The workshop is two-thirds faster now, but the man-hours on the quotation sheet are still from last year's book

Last month I went to a factory that makes injection-molded parts. The workshop director pulled me over to see the newly installed system: after handing injection molding machine parameter tuning over to an AI agent, it took only 5 trials for a mold to produce qualified parts, down from more than ten; debugging time was cut by two-thirds. When he said this, his eyes lit up, because he had waited three years for this number.

Three months later, at the finance meeting, the problem had flipped: gross margin hadn't moved. Quotes were still given at the old prices, costs were still calculated using the old norms, and the workshop was indeed faster—but the part that got faster hadn't made it into any table.

This is not an isolated case. On November 17, CCTV's Morning News aired a report on Xiaomi's Wuhan smart home appliance factory: 120,000 data collection points, collecting more than 1.1 billion pieces of data every day, and more than 600 measurement points inside a single smart injection molding machine. After AI parameter tuning, it similarly "only takes 5 runs to produce qualified products, and debugging time is also shortened by two-thirds." Factory director Xu Ying said something very apt in the interview: Using AI is not about having fewer people or no people; the key is to keep making the granularity smaller and more precise.

This statement is about manufacturing. But for corporate management, the real question is the other half: the granularity in the workshop has become finer—has the granularity of bookkeeping followed suit?

1. Those 120,000 points are not actually measuring "efficiency"

If you break down CCTV's on-site details, you'll find that what it describes is far more specific than "efficiency improvement."

More than 600 measurement points inside the injection molding machine sense temperature, humidity, and raw materials. These data generate causal chains in real time, and the machine uses them to determine where the process parameters should be adjusted at this moment. What is changed is not "how fast people work," but "where the parameters start from." In the past, parameter tuning relied on the feel of experienced masters; a mold would be tried a dozen or so times, relying on intuition accumulated from experience; now, starting from the causal chains in the data, it converges in 5 attempts.

In other words, what is compressed is the number of trial and error attempts; when it comes to the process, it is the time for the "mold trial" step.

Trial molding time plays an awkward role in cost accounting. It is not normal production labor hours; many companies allocate it proportionally into manufacturing overhead, or record it separately as a "development fee" attached to the project. Either way, it has a default quantity—in the past, how many hours corresponded to a dozen or so trial runs; this number was written into the standard cost and indirectly into the quotation.

Now the number of mold trials has dropped by two-thirds. If this change is not synchronized into the standard quota, a hard-to-notice consequence will arise: quotes are marked up based on the cost of "needing more than a dozen trials," while in reality only 5 trials were run, and the price difference in between does not become profit but instead becomes "we are more expensive than others." When customers take it for price comparison, what they compare is exactly this difference.

A counterintuitive point: A faster workshop looks like a good thing in the short term, but in the medium term it means distorted costs. Because the faster part has no corresponding account item: it appears neither in labor hours nor in expenses, but quietly disappears between the two statements.

II. From "quick adjustment" to "accurate calculation," there are three procedures in between

In several projects, I have repeatedly seen the same sequencing problem: deploy the equipment first, then the systems, and only at the end remember that the quotas need to be changed. And changing quotas is not as simple as opening a field and changing a number.

First: The process must be broken down enough to be timed

Is mold trial one process step or three? If the process route has only “injection molding” as the single step, and mold trial is mixed with production, then debugging time drops by two-thirds, but you can’t even pinpoint “which number should be changed.”

The approach of Gaoke Rubber & Plastics in Jing County, Hebei, is worth referencing: a November 17 report in Guangming Daily said that this enterprise has connected five systems, including ERP and MES, with online work reporting for 13 processes and four layers of quality inspection interception. Thirteen processes may not be the optimal solution, but it provides a lower bound—only when processes are broken down properly does work reporting have a place to land; and only when work reporting has a place to land can actual working hours be numbers rather than estimates.

Second: quotas need someone responsible for revising them

In Odoo, the object corresponding to this operation is the operation line on the routing (mrp.routing.workcenter), where the labor/machine time has a mode choice between "Manual" and "Calculate based on actual". Most companies choose Manual when going live, enter an estimated value, and then no one ever touches it again.

