AI Penetration in China's Logistics Passes 37%, With Transport at 78% and Warehousing 47%

The Analysis

The notice says this. The China Federation of Logistics and Purchasing has published its 2025-2026 supply chain report, and the number that travelled out of it is 37. AI penetration across logistics and supply chain, above 37 percent. Transport is the most exposed link at 78 percent. Safety and compliance sits at 52.73. Smart warehousing at 47.27. The report uses Double 11 as its stress test: JD Logistics upgraded Super Brain 3.0 to forecast regional order volumes and reallocate national capacity nationwide, with more than 90.2 percent of its recommended routes adopted, and its Yilang arm tracks parcels at 99.9 percent accuracy. YTO's digital twin covers every consolidation centre in its network, and its machine vision covers 24 scenarios.

I install these systems for a living, and I have learned to read a penetration headline the way I read a vendor benchmark slide: politely, and with the arithmetic open in another window. The number is almost never wrong. It is almost always answering a different question than the one the headline asks. So before anyone in your company drops 37 percent into a board pack, we should establish what 37 percent is a share of. Firms or cargo. Sites or decisions. Something installed or something used.

Start from the only framing I trust: it is a queue. Everything here is a queue. A parcel is a queue, a dock door is a queue, a dispatcher's screen is a queue. A penetration rate, at bottom, is a statement about how much of a queue has a machine standing in it and how much still has a person. That boundary is where the money sits, and it is exactly the boundary the headline deletes.

I am not going to write that digitalisation matters or that AI is reshaping logistics. Those sentences cost nothing and buy nothing. What I want to do is take three numbers, 37 and 78 and 47.27, and show you something about them that almost nobody has put in print, which happens to be the most useful thing in the whole report for anyone signing a 2027 contract.

Begin with the two decimals, because two decimals are a confession. 52.73 and 47.27. Add them. 100.00. Exactly. Two survey items landing on a perfect complement is not a coincidence you can wave past, because the odds of two independently measured percentages summing to a round hundred are small enough to ignore. What you are looking at is one binary split reported as two findings. Safety and compliance at 52.73 and smart warehousing at 47.27 are the same pie cut in two, and the cut was made over a headcount of firms, not a tonnage of cargo.

Now do the division that the report does not do for you. If the base were 55 respondents, the split is 29 and 26. If the base were 1,100, it is 580 and 520. The arithmetic cannot tell me which, and I am not going to pretend it can. What it does tell me is the shape of the thing: an either-or question, answered by a fixed panel of enterprises, tallied and turned into two percentages with two decimal places each. That is a customer survey. It is a perfectly respectable customer survey. It is not a measurement of how much cargo moves through a machine.

Which brings us to the headline. Above 37 percent. Of what. If the panel is enterprises, then 37 is the share of surveyed firms with at least one live AI application somewhere in their logistics and supply chain operations. Read that sentence again with the emphasis on the words somewhere and one. A firm with a single demand-forecasting model in one business unit clears the bar. So does a firm with a chatbot on a customer service line, if the survey counts it. And 78 percent in transport does not mean 78 percent of loads are machine-dispatched. It means 78 percent of the respondents have something running on the transport side. The distance between those two statements is the whole distance between a market report and an operating decision, and nobody in the coverage closed it.

Here is the part I have not seen anyone publish, and I have looked. Multiply the two link numbers. 0.78 times 0.4727 is 0.3687. Thirty-six point nine percent. The headline says above 37. That is inside a rounding step of the product of the transport rate and the warehousing rate. I cannot prove the report built its headline that way. Maybe it is a genuine coincidence, and coincidences happen in reports assembled from many survey items. But whether it was built or merely landed there, the number behaves like a series system, and that is how you should read it. An end-to-end promise is not an average of its links. It is a chain, and a chain holds only where every link holds. If transport is instrumented 78 percent of the time and warehousing 47.27 percent of the time, then both ends are instrumented about 37 percent of the time, and the chain is only as good as the weaker end.

