your food truck order has a retail cousin in alibaba's qwen office 2.0 release. it puts the agent inside dingtalk group chats, takes tasks from members, and starts on its own when a rule fires, like a store clerk posting about a stock-out. the case they gave was intime department stores, where five people used to copy submissions from more than 60 store chats into a spreadsheet by hand. which app does the tai'an factory actually run its orders through? the release only mentions dingtalk groups.
Yuzu — good catch, that's the retail cousin of the same problem. Qwen Office 2.0 (Nov 9) works because the work is already inside DingTalk groups — the agent lives where the conversation is and fires on rules. The Tai'an factory is messier: the order lives across WeChat, email, Excel and a salesperson's head, no single app to live inside. That's why my call is a harness that reads across all of them rather than an agent inside one of them. The Intime case you mention — five people copying submissions out of 60+ chats — is the same shape: the cost isn't the copying, it's that no one place holds the order.
casic's industrial cloud chief scientist told a huxiu closed session that some defense plants' scheduling software barely gets used. how does the harness pick up the part that's only in the salesperson's head?
that's the exact failure the harness is built for. Scheduling software sits unused because it asks the plant to move its work into the system; the work stays in WeChat, Excel, and the salesperson's head instead. The harness doesn't pretend to read a mind — it drafts the order from every trace that does exist (chats, emails, drawings, past quotes), then puts the draft in front of the salesperson to correct before anything is acted on. The part that's only in their head gets picked up at the approval step, not by extraction. If there's no trace and no one to ask, it stops and asks rather than guessing — that's the pre-check. The unused system is what happens when software skips that step and assumes the order is already written down somewhere.
great analysis, breaking down the application layer- is there an interface close to shop floor that allocates, assigns tasks like creating jira ticket or completing specific work-order?
Progress build videos are critical for cross functional alignment especially during NPI builds - almost comical how CapCut caps out for this massive time sink task right now.
Ishita — good question, and it's the right one to ask. In the picture I sketched there isn't a Jira-like screen to babysit: the skill itself does the allocating — it reads the order, drafts or updates the work order, and a person signs off before anything reaches the floor. The interface is the approval step, not another dashboard. The queue is the order itself.
And you're right about the videos — NPI builds are where that time sink is worst, and CapCut capping out is exactly the symptom. That's why the video skill matters: footage filed by order number, cut and captioned by the agent, a person looks before it goes out. Same logic as the work orders.
The scattered order (WeChat, email, Excel, a salesperson's head) is familiar from Japanese plants too, except there the head usually belongs to a veteran team leader. On your fourth risk, who builds it: in the plants I walk, the person who spots a problem first is the floor leader, and the call is often made the next morning rather than in the moment. I would be curious whether the system can show that person the whole order history in a form they can argue with, before it acts. That is where I think trust would start.
Kei — thank you, that's the sharpest version of the question. The Japanese parallel is exactly right: in those plants the "head" is usually a veteran team leader, and when he retires the order goes with him. Same shape as Tai'an, different address.
On the fourth risk, I think your observation is the answer. If the person who spots the problem first is the floor leader, and the call is made the next morning rather than in the moment, then the harness's job is to have the whole order history ready by that morning — every change, every drawing revision, every cost — in a form he can argue with, before anything is acted on. The approval step is where trust starts. If he can't challenge it, it's just another system telling him what happened.
Thanks, Yuanwei. One more reason I like the morning version: when the team argues with the same picture every day, the younger leaders slowly learn to see what the veteran sees. I think that is how the order stops leaving with the person who retires.
The penetration of AI will reshape organizations and ultimately turn them into observable rules and regulations. I believe the process of teaching and learning will be disrupted, which is exactly the current situation in China and Silicon Valley: the living space for junior engineers has been massively squeezed. We also need to find a better way to get people involved in more meaningful activities, but I don't have the answer yet.
I don't have the full answer either, and I agree the squeeze on junior people is the real risk. At a car plant I know well, the floor dashboard deliberately did not show the reason for each stop. The veteran leaders said what they used to know by feel was now on screen, so younger leaders could see it too, while finding the reason stayed the team's job. So my working answer is to let AI hand people the picture and leave the "why" for them to argue out, because that argument is where juniors keep learning. Do you see anything like that in the plants you know in China?
your food truck order has a retail cousin in alibaba's qwen office 2.0 release. it puts the agent inside dingtalk group chats, takes tasks from members, and starts on its own when a rule fires, like a store clerk posting about a stock-out. the case they gave was intime department stores, where five people used to copy submissions from more than 60 store chats into a spreadsheet by hand. which app does the tai'an factory actually run its orders through? the release only mentions dingtalk groups.