The problem isn't in the fields; it's in accountability. Who changes the quota, under what conditions, who reviews it after the change, and whether sales should be notified after the change—if these four things aren't written into a process, the quota will forever remain at the version from the day of project acceptance. I saw something even more extreme at a customer: for the same part, the work hours in the quoting system and the work hours in MPS scheduling are not the same number; two separate teams calculate their own, neither is wrong, but together they are wrong.

Third: Losses must have a reason code

The most deceptive thing about work reporting data is that it only records "how long was used" and not "why it took so long." In Odoo's operation work reporting records (mrp.workcenter.productivity), there is a loss reason field that can distinguish effective working time, performance loss, and availability loss. If this field is left blank, the reported working hours are just a total number.

Here's a concrete difference: a certain process has a standard of 40 minutes but actually takes 55 minutes. If you only record 55 minutes, you only know it's 15 minutes slower; you don't know whether it was equipment failure, waiting for materials, changeover, or inexperienced operation. The improvement actions corresponding to these four causes are completely different—equipment failure goes to the equipment ledger and maintenance plan, waiting for materials goes to inventory and procurement, changeover goes to process engineering, and inexperienced operators go to training. Merge everything into a single number, and management actions lose their anchor.

3. Those few numbers in Odoo—change one and it sets off a chain

Many companies think that 'changing quotas' is a finance matter, but in fact it affects at least five areas. Taking the models from Odoo 17 to 19 as an example (fields may be slightly adjusted in version 20; verify before implementation), it is roughly a chain like this:

Change thisWhat does it affect?What happens if it is not changed?
Operation hours (the hours field on the routing)Standard cost, quotation, production scheduling takt timeQuotes are either too high or too low, production scheduling follows the old takt, and delivery commitments are conservative.
Work Center Capacity and Efficiency Factors (capacity and time efficiency of mrp.workcenter)Capacity load, process scheduling, overtime budgetProduction capacity is underestimated, and we dare not take orders we should accept.
Work Center Hourly RateManufacturing overhead allocation, process costsThe cost structure is wrong, but the total may happen to be correct.
Equipment Register (maintenance.equipment) and Maintenance PlanBreakdown downtime, spare parts consumption, depreciation allocationThe status collected from 120,000 points has no owning object.
Product Standard Cost (product.standard_price) and Costing MethodInventory value, gross profit, period-end carryforwardWorkshop costs have been reduced, but inventory value is still at the old price.

The easiest thing to overlook in this table is the fourth row. Xiaomi's plant deployed 120,000 collection points, and the equipment can perform self-inspection—automatically add oil when low, automatically pressurize when air pressure is insufficient, and automatically start up once debugging reaches OK. For these capabilities to enter cost accounting, the prerequisite is that every piece of equipment in the ERP has a card: how much it is worth, how much it depreciates per hour, how long its maintenance cycle is, and when its last failure occurred. If that card contains only a name and location, the status coming from the collection points is merely a monitoring screen and cannot enter the accounts.

A common pitfall: You only changed the labor hours on the routing and forgot to synchronize the work center's capacity and efficiency factors. The result is that costs are calculated correctly, but scheduling still runs at the old takt—the promise dates sales sees are later than what can actually be produced, so orders are lost for nothing; or worse, scheduling is compressed to the new takt, which cannot actually be achieved, and delivery is delayed again and again.

IV. The "minimum viable version" provided by Jing County may be more suitable for most enterprises

That Xiaomi factory invested more than RMB 2.5 billion, with 120,000 data collection points, more than 130 sets of AI visual inspection, and 94% coverage of smart logistics across the factory. Most companies cannot replicate this, nor do they need to. What is truly copyable is the "small, fast, light, and precise" approach used in Jing County, Hebei.

