Supply Chain Action Points
Read this first — the conclusion, and the moves to make:
- Log five timestamps per load for four weeks, from order release to gate-out, and rank every dock by ninetieth percentile minus median wait; anything above 60 minutes goes on the fix list before any dispatch tender is issued.
- Cut ninetieth percentile dock wait by 40% within one quarter using appointment windows and a published priority rule, measuring on the same lane, same month and same volume as the baseline.
- Put the dispatch contract's acceptance on loaded-kilometre ratio and on-time pickup percentage, for example empty running down three points and on-time pickup at 95% within two quarters, with no renewal if either misses.
- Hold cross-border dispatch spending until document first-pass rate is at or above 95% and border dwell variance fits inside one shift; below those numbers the inland gain is arithmetically invisible to the customer.
- Ring-fence any low-altitude or green pilot with a fixed budget, a six-month window and a written kill criterion, and keep it out of the capacity plan until it clears both.
The 2026 Modern Commercial Logistics High-Quality Development Conference opened 10 July in Tongcheng, Hubei, releasing the '2026 China Road Transport Capacity Development Report.' Sessions focused on AI large-model applications in dispatch and route optimization, cross-border transport, low-altitude-economy commercialization and green transport under the 15th Five-Year Plan blueprint.
The Analysis
A conference is a press release with a stage and better catering. That is not a reason to skip it, but it is a reason to read the document instead of the agenda.
The 2026 Modern Commercial Logistics High-Quality Development Conference opened on 10 July in Tongcheng, Hubei, alongside the sixth capacity ecosystem conference, and released the 2026 China Road Transport Capacity Development Report. The sessions went where these events usually go: large models in dispatch and route optimisation, cross-border transport, commercialisation of the low-altitude economy, and green transport under the 15th Five-Year Plan.
Here is the part worth modelling. Dispatch is an assignment problem on a graph, and the weights on that graph are not travel times. On most domestic road lanes the largest term in a truck's cycle is not the drive. It is the wait.
Take one tractor doing a 500 kilometre round trip. Eleven hours of driving is a reasonable assumption at legal speeds with a break; four hours of loading, unloading, gate and dock waiting is another. Both figures are mine, not the report's. Dispatch optimisation touches the eleven, and might recover 8% of it, which is 53 minutes. An appointment window that actually holds touches the four, and cutting it in half is two hours. The second change is worth more than twice the first and it needs no model at all. It needs a calendar and a penalty clause.
That ratio is why I read AI dispatch news with one question: what is the service time at the nodes? A model that assigns trucks to loads is a scheduler sitting in front of a queue. Schedulers are good at sequencing and helpless against physics. If the dock takes three hours because the receiver has one forklift and no appointment book, no assignment rule shortens it. You get a better-ordered queue, and the same day ends at the same hour.
Capacity reports also have a population problem. Registered capacity is not available capacity. A truck that is licensed, insured and parked is in the denominator and out of the numerator, and seasonal utilisation swings sit inside an annual average where nobody can see them. Ask for the monthly series and the loaded-to-empty ratio by region before you quote the headline anywhere near a budget.
Behind all of it sits the variable that actually sets your freight cost: the ratio of loaded to empty kilometres, and behind that, the variance of dwell time. A fleet with a 12% empty-run ratio and a stable two-hour dock turn is easier to buy from than a fleet at 8% empty running whose dock turn swings between one hour and six, because the second one cannot be promised to a customer. When you sign a service level you are not buying the average. You are buying the tail.
Who gains depends on the size of the data rather than the size of the company. A shipper moving a handful of loads a week gets price transparency from a digital platform and very little optimisation, because the platform is applying somebody else's history to your lane. A shipper running 20 to 100 trucks on dedicated contract gets real value, since four weeks of clean records is enough to learn a lane's behaviour. A cross-border shipper gets almost nothing from dispatch until border dwell and document accuracy are fixed, because a two-hour inland gain disappears into a two-day wait at the port.
Low-altitude commercialisation and green transport belong in the same sentence for a practical reason: both are options with a strike price, not capacity you can book this quarter. Treating either as a planning input today is how a transport budget acquires a line item that never pays back. The near-term use of both is a bounded pilot with a written kill criterion, kept outside the capacity plan until it clears.
You will feel this on a contract cycle, not a news cycle. Tender, integration, data cleanup and the first measurable quarter add up to something like six months, and the first invoice you receive is for cleaning up appointment data that nobody owned. The capacity gain shows up in the second quarter of clean data, which is roughly when a shipper discovers the savings were in the dock the whole time and were never going to come out of the routing engine.
Put a dispatch model on top of a dock with a three-hour ninetieth-percentile wait and watch what happens. The model does not remove the variance; it learns it. Now you have a confident ETA built from historical waiting, a promise made to a customer on that ETA, and a miss rate that gets worse before it gets better, because the algorithm has absorbed your worst days and treats them as normal. Digitising a queue is not the same as shortening it, and the first is much easier to buy.
The trap is buying the demo. Demos run on clean lanes with cooperative drivers and a dispatcher who knows which receiver answers the phone before ten. Your lanes have a receiver who takes lunch at the wrong hour. Instrument five timestamps per load for four weeks, order release, gate-in, dock-start, dock-end and gate-out, then compute median and ninetieth percentile per dock. Where the gap between the two is over an hour, that gap is the cheapest capacity you will find this year, and it costs a calendar rather than a licence.
↑ Back to the key points