The AI that actually makes money in logistics is not a large model. It is a decades-old discipline

In the automated warehouse projects I have worked on, the biggest efficiency gain used no new buzzwords at all: we tuned equipment scheduling and wave planning together, and overall efficiency rose by about 20%.

The most impressive demo was a digital twin. In my projects, it never made it into real use.

This post covers two things: which kinds of AI actually earn money in logistics, and which kinds are overrated.

1. What earns money is the AI that “computes the best answer”