The figures published by Guangming Daily are plain: an annual fee of 20,000 yuan, 20% of the price of comparable products, an upgrade cycle of 5 to 15 days, and 152 enterprises already upgraded. For a specific company, after Hebei Xiboli Special Rubber launched the system in April this year, production scheduling changed from "salesperson to order tracker to planning and workshop" to a one-click direct connection to the production line, and the scheduling cycle was compressed from 48 hours to 2 hours; each rubber hose has a digital ID, and scanning the code shows which day it was produced, which worker, which equipment, and which batch of rubber, while quality traceability efficiency increased fivefold.

This setup isn't very technically sophisticated; what makes it valuable is that it chose the right granularity: batch + process + person + machine. By recording all four dimensions together, you establish the basis for cost, traceability, and quality. Conversely, if you record only output and not the machine, traceability breaks down on the shop floor; if you record only the machine and not the batch, material mixing cannot be detected.

A word of caution: for 5 to 15 days and an annual fee of 20,000, what you usually get is work reporting, dashboards, traceability, and a few management reports—not a complete system that can perform process costing, multi-company consolidation, and support quoting. It solves "visibility," not "accurate calculation." The price difference between these two is an order of magnitude; don't accept the former as the latter.

V. If You Only Do Three Things

1. Take a quotation from the last three months and trace back its labor-hour quota. Ask three questions: in what year was this figure set, who made the most recent modification, and how much do the workshop's current actual labor hours differ from it. If this equipment just deployed an AI agent this year and debugging time has dropped by two-thirds, while the quota has not changed, then your quotation is systematically too high—the small gap the customer sees when comparing prices is very likely this.

2. Make the loss reason in work reporting a required field. You don’t need to make it very detailed right from launch; start with four categories: equipment failure, waiting for materials, changeover, and quality rework. After a month, you’ll have a table you didn’t have before—where the lost time went. This table will tell you where to spend money better than any AI proposal.

Third, establish a time rule for quota review. For example: whenever the process parameters of any operation undergo substantive changes (new equipment, new molds, new parameter sets), the quota must be reviewed within 30 days; the reviewer signs, and any changes are recorded. A quota is not static data that is filled in once and done; it can expire. Processes are changing, equipment is being replaced, and AI is tuning parameters—the only thing that does not move is that field, which itself shows that the process is missing a link.


Go back to the injection molding factory from the beginning. Later they did a very small thing: they separated trial molding from the "injection molding" process and set it up as a separate step, requiring loss reasons to be filled in when reporting work hours; three months later, when standard costs were recalculated, the cost of the same product dropped by 6%, and the quotation was adjusted by one tier accordingly. The workshop did not install another sensor; the money was picked back up from the gaps in the process.

AI makes the granularity of the workshop smaller—this is what technology brings. Making the granularity of the ledger smaller along with it is what management must do. The former costs budget, the latter costs patience—and most of the time, what holds companies back is the latter.

Data sources in this article: CCTV's “Morning News” report broadcast on 2026-11-17 on digital-intelligence-enabled production line renewal, and the companion visit article from Wuhan Economic and Information Bureau (Xiaomi Wuhan Smart Home Appliance Factory: 120,000 collection points, more than 1.1 billion data records collected per day, qualified products after 5 trial mold runs, debugging time shortened by two-thirds, first-pass yield 99.98%, and energy consumption per unit down about 10%); Guangming Daily, 2026-11-17, page 09 (Jing County Industrial Brain; Gaoke Rubber & Plastics' online reporting for 13 processes and 4 quality inspections; Xiboli's production scheduling cycle reduced from 48 hours to 2 hours; “small, fast, light, and precise” annual fee of 20,000 yuan and a 5- to 15-day transformation cycle). Odoo models and fields are based on versions 17 to 19 as examples; version 20 may change them, so please check the official documentation before implementation.

Zhongdi Information · Enterprise Management Digitalization Observations · The enterprise cases in this article are cited from public reports and are not accompanied by third-party audits.

关于我们

​我们致力于帮助中小企业实现数字化转型,我们的团队由一群充满激情和创新思维的专业人士组成,他们具备丰富的行业经验和技术专长。

扫一扫获取顾问以及手册

Sign in to leave a comment
When an agent retires, who owns its memories?