标签看板一键刷新Multi-tab dashboard refreshed
💡 无人值守Hands-Off

用系统的定时任务每天定点触发,错过了也会自动补跑。需要立刻看最新数据时,也可以手动一键刷新。A scheduled task fires it at a set time each day and catches up automatically if a run is missed. Anyone who needs fresh numbers right away can also trigger a one-click refresh.

4. 一个看板,六个视角One Dashboard, Six Views

所有结果汇总在一个看板里,按标签切换。下面是每个视角"回答什么问题"——不涉及任何具体数字。

Everything lands in one dashboard, switchable by tab. Here's the question each view answers — no specific figures involved.

🚨 缺料预警Shortage Alerts

哪些材料正走向短缺、大概哪天断、该采取什么动作。点开某一行能展开它的未来走势曲线。

Which materials are heading for a shortage, roughly when, and what to do. Click a row to expand its forward trajectory curve.

📉 净位置Net Position

每种铝材未来一段时间的"手上+在途 减 需求"曲线,采购到货当天会出现一个台阶式跳变。支持按系列、窗型筛选搜索。

A forward curve of "on-hand + inbound minus demand" for each profile, with a step jump on the day a PO arrives. Filterable and searchable by series and window type.

♻️ 损耗率Loss Rate

实际用量与理论用量的对比,帮助发现哪些环节损耗偏高、值得追查。

Actual versus theoretical consumption, surfacing where loss runs high and is worth investigating.

🎯 报价胜率Win Rate

从客户、产品系列、销售等多个视角看报价转化情况,给需求预测提供参考。数据不足时会显式标注,绝不静默猜测。

Quote-to-order conversion seen by customer, product series, and salesperson — feeding demand forecasting. When data is thin it says so explicitly, never silently guessing.

📐 备料长度建议Stock Length

基于历史切料情况,建议每种型材按哪个长度备货最省料,并给出操作上可行的方案。

Based on historical cutting patterns, recommends the most material-efficient stock length per profile — and one that's practical on the floor.

📦 采购跟单PO Tracking

把每张采购订单的状态、装柜分配和催货优先级集中展示,让采购知道"先催哪一单"。

Centralizes each PO's status, container allocation, and expedite priority so purchasing knows which order to chase first.

5. 几个关键的思路A Few Key Ideas

这个项目真正的价值不在"代码",而在把业务里的隐性经验变成可计算的规则。几个例子:

The real value isn't the code — it's turning tacit business know-how into computable rules. A few examples:

🪣 按"采购路径"分桶,而不是按料号Bucket by procurement path, not by part number

几千个料号一个个估交期不现实。改成按"采购方式"把材料归成几类,每类用它自己的真实历史交期,既准又好维护。Estimating lead time for thousands of part numbers one by one isn't realistic. Instead, materials are grouped into a handful of buckets by how they're procured, each using its own real historical lead time — more accurate and far easier to maintain.

🔁 动态补货点,跟着真实交期走A dynamic reorder point that tracks real lead time

老的补货点是个固定值,且远短于材料真正到货所需的时间——这正是系统性缺料的根因。新方法让补货点跟随每桶的真实交期和波动自动调整,体现"不缺料优先"。The old reorder point was a fixed value, and far shorter than how long materials actually take to arrive — the root cause of systematic shortages. The new approach lets it adjust to each bucket's real lead time and variability, reflecting a "don't run out" priority.

🏭 合成库存:把两个仓库当一个池子Synthetic on-hand: treat two warehouses as one pool

材料分散在批发仓和工厂仓。系统把两边合并,并按入库、盘点、日常消耗三种事件自动调整,避免"这边显示没货、那边其实有货"导致的误判空运。Material sits in both a wholesale and a factory warehouse. The system merges them and auto-adjusts for three kinds of events — inbound, physical count, and daily consumption — avoiding the false "air-freight now" calls that come from one side showing empty while the other has stock.

🚦 催货优先级:先做最急的Expedite priority: do the most urgent first

采购同事手里的在途订单和跟单记录被整理成统一清单,每个未发货项目按"还能撑多久"自动分高/中/低优先级,采购可以直接拿去跟供应商说"先做这几单"。Purchasing's in-transit orders and follow-up records are pulled into one list, where each unshipped item is auto-ranked high / medium / low by how long stock will last — so purchasing can tell the supplier exactly which to make first.

🛡️ 永不静默猜测Never silently guess

数据缺失时绝不偷偷用默认值,而是显式标出"这里没数据"并单独列出来。错误的结论比"暂时没有结论"危险得多——宁可空着,也不蒙混。When data is missing, the system never quietly substitutes a default — it explicitly flags "no data here" and lists it out. A wrong conclusion is far more dangerous than "no conclusion yet" — better blank than bluffed.

6. 他们是怎么"不写代码"做出来的How They Built It Without Coding

和我们其它案例一样:业务的人讲清楚规则,Claude 写引擎、调试、改需求。下面是真实迭代过程的缩影。

Like our other cases: the business owners explain the rules, and Claude writes the engine, debugs, and revises. Here's a condensed look at the real back-and-forth.

