Supply Chain Action Points
Read this first — the conclusion, and the moves to make:
- By 30 September 2026, put your top five lanes on one page with parcels per day and the share of volume held by the largest single provider; any lane above 60% concentration goes to a second quote in October.
- Consolidate annual volume from six providers to no more than three before the next tender, and take the density gain as a rate commitment rather than a promise of better service.
- Request lane-level unit costs, not network averages, for every quote above 50,000 parcels a month; reject any figure that arrives without its volume denominator attached.
- Before 31 December 2026, test your carton profile against the dimension and weight range the sorting cameras are specified for, and cap out-of-range units at 10% by count.
- Recalculate your own break-even each quarter using your actual parcels per day; where volume sits below 87,000 a day, treat automation-led savings as a supplier claim rather than a budgeting assumption.
Chinese operators are scaling AI beyond demos: YTO's AI-vision sorting centers now run million-order days with one-third the labour, SF's predictive fleet maintenance cut breakdown downtime 40%, and SF has deployed over 3,200 driverless vehicles across 150+ cities. Single-parcel transport cost fell to 0.36 yuan and centre handling to 0.26 yuan, showing AI is now embedded in equipment and workflows.
The Analysis
Three numbers belong together: one third of the labour, forty per cent less breakdown downtime, and a transport cost of 0.36 yuan per parcel with centre handling at 0.26. The first two are rates of change. The third is a quotient, and a quotient is the only one of the three that tells you what the system costs to run.
Chinese operators have moved AI out of the demonstration hall and into the equipment. Vision sorting at YTO runs million-order days on a third of the people. Predictive maintenance at SF cuts fleet breakdown downtime by 40%, and SF has put more than 3,200 driverless vehicles on the road across 150-plus cities. None of that is a pilot any more. All of it is now a cost structure.
A cost structure is a machine with a break-even buried inside it, and the break-even is the part nobody quotes.
The three deployments get reported as one trend and they are three different machines. Vision sorting attacks labour hours per thousand parcels, and the constraint it relieves is a headcount that scales with volume. Predictive maintenance attacks the variance of breakdowns rather than the mean, so the asset it protects is not the truck but the promise made on top of the truck. Driverless last-mile attacks the marginal cost of a stop, the smallest unit in the chain and the one that resisted automation longest, because a stop involves a doorway, a lift and a signature.
Read them as a single trend and you arrive at a single bad conclusion, which is that AI reduces cost. Each one reduces a specific term in an equation, and only one of the three has a number attached that can be compared across operators.
Add the two published unit costs, 0.36 for line-haul and 0.26 for centre handling, and you get 0.62 yuan of logistics cost per parcel, on the assumption that the two are additive and do not overlap. That is my assumption, not something the source establishes. Sixty-two fen is roughly what a manual network spends on the same two steps before it spends anything at all on the last mile.
The interesting property of that number is not its size. It is that it has a denominator. Cost per parcel equals fixed cost divided by volume, plus variable cost per parcel. AI moves money from the second term to the first. Vision sorting swaps pickers for cameras, and cameras are capital. Predictive maintenance swaps roadside repair for sensors and models, and those are capital. Driverless vehicles swap drivers for vehicles, and a vehicle depreciates on a Sunday.
Which makes the break-even the number that decides whether any of this is a strategy or a decoration. Suppose a network carries 12 million yuan a year of depreciation for its vision and maintenance stack, and the saving against a manual operation is 0.38 yuan per parcel. Both are my assumptions, picked to make the arithmetic readable. Break-even is 31.6 million parcels a year, about 87,000 a day, every day, not on the peak day. An operator doing million-order days clears that line by an order of magnitude. An operator doing 40,000 a day does not clear it at all, and for that operator the same technology is a cost increase wearing the costume of an efficiency programme.
This is why the two unit costs cannot be read as benchmarks. They are the output of a volume that most readers do not have. A forwarder or a mid-sized exporter looking at 0.62 yuan and asking an equipment vendor to match it is asking the wrong party a question about the wrong variable. The vendor can sell the cameras. Nobody can sell the density.
For importers and exporters the practical consequence runs through consolidation rather than technology. A shipper whose annual volume is split across six forwarders is, in effect, renting six sub-scale networks. Consolidating onto two does not buy AI. It buys the denominator that makes somebody else's AI cheaper for you, because the volume you hand over is what lets an operator amortise the capital and then quote a rate that reflects it.
You will feel this through the rate card before you feel it through the invoice. Automated cost structures propagate as price competition on the lanes where density is highest, so the first visible effect is a widening gap between the busy lane and the thin one, not a uniform reduction. Expect that gap to widen across the next two to four quarters on Chinese domestic lanes, and expect it not to appear at all where no operator has density. That asymmetry is the signal: when one lane gets 15% cheaper and the lane beside it does not move, the difference is not technology, it is volume.
The part the coverage skips is inside the 3,200 vehicles. Spread across more than 150 cities, that is about 21 vehicles per city, and 21 is not a fleet. It is a learning cell. The economics of a driverless route turn on stops per vehicle per day, and at 21 vehicles a city the operator is not yet running a transport network. It is running a data collection programme with a delivery function attached. That is a rational thing to do at this stage, and it means the published unit costs describe a network average: dense districts where the model works, thin ones where it does not, blended into one number that is quoted to you as if it were universal.
Ask for the lane-level figure, not the network figure. An operator quoting one cost per parcel across a national network is quoting a mean over a distribution, and the distribution is where your money sits. If your lane is in the thin tail, the 0.36 you are quoted is being subsidised by a lane you are not on, and the subsidy ends the quarter the operator is asked to show a profit by route.
There is a related reading of the 40% that changes how you should use it. A 40% cut in breakdown downtime is a reduction in variance, not in the mean repair bill. Variance is what you hold assets against: the buffer fleet, the standby driver pool, the safety stock that exists because a truck might not arrive. Cut the variance and you can shrink all three. The money is in the fleet you no longer need, and almost every buyer negotiates the repair bill instead.
Two conditions flip the arithmetic. If volume falls below the break-even, fixed cost per parcel rises faster than any labour saving can offset, and the technology becomes an amplifier of a bad quarter rather than a cushion against one. If the parcel mix moves toward oversized, fragile or highly variable units, vision accuracy drops, exception handling returns to people, and one third of the labour becomes a fraction of a fraction. Both are measurable before you sign anything: your own volume against the operator's break-even, and your own carton profile against the specification the cameras were trained on.
Standing still is not the neutral option. A competitor with density is repricing the lane you share, and the gap compounds, because lower cost funds more volume which funds lower cost. The choice was never whether to adopt AI. It is whether to be the operator with the density or the shipper who buys from one.
Start with a number you already own: parcels per day on your heaviest lane, and the share of your volume sitting with a single provider. Below 40%, you are not buying automation. You are renting somebody else's scale at retail.
↑ Back to the key points