JD Logistics shows where AI must own the whole job

JD AI Shopping, JD.com's standalone conversational app, puts the question plainly: can AI own a customer's goal rather than merely assist inside a screen? According to geekpark, Chen Yu's team built it as a native shopping experience based entirely on conversation, while also placing AI into the established shopping flow through search intent enhancement, recommendations, and a scenario prompt above the cart. A full pilot showed the limit: conversation fits complex product-selection consultations, not every shopping situation.

That distinction is the useful one. woshipm's author argues that the dividing line is not the presence of AI, but completion of a complete, high-value task; Zeng Ming frames an Agent as a system that can pursue a goal without step-by-step human direction. The harder work begins after the model responds: redesigning the workflow, linking decisions to execution, and assigning responsibility when the system cannot finish the job.

A copilot is not an outcome owner

A copilot improves a moment inside an existing product: it helps interpret a query, surfaces relevant information, or offers a prompt. An Agent has a different contract. According to woshipm's account of Zeng Ming's argument, it is defined by its ability to pursue a goal and finish the associated task without requiring step-by-step human direction. The unit of value is therefore not a better screen interaction, but completion of a complex, high-value job.

That distinction shifts accountability from assistance to an outcome.

JD.com's shopping experiments show why conversation does not automatically replace an established flow. Chen Yu's team launched the standalone JD AI Shopping app to build a shopping experience based entirely on conversation. After a full pilot, geekpark reports, the team concluded that conversational shopping works especially well for complex product-selection consultations, rather than every purchase. A shopper deciding among complicated options may need dialogue; a routine purchase may not.

JD.com instead placed AI within its existing shopping flow. Its search gained intent enhancement, recommendations gained time-, scenario-, and user-based information, and a scenario-prompt entry appeared above the shopping cart. These are useful assistant capabilities because they improve decisions at points where shoppers already act. But woshipm's author argues that the meaningful dividing line comes later: AI must complete a whole high-value task, not merely add intelligence to a product feature. Conversation can be an interface. Outcome ownership is the stronger claim.

The task boundary reaches into the warehouse

A warehouse task does not end when a model recognizes an item. At JD.com, more than 100,000 SKUs create a physical operation in which planning must reach the shelf, the robot, and the exception path. According to geekpark, Lin Che's team built the World Intent Model to separate intent generation from frequent execution. A large model sets the intent; an execution specialist carries out repeated actions. When an unexpected event interrupts the flow, a controller calls the large model back into the loop.

That structure makes the task boundary operational rather than screen-level.

Sensing is part of ownership because the system must know when its plan no longer fits the object in front of it. woshipm describes JD Logistics' Yilang robotic arm as using visual and tactile signals to choose a grasp and check that the item is secure. geekpark reports JD.com's claim that Canglang reached 99.9% grasping accuracy, covered more than 85% of products, and picked 80 items per hour. Those measures connect intelligence to a warehouse result, while the controller provides a route back to higher-level judgment when normal execution fails.

The same boundary appears beyond the warehouse. woshipm says UPS uses ORION operations-research algorithms for last-mile route optimization, while FedEx Network 2.0 weighs weather, traffic, shipment volume, and air and ground capacity together. DHL's Agentic AI is described as detecting operational changes, acting on decisions, and learning from what follows. JD Logistics' Super Brain 2.0 similarly coordinates planning from data across the logistics process. In each case, useful AI has to connect a plan to movement in the network, then revise that plan when conditions change.

From AI-assisted interfaces to operational execution across Chinese industry

Autonomy needs an escalation contract

An AI system should receive authority according to the cost of a wrong action, the reliability of its inputs, and the ability to reconstruct what happened. In physical operations, a useful answer is insufficient if it cannot be executed safely or reversed when conditions change. The system needs a defined point at which it stops acting alone and routes the case to a person or a higher-level decision process.

