AI on the factory floor starts with the numbers people already report车间里的 AI，从大家已经在报的数开始
No sensors first. In most plants the numbers that matter are already said out loud every shift. The first job is to give them one definition and an owner.不必先上传感器。大多数工厂里，要紧的数每个班次都有人在报。第一件事，是给它们一个口径和一位负责人。
October 7, 2026 · 4 min read2026 年 10 月 7 日 · 阅读约 3 分钟
Most pitches for AI in manufacturing start with sensors: connect the machines, stream the signals, build the model. It is a reasonable vision and a slow first step. Many plants that would benefit most from AI do not have the budget, the integration team or the patience for a year of instrumentation before the first useful answer.
We start somewhere else. In almost every plant, the numbers that matter are already being reported, out loud, every shift. "Line 2 down forty minutes, mould change." "Machine 3 waiting for material." "Night shift: 1,840 good, 22 scrap." They live in group chats, in spreadsheets and in the ERP, and they are rarely defined the same way twice.
The first job of AI on the factory floor is to give those numbers one definition and an owner.
Step one: let the floor report the way it already does
A shift lead should not have to learn a new system to report output. In an ADIS deployment, the production team's digital employee sits in the same chat the shift leads already use.
When someone writes "machine 3 is waiting for material", the digital employee updates the machine's status. When a shift lead posts the night's numbers, it turns them into an output report against the right work order, and when a number is missing it asks for it instead of guessing. Downtime older than a day gets a reminder; maintenance that is due gets one too.
None of this needs new hardware. The reports were already being written. Now they become records.
Step two: one definition of OEE
Every plant measures overall equipment effectiveness, and almost every plant has three versions of it. The plant manager's spreadsheet, the ERP report and the number a shift lead quotes in a meeting rarely agree, because each one counts planned time, rework or changeovers a little differently.
So the definition is written down, once, and computed from the same records for everyone:
- Availability is run hours divided by planned hours.
- Performance is everything produced, good, scrap and rework, divided by run hours times the machine's rated output.
- Quality is good parts divided by everything produced.
OEE is the product of the three. The definition is also part of the acceptance criteria of the deployment, so "OEE went up" means the same thing to the plant manager and to the people who signed the contract.
Step three: count what an idle machine costs
An idle machine is not free. Depreciation, interest, insurance and floor space are paid whether it runs or not, and every idle hour also loses the contribution that an hour of running would have earned.
Take a $1 million imported machine, financed at about 6.5% and depreciated over ten years. It costs roughly $165,000 a year to hold. Over 4,800 planned hours a year, ten more points of OEE add 480 productive hours, which recover about $16,500 of holding cost before a single extra unit of margin is counted.
Once finance enters each machine's holding cost and hourly contribution, idle loss is computed per machine, every day. The question on the plant manager's desk changes from "which line feels slow?" to "which machine is losing the most money, and why?"
Step four: people still schedule
It is tempting to let an algorithm write the schedule. In our deployments, the planner keeps the schedule and the system does the arithmetic.
Each morning the planner sees the work orders due within three days that are not finished, the load on every machine for the next seven days and the delivery risk on each order. When a machine goes down, the planner reassigns the work. The reassignment is an action on record: a reason is required, and the old assignment stays visible, so nobody has to reconstruct later why an order moved.
Step five: connect the floor to the rest of the company
A factory does not end at the shipping dock. The same records connect upstream to quotes and orders, where a quote below the margin floor needs approval, and downstream to invoices, receivables and cash. The value stream view shows every open work order by stage, how long it has been there and how many are past due.
That is where the plant stops being a black box to the rest of the business. Sales can see why an order is late. Finance can see what idle capacity costs this month. The owner sees one set of numbers.
In a factory, the first AI win is rarely a smarter model. It is everyone finally arguing about the same number.
What changes, role by role
- Shift leads report the way they always have, in chat, and stop retyping it into spreadsheets.
- Planners start the day with a list of risks instead of a hunt for them.
- Maintenance gets reminders from the records rather than from memory.
- Plant managers decide what to fix first by cost, not by noise.
- Finance sees capacity loss in the cash view, not in a quarterly surprise.