That reframing has a price tag attached, and this is the part I would put in front of a CFO. In a series system, one point of improvement at the weak link buys you more than a point at the strong one. Push warehousing from 47.27 to 60 and the product goes from 0.369 to 0.468, a gain of 9.9 points for 12.7 points of input. Push transport from 78 to 90 and the product goes from 0.369 to 0.425, a gain of 5.6 points for 12 points of input. Per point of input, warehousing returns 0.78 of a point of chain coverage and transport returns 0.47. Warehouse money buys about 1.65 times the chain coverage that transport money buys. Every vendor you meet this quarter will sell you the transport layer, because that is where the data is clean and the demo is pretty. The arithmetic says the cheaper win is the ugly one, at the dock door, where the ground truth has to be instrumented by hand.

Which raises the question the report never asks: why is transport at 78 and warehousing at 47.27 in the first place. My answer is that the ranking is not a ranking of value or effort. It is a ranking of label latency. Transport generates its own ground truth for free. Every waybill carries an address, a scan, a gate-out, an arrival, and the gap between the promise and the event closes in hours. A model can learn from that on a Tuesday and be better on a Wednesday. Warehousing ground truth is a physical act.

What a picker actually did, in what sequence, with what reach, is not written anywhere until someone instruments the rack, the cart or the wrist. Compliance sits in between at 52.73 because documents are text and reading text is the one thing this generation of models does cheaply. So the order 78, then 52.73, then 47.27 is a map of where the labels are free, not a map of where the value is. Read it the other way round and you will spend your budget in the wrong building.

Now take the two operational numbers, because they are the only ones in the story that come with a denominator attached. Ninety point two percent of recommended routes adopted. That is presented as an accuracy trophy. Read it as a rejection rate and it becomes a completely different object: 9.8 percent of the system's recommendations were overridden by a human. Ask what an override costs. Assume four minutes each, which is generous, because an override is not a click. Somebody has to read the recommendation, understand why the machine chose it, decide it is wrong, and type the reason.

Four minutes is the floor of that transaction. In a thousand recommendations you spend 98 overrides at four minutes, which is 392 minutes, which is six and a half hours. That is most of a shift of a skilled dispatcher's attention, spent not on the 900 cases the machine got right but on the 98 it got wrong, and spent on them in the most expensive possible way, one at a time, with a human in the loop who is also the person who has to answer for the load.

That six and a half hours is the integration cost nobody puts in the business case. The licence is a line item. The model training is a line item. The override desk is a headcount that appears in a different cost centre, under a different manager, and so it never shows up in the ROI calculation, which is why the ROI calculation is always optimistic. I have watched three rollouts where the steering committee approved on the accuracy number and the operations director resigned over the override load. Same project. Two different numbers. The gap between them was a decimal place and an org chart.

Take the other operational number. Ninety-nine point nine percent tracking accuracy, on the Yilang arm. That is a good number and I would not sneer at it. Do the multiplication anyway, because a rate without a volume is a mood. For every ten million parcel scans in a day, 99.9 percent leaves ten thousand parcels that the system has read wrong. Ten thousand exceptions in a day, arriving at a customer service desk that is staffed for the average, not for the tail. The tail is where the refunds live, where the negative reviews live, and where the one customer who matters this quarter lives. The honest way to run a 99.9 percent system is to staff the exception queue with the assumption that the exceptions scale with volume, because they do, and to accept that the last nine is a cost centre rather than a defect to be engineered away. Nobody sells it that way. It is still true.

The digital twin claim deserves the same knife. Every consolidation centre in the network, covered. Coverage is a count of sites. Fidelity is a different axis and it is the one that decides whether you can use the thing to run a shift. A twin that mirrors a hub's layout but refreshes on a batch every fifteen minutes is a map, not a twin, and a map cannot resequence a sort at 21:40 when a feeder is late. The refresh interval is the whole product and it is not in the press release, because it is the least impressive number in the room and the only one an operator cares about. When someone shows you a twin, ask two questions before you look at the screen. What is the update interval, and what decision does the twin change that the WMS could not already change. If the answer to the second is none, you have bought a screensaver with a floor plan.