JD's World Intent Model offers a practical split, according to geekpark. A large model generates intent, while an action expert handles high-frequency execution. A controller watches for unexpected events and calls the large model again when the routine no longer fits. That arrangement avoids asking one model to make every decision at every moment. It also makes the controller an accountability boundary: normal actions can proceed, while exceptions receive more scrutiny.

That is the escalation contract.

Performance claims make the boundary concrete rather than philosophical. JD.com says its Canglang robot reaches 99.9% grasping accuracy, covers more than 85% of products, and picks 80 items per hour. Those figures describe a capable operating range, not permission to treat every item or incident as equivalent. A system should identify when it is within that range, record the decision, and expose an override path when it is not.

Human oversight also has to be operational, not ceremonial. JD.com's Nirvana Plan provides technical training for its existing 700,000 couriers and other blue-collar workers through 120 schools nationwide, geekpark reports. That matters because escalation lands with people who need enough context to intervene. woshipm describes DHL's Agentic AI as sensing changes, making and executing decisions, then learning from outcomes; the learning loop still needs explicit limits on when autonomous execution may continue.

Measure the value loop, not the model demo

An ROI test begins after the model demo. JD.com claims its Canglang robot reaches 99.9% grasping accuracy, covers more than 85% of products, and picks 80 items per hour. Those are meaningful operating measures, but they do not by themselves show the cost of connecting the robot to warehouse systems or repairing bad operational data.

A good metric must survive ordinary work.

woshipm reports that JD Logistics cut calculations that had taken two or three days to two hours, with some work moving toward 30 minutes. That gain matters only if redesigned workflows let planners act on the result, exception cases reach accountable people, and the system remains maintained as conditions change. The calculation is one component of the return; integration, training, and responsibility for the resulting outcome belong in the same calculation.

The training bill is visible even when it is not presented as a software cost. According to geekpark, JD.com's Nirvana Plan offers technical training to its existing 700,000 couriers and other blue-collar workers through 120 schools nationwide. That scale suggests deployment is a workforce change as much as a model deployment. A credible ROI review should therefore track the data work required before launch, the operating redesign required afterward, and who owns failures that cross from software into physical execution.

The warning is industrial deployment itself. 36kr attributes to Guo Yingzi research finding industrial AI projects generally fail at rates above 70%, while Wang Guolong argues that scale takes time when the measure is completion of a value loop. woshipm also reports about $300 million in savings from UPS's ORION system. Strong outcomes can exist, but short-term ARR alone cannot establish long-term potential, as Zeng Ming argues.

Choose AI by the outcome it can own, not the screen it can decorate

• A customer is making a complex, consultative purchase decision involving many product attributes. Use a conversational AI experience as the primary interface. JD.com's pilot found conversational shopping well suited to complex product-selection consultations; keep conventional search, recommendations, and shopping flows for other cases rather than forcing every transaction through chat.
• A warehouse task is repetitive and high-frequency, but exceptions are unavoidable. Split planning from execution. JD.com's World Intent Model generates intent with a large model, delegates routine execution to an action expert, and calls the large model again when an unexpected event occurs. This is a stronger operational pattern than asking one model to control every movement continuously.
• A logistics or industrial team wants AI to improve a full operational process rather than automate an isolated step. Start with a measurable business outcome, operating constraints, and domain knowledge, then connect the system to the data and execution layers needed to act on its plan. JD Logistics describes Super Brain 2.0 as integrating full-process data for coordinated planning, while Li Yaman argues against predefining every step before involving AI.
• An organization is considering an industrial AI deployment but cannot yet show how it completes a value loop. Treat it as an implementation and workflow-redesign project, not a model procurement exercise. One conference participant said industrial AI projects generally have a failure rate above 70%, and another said scaled deployment still needs time when judged by whether it completes a value loop.
• A team is evaluating whether a chatbot feature qualifies as an agentic product. Require the system to complete a task around a goal without step-by-step human direction, while defining when humans handle exceptions and accountability. Zeng Ming's distinction between a chatbot and an Agent is goal completion without continuous human instruction; the opportunity, in this view, is complex, high-value tasks rather than isolated simple steps.