K
我想知道每种铝材按现在的库存和在途采购,未来一段时间哪天会断料I want to know, for each aluminum profile, which day it'll run short given current stock and inbound POs
AI
明白。每种料的日消耗用什么口径?是按订单×BOM 算理论用量,还是用工厂实际的用量?Got it. What should daily consumption be based on — theoretical usage from orders × BOM, or the factory's actual usage?
K
用工厂的实际月用量更准。库存要把批发仓和工厂仓合起来算The factory's actual monthly usage is more accurate. And stock has to combine the wholesale and factory warehouses
AI
好。我先做一版:合并两仓库存 + 在途采购 − 实际日耗,画出每种料的未来曲线,跌破零就标预警。跑出来你看数字对不对?Sure. First pass: merge both warehouses + inbound − actual daily usage, plot each material's forward curve, flag where it crosses zero. Run it and check the numbers?
M
大方向对,但有几种料其实工厂仓有货,系统却报了空运——是不是漏算了工厂仓?Right direction, but a few materials show "air-freight" when the factory warehouse actually has stock — is it missing the factory side?
AI
对,之前只算了批发仓。已经把工厂仓合并进来,并按入库/盘点/消耗自动调整。再跑一次,假空运应该消失了Right — it was only counting the wholesale side. I've merged the factory warehouse in and made it adjust for inbound / count / consumption. Re-run it; the false air-freight calls should be gone
M
好多了。能不能把我维护的那张在途订单表也接进来,顺便标出哪几单最急?Much better. Can you pull in the in-transit order sheet I maintain too, and flag which ones are most urgent?
AI
可以。我把你的表解析进来,每个未发货项目按"还能撑多久"分高/中/低优先级,放进采购跟单看板,点开能看每条明细和装柜分配Done. I parse your sheet, rank each unshipped item high / medium / low by how long stock lasts, and put it in the PO-tracking view — click to see each line and its container allocation

7. 先"陪跑",再信任Shadow First, Trust Later

👥 只给排产和采购Planning & Procurement only

材料告警只推给排产和采购两个角色,不直接进销售环节——避免半成熟的判断影响对客户的承诺。

Material alerts go only to planning and procurement, never straight into sales — so half-baked judgments don't shape customer commitments.

🌓 影子模式验证Shadow-mode validation

新算法先和现有流程"并行陪跑"一段时间,持续比对准确度,达标后再考虑让它正式参与决策。

New logic runs in parallel with the existing process for a while, accuracy compared continuously, before it's trusted to formally inform decisions.

🗂️ 数据分级Data classification

项目严格区分可共享数据与敏感数据(单价、库存金额、供应商报价等),敏感信息只在受限范围内流转。

The project strictly separates shareable data from sensitive data (unit prices, inventory value, supplier quotes, etc.); sensitive items stay within a restricted circle.

🧩 分清谁是源头Clear source of truth

需求看 Cantor、采购库存看 NetSuite,同一份数据只认一个"源头系统",避免两边打架。

Demand from Cantor, purchasing & inventory from NetSuite — each dataset has exactly one system of record, so the two never contradict.

8. 这个案例告诉我们什么What This Case Shows

🧠

经验可以变成规则Tacit → Rules

老师傅"凭感觉"的判断,讲清楚后就能变成每天自动跑的逻辑

A veteran's gut call, once explained, becomes logic that runs every day

🔗

打通数据孤岛Bridge silos

不替换系统,用 AI 在两个系统之上做"翻译"和综合判断

Don't replace systems — let AI translate and reason across them

⏰

从救火到预警Firefight → Forecast

把"事后才发现"变成"提前几周就知道",留出行动时间

Turn "found out too late" into "known weeks ahead," with time to act

🤝

业务人主导Business-led

排产和采购的人自己主导,IT 提供数据接口配合

Planning and procurement lead it; IT supports with data access

🛡️

谨慎上线Cautious rollout

先陪跑验证、只给相关角色、敏感数据分级——稳扎稳打

Shadow first, scoped audience, classified data — step by step

📈

从小处滚起来Start small

先做一个最痛的点,跑通了再一块块往外扩

Solve the most painful piece first, then expand one block at a time

🌱 最懂业务的人 + AI = 最强组合🌱 The People Who Know the Business + AI = the Strongest Team

这个项目最关键的不是技术,而是排产和采购的同事最清楚"哪里痛、规则是什么"。你不需要会写代码——只要把业务讲清楚,AI 就能把它变成每天帮你干活的工具。从你手上最头疼的那件事开始。The key to this project wasn't the technology — it was that the planning and procurement team knew exactly where it hurt and what the rules were. You don't need to code: explain the business clearly, and AI turns it into a tool that works for you every day. Start with whatever frustrates you most.

VWD AI Guide v3.9 | VWD | 2026 | Created by IT depart. | 有问题发邮件到Questions? Email it@valuewds.com