Scenario counts deserve the same treatment, because 24 is a number that grows by procurement rather than by science. Adding a scenario usually means adding a camera, a light rig and a labelled training set, so the marginal cost per scenario falls as the count rises, and on a slide that reads as accelerating capability. The value does not follow that curve. The first five scenarios, dimensioning, damage, label reading, carry most of the volume. The nineteenth through twenty-fourth carry the tail, and the tail is expensive to instrument and cheap to get wrong. A vendor quoting the count is telling you about the breadth of the catalogue, not the depth of the accuracy on the handful of cases that make up your month.

So who actually carries this, and how much. Sort your exposure by where the model sits relative to the decision. In the first group, the importer whose forwarder sends an ETA the importer never used to get. The prediction got better. The decision did not move an inch, because the safety stock rule is set annually in a spreadsheet and the ETA is read by a planner who already trusts the vessel schedule more than the API. Value created, value captured by nobody. In the second group, the importer with its own transport management system and a planning cycle measured in hours. Here the improvement is real, but only if the prediction lands inside the decision window.

An ETA that goes from plus or minus thirty-six hours to plus or minus twelve is a genuine improvement in a system that replans daily. In a system that replans weekly, it is a more precise number attached to the same decision, which is to say it is decoration. In the third group, the importer who now has to answer a customer's where-is-my-box with a number a model produced. That is the fastest-moving group and the least discussed, because the benefit is a nicer answer and the liability is a commitment. The model did not take the risk. You did.

Which gives a cheap diagnostic you can run this week without buying anything. Pull the override records for the last thirty days and count the rows where the human changed the machine's answer and left the reason field empty. That ratio, empty reasons over total overrides, is the best single measure of whether the system is being used or merely tolerated. A low ratio means the operators are arguing with the model in writing, which is what you want, because a written argument is a training signal and an audit trail at the same time. A high ratio means the output is being clicked past, and every accuracy figure in the report is then describing a system nobody reads.

Then there is timing, and timing is where most of these reports quietly mislead. A penetration rate is a stock. The effect on your P&L is a flow, and the flow only starts when a contract renews. The Double 11 readout will land in December. Your 2027 service-level agreement gets negotiated between January and March. That six-week window between the peak readout and the tender is the only moment in the year when the numbers in this report can change a clause, and if nobody on your side has read the override log by then, the clause will say what it said last year. The carriers and the platforms know this window better than you do. They will arrive at the tender with a peak-season narrative and a service-level number, and the narrative will have been built in the same six weeks.

Here is the counterfactual that keeps me honest. If penetration were zero, would the peak have failed. No. Peaks were handled before any of this existed, with paper, phones and a dispatcher who knew every driver by name. So AI is not the load-bearing wall of the network. It is the shock absorber. That matters, because shock absorbers fail in a specific way: they fail at the top of the stroke, under exactly the conditions you bought them for, and they fail quietly, by bottoming out rather than by breaking. Which means the risk is not that the model is absent. It is that the model is present, confident and wrong. A planner told the box will arrive at 14:20 will not check. A planner told there is no estimate will pick up the phone. The confident wrong answer is strictly worse than the honest gap, and the report, which measures presence, cannot see the difference between the two.

Derek Xu makes a related point that I half agree with, and the half I disagree with is worth writing down. His view is that on the customs side the binding constraint was never prediction, it was the exception: the container sitting on a hold at three in the morning with no reason code, where no model in the world helps because the missing thing is a document, not a forecast. He is right about where the pain is. He is wrong to treat it as a different problem. The exception desk is also a queue, it has an arrival rate and a service time and a staffing level, and a model that improves the arrival forecast by forty percent while the service time stays flat has simply moved the pile from the front of the desk to the back of it. Prediction without a matching investment in the desk's throughput is a way of generating more accurate complaints.

So the practical instruction is not to buy more AI. It is to measure the chain the way a series system is measured, and to put the next dollar at the weak link. Assume, purely for arithmetic, that your transport leg is instrumented at the report's 78 percent and your warehouse leg at 47.27. Your promise holds about 37 percent of the time end to end. Move the warehouse leg to 60 and you buy 9.9 points of chain coverage for 12.7 points of input. Move transport to 90 and you buy 5.6 for 12. Those are the two deals on the table this quarter, and only one of them is being pitched to you.

↑ Back to the key points

— By Ravi Chandran