Supply Chain Action Points
Read this first — the conclusion, and the moves to make:
- Before you trust any forecast percentage, request the metric definition in writing within 30 days: pin down the granularity, the horizon and the unit of measure, and treat a number without those three as a network-level figure until proven otherwise.
- Run your own backtest on the operator's promise: take 90 days of your own order history and compute error at SKU-location-day level, then compare it against the aggregate number you were quoted.
- Ask for the re-plan frequency, not the re-plan speed: find out how many times a day the routing can be re-solved and who is allowed to trigger it, and set a target of at least one intra-day re-plan on your highest-volume lane.
- Treat the manual tail as your own risk, not the operator's: identify the exception categories inside the non-automated share, and agree a named contact and a response time for each before the next peak.
- Set a quarterly review that tracks tail error, not average error: hold the baseline from your own backtest and re-price or re-route if the tail accuracy slips for two quarters running.
At the Beijing International Transport Expo (26-28 August), SF showcased its 'Apu' AI agent cluster with daily demand-forecast accuracy above 98%, while JD Logistics launched 'SuperBrain 3.0', cutting end-to-end routing for 100 million parcels from minutes to seconds across 3,600+ warehouses and lifting vehicle auto-dispatch to 86%. Both signal a shift from experience-led to algorithm-led logistics.
The Analysis
At the Beijing International Transport Expo, held 26 to 28 August, two of China's biggest logistics operators put their numbers on the wall. SF showed its Apu AI agent cluster, claiming daily demand-forecast accuracy above 98%. JD Logistics launched SuperBrain 3.0, which it says cuts end-to-end routing for 100 million parcels from minutes to seconds, spans more than 3,600 warehouses, and lifts vehicle auto-dispatch to 86%.
Two numbers, two claims, and a headline that says logistics is moving from experience-led to algorithm-led. I would slow down before accepting either number at face value, and not because I doubt the operators. Because the numbers do not mean what the headline wants them to mean.
Take it apart the way you would take apart any throughput claim. What is being measured, at what granularity, and where does the queue move to.
The forecasts first. Accuracy above 98% on a daily demand forecast is a strong number right up until you ask at what level it is measured. Forecast accuracy at the level of a whole network on a given day is a much easier number than accuracy at the level of a single SKU at a single location on a given day, and both can be reported honestly as over 98%. The first tells you the aggregate will roughly land. The second tells you the picker will find the right stock in the right bin. Those are different businesses. Whenever you see one accuracy figure and no granularity, assume it is the friendlier of the two until you are shown otherwise, and ask for the SKU-location-day number specifically.
Now the routing claim. Minutes to seconds on 100 million parcels is a latency number, and latency is not an outcome. Nobody's customer feels a 10-second re-plan. What the customer feels is whether the parcel arrives when promised, and whether the exception gets caught before it becomes a complaint. So the honest question is not how fast the plan runs. It is how often, given that speed, the plan can be re-run, and whether re-running it actually improves anything. The value of a plan that used to be too slow to bother with is that you can now bother. That is the real claim hiding under the stopwatch.
The forces here are boring and that is a compliment. Order data went digital years ago, warehouse scanning turned every pallet into a row, and the compute to re-solve a large routing problem dropped below the cost of a planner's coffee break. When the machine is cheap and the data is clean, planning frequency stops being a scarce resource and starts being a dial you can turn up. Both operators are turning it up, and the reason is the same reason any queue gets attention: the wait was longer than the work.
Who feels it first. The operator itself, inside its own network, feels it first and most, because it owns the warehouses, the fleet and the data, and can re-measure without asking anyone. Shippers using that operator feel it second, through the service level: cut-off times, dwell, how fast an exception is caught. Who barely notices: the shipper moving predictable, high-volume freight on stable lanes between fixed points, whose problem was never forecast accuracy but a receiver who takes deliveries only on Tuesday mornings. If your constraint is a delivery window, a better forecast does not move it.
Timing runs on the planning calendar, not the news cycle. A 98% forecast that is only good at network level will show up in your service terms as slightly steadier cut-offs before it shows up as anything you can see. The seconds-level re-plan arrives sooner, if at all, because it only changes what the operator does internally. Expect the internal benefit in one to two quarters and any customer-visible effect four to six quarters out, and expect it to appear first on the lanes where the operator has the most volume and the most freedom to reshuffle.
Here is what both headlines skip, and it is the same point twice. A percentage that sounds absolute rarely is. The 98% is only as good as its granularity, and the 86% auto-dispatch means 14% still goes to a human. That 14% is not the boring leftover, it is the tail, and the tail is where the money is spent. The easy dispatches automated first, because they were easy. The remaining 14% are the ones with the weird receiver, the damaged case, the address nobody can parse, the driver who quit mid-shift.
Pushing auto-dispatch from 86% to 95% is not 9% more work, it is the hardest 9% left, and the cost curve turns steep exactly where the number looks almost done. The second skipped point is the same idea applied to the forecast. A model trained on stable patterns does not get more accurate when the pattern breaks. It degrades at the exact moment accuracy matters most, which is during a promotion, a new product launch, or a port disruption.
Put a number on it, assumptions stated. Take the 100 million parcels a day. Assume full-network routing used to run as a ten-minute batch and now runs in ten seconds. The direct saving is trivial: a planner who waited ten minutes now waits seconds, maybe thirty minutes a day back, worth a couple of hundred dollars a month. The real figure is the frequency it unlocks. If a ten-minute run allowed one re-plan per shift and a ten-second run allows twelve, you have eleven extra re-plans a day. Assume each re-plan recovers 0.01% of line-haul cost, and assume line-haul spend is about 0.50 dollars per parcel, so 50 million dollars a day.
That is 5,000 dollars per re-plan, about 55,000 dollars a day, or roughly 20 million dollars a year. Now be honest, because the improvement per re-plan decays fast: the first few re-plans catch the big errors and the eleventh mostly reshuffles noise. Cut the assumed recovery to a third and you still clear 6 million dollars a year. The latency win is real, but it lives in the re-plan frequency, not in the seconds.
Reverse one assumption and the conclusion flips. If the operator's forecast is measured at SKU-location-day level and still clears 98%, that is a genuinely hard number and you should expect the service to hold up under promotion and disruption. If the number is measured at network-day level and the SKU-level error is materially worse, then the headline is a marketing artefact and your safety stock still does the real work. The condition to watch is not the headline percentage. It is the forecast error in the tail, on the days and the SKUs that hurt.
↑ Back to the key points