Sensors and machine connections can come later, and they will be more useful when they arrive, because the definitions they feed already exist. The value starts earlier, with the numbers people are already saying out loud.
讲制造业 AI，大多数方案从传感器开始：把设备连上，把信号接进来，再建模型。这个方向没错，但第一步很慢。最需要 AI 的工厂，往往没有预算、没有集成团队，也等不起一年的改造才看到第一个有用的答案。
我们从别处开始。几乎每家工厂，要紧的数每个班次都已经有人在报：「2 号线停了 40 分钟，换模具。」「3 号机等料。」「夜班：良品 1,840，报废 22。」这些数在群里、在表格里、在 ERP 里，而且很少有两次是按同一个口径算的。
AI 进车间的第一件事，是给这些数一个统一的口径，和一位负责人。
第一步：让车间照原来的方式报
报产量不该让班组长再学一套系统。在 ADIS 的部署里，生产部的数字员工就在班组长们本来就用的那个群里。
有人说「3 号机等料」，数字员工就更新这台设备的状态；班组长发夜班的数，它把数变成对应工单的报工记录；缺了哪个数，它会追问，而不是猜。停机超过一天会提醒，保养到期也会提醒。
这些都不需要新硬件。这些数本来就有人在报，现在它们变成了记录。
第二步：OEE 只有一个口径
每家工厂都算设备综合效率（OEE），几乎每家都有三个版本：厂长的表、ERP 的报表、班组长开会时说的数，很少对得上，因为每个版本对计划时间、返工、换型的算法都略有不同。
所以口径只写一次，所有人都从同一份记录算：
- 可用率 = 开机小时 ÷ 计划小时；
- 性能 = 全部产出（良品 + 报废 + 返工）÷（开机小时 × 额定产能）；
- 质量 = 良品 ÷ 全部产出。
OEE 是三者相乘。这个口径同时写进部署的验收标准，于是「OEE 提升了」对厂长、对签合同的人，是同一个意思。
第三步：算清楚一台闲置设备要花多少钱
闲置的设备不是免费的。折旧、利息、保险、厂房面积，开不开机都要付；每闲一个小时，还少了一小时开机本该挣的边际贡献。
以一台 100 万美元的进口设备为例，按约 6.5% 融资、10 年折旧，一年的持有成本约 16.5 万美元。按一年 4,800 个计划小时算，OEE 每提高 10 个点，就多出 480 个有效小时，先收回约 1.65 万美元的持有成本，这还没算多出来的产品毛利。
财务把每台设备的持有成本和每小时边际贡献填进去之后，闲置损失按设备、按天算出来。厂长桌上的问题，从「哪条线感觉慢」变成「哪台设备亏得最多，为什么」。
第四步：排产还是人来排
让算法直接排产，很诱人。在我们的部署里，排产归计划员，系统负责算。
每天早上，计划员先看到三天内到期还没完成的工单、未来 7 天每台设备的负荷，以及每张订单的交付风险。设备坏了，计划员改派。改派是一次留痕的动作：必须写原因，原来的安排也保留着，以后没人需要去回忆一张单为什么挪了。
第五步：把车间接到公司的其他部分
工厂不止于发货口。同一份记录往上接报价和订单——报价低于毛利底线要审批；往下接发票、应收和现金。价值流视图按阶段列出每一张在制工单：在这个阶段停了多久、有几张已经逾期。
从这里开始，车间对公司其他部门不再是一个黑箱。销售看得到订单为什么晚了；财务看得到这个月闲置产能花了多少钱；老板看到的是一本账。
在工厂里，AI 的第一场胜利很少来自更聪明的模型，而是大家终于在争论同一个数。
每个岗位变了什么
- 班组长照常在群里报数，不用再往表格里抄一遍。
- 计划员每天从一张风险清单开始，而不是到处找风险。
- 设备维护的提醒来自记录，不靠记性。
- 厂长按成本决定先修什么，而不是按谁嗓门大。
- 财务在现金视图里看到产能损失，而不是等季度末的意外。
传感器和设备直连可以以后再上。等它们来的时候会更有用，因为它们要喂进去的口径已经在了。价值来得更早——从大家已经在说出口的那些数开始。
AIDC · AI Deployment CompanyAIDC · AI Deployment Company