Yuzu — good catch, that's the retail cousin of the same problem. Qwen Office 2.0 (Nov 9) works because the work is already inside DingTalk groups — the agent lives where the conversation is and fires on rules. The Tai'an factory is messier: the order lives across WeChat, email, Excel and a salesperson's head, no single app to live inside. That's why my call is a harness that reads across all of them rather than an agent inside one of them. The Intime case you mention — five people copying submissions out of 60+ chats — is the same shape: the cost isn't the copying, it's that no one place holds the order.
casic's industrial cloud chief scientist told a huxiu closed session that some defense plants' scheduling software barely gets used. how does the harness pick up the part that's only in the salesperson's head?
that's the exact failure the harness is built for. Scheduling software sits unused because it asks the plant to move its work into the system; the work stays in WeChat, Excel, and the salesperson's head instead. The harness doesn't pretend to read a mind — it drafts the order from every trace that does exist (chats, emails, drawings, past quotes), then puts the draft in front of the salesperson to correct before anything is acted on. The part that's only in their head gets picked up at the approval step, not by extraction. If there's no trace and no one to ask, it stops and asks rather than guessing — that's the pre-check. The unused system is what happens when software skips that step and assumes the order is already written down somewhere.
great analysis, breaking down the application layer- is there an interface close to shop floor that allocates, assigns tasks like creating jira ticket or completing specific work-order?
Progress build videos are critical for cross functional alignment especially during NPI builds - almost comical how CapCut caps out for this massive time sink task right now.
Ishita — good question, and it's the right one to ask. In the picture I sketched there isn't a Jira-like screen to babysit: the skill itself does the allocating — it reads the order, drafts or updates the work order, and a person signs off before anything reaches the floor. The interface is the approval step, not another dashboard. The queue is the order itself.
And you're right about the videos — NPI builds are where that time sink is worst, and CapCut capping out is exactly the symptom. That's why the video skill matters: footage filed by order number, cut and captioned by the agent, a person looks before it goes out. Same logic as the work orders.
The scattered order (WeChat, email, Excel, a salesperson's head) is familiar from Japanese plants too, except there the head usually belongs to a veteran team leader. On your fourth risk, who builds it: in the plants I walk, the person who spots a problem first is the floor leader, and the call is often made the next morning rather than in the moment. I would be curious whether the system can show that person the whole order history in a form they can argue with, before it acts. That is where I think trust would start.
Kei — thank you, that's the sharpest version of the question. The Japanese parallel is exactly right: in those plants the "head" is usually a veteran team leader, and when he retires the order goes with him. Same shape as Tai'an, different address.
On the fourth risk, I think your observation is the answer. If the person who spots the problem first is the floor leader, and the call is made the next morning rather than in the moment, then the harness's job is to have the whole order history ready by that morning — every change, every drawing revision, every cost — in a form he can argue with, before anything is acted on. The approval step is where trust starts. If he can't challenge it, it's just another system telling him what happened.
Thanks, Yuanwei. One more reason I like the morning version: when the team argues with the same picture every day, the younger leaders slowly learn to see what the veteran sees. I think that is how the order stops leaving with the person who retires.
The penetration of AI will reshape organizations and ultimately turn them into observable rules and regulations. I believe the process of teaching and learning will be disrupted, which is exactly the current situation in China and Silicon Valley: the living space for junior engineers has been massively squeezed. We also need to find a better way to get people involved in more meaningful activities, but I don't have the answer yet.
I don't have the full answer either, and I agree the squeeze on junior people is the real risk. At a car plant I know well, the floor dashboard deliberately did not show the reason for each stop. The veteran leaders said what they used to know by feel was now on screen, so younger leaders could see it too, while finding the reason stayed the team's job. So my working answer is to let AI hand people the picture and leave the "why" for them to argue out, because that argument is where juniors keep learning. Do you see anything like that in the plants you know in China?