Build a smaller feedback loop first

A smaller firm can begin with a bounded task that still contains a complete business outcome. The point is not to automate a single click, but to define what acceptable completion looks like, what constraints apply, and what knowledge the task requires. Li Yaman argues in woshipm that product managers should set those conditions before asking AI to find a solution. That reverses the familiar impulse to split work into predetermined steps and then attach a tool to each one.

Start where the acceptance test can be checked against completed work.

The initial loop should connect the AI's proposed or completed work back to the result it produced. That feedback supplies the task-specific learning Zeng Ming identifies in woshipm, alongside context management, long-term memory, and interaction methods. Human owners should retain responsibility where the task calls for judgment, especially when the result fails its acceptance test or falls outside the stated constraints. Authority can widen only after the team can see the loop close reliably.

A bounded starting point does not mean choosing a trivial step. Zeng Ming argues that AI's larger opportunity is in complete, complex, high-value tasks rather than isolated simple actions. Chen Yu's team reached a related limit after a full pilot, according to geekpark: conversational shopping fit complex product-selection consultations, but it did not fit every shopping situation. The useful boundary is therefore specific, not universal. A firm can assign AI a defined outcome, preserve human judgment at the edges, and redesign collaboration around the task. As woshipm's author argues, that redesign should come before buying tools.

Start with a complete, high-value user task, then re-decompose the work and the collaboration around it before buying another AI tool. Ask where context is lost, who owns the task, and what knowledge must persist between attempts. According to woshipm, an Agent company's technical accumulation must cover context management, long-term memory, interaction methods, and task-specific learning mechanisms. Judge progress by sustained task completion, not short-term ARR, which Zeng Ming argues cannot establish long-term market potential.

Watch for work that can move from role-centered ownership to task-centered ownership. Treat education, health, enterprise growth, and organizational brains as areas to investigate, not proven tracks.

For readers outside China

• Availability: The source material discusses JD.com, JD Logistics, JoyInside, and industrial deployments in China, but does not say whether their AI shopping, warehouse, or logistics systems are available outside China. It also does not describe international access, supported languages, APIs, or enterprise purchasing arrangements.
• Pricing: Pricing is not disclosed in sources.
• Closest Western equivalents: UPS ORION, which uses operations-research algorithms to optimize last-mile delivery routes; FedEx Network 2.0, which evaluates weather, traffic, shipment volume, and air and ground capacity in network planning; DHL's described Agentic AI system, which senses operational changes, makes and executes decisions, and learns from outcomes
• Data residency: Data residency, cross-border data transfer, cloud-hosting location, retention, and compliance arrangements are not disclosed in sources. For overseas organizations, those questions would need to be resolved directly with the relevant provider before connecting operational data or execution systems.

Sources

• 36kr 圆桌：新能源：后时代的深水淘金 | 36氪 2026产业未来大会 https://36kr.com/p/3984491452824578
• geekpark 京东押注物理 AI，冲在前面的是一群 95 后 https://geekpark.net/news/370519
• woshipm 阿里前总参谋长曾鸣万字访谈总结，关于 AI 未来的 42 个判断 https://woshipm.com/ai/6461475.html
• woshipm 京东物流李亚曼：当AI进入物流全链路，产品经理要重新定义自己的价值 https://woshipm.com/ai/6455237.html
• 36kr 圆桌：探路-产业分工的新图谱，看到新未来--国家力量如何转化为产业价值 | 36氪 2026产业未来大会 https://36kr.com/p/3984472921783045
• 36kr 圆桌：先进制造：给世界工厂装一颗AI大脑 | 36氪 2026产业未来大会 https://36kr.com/p/3986931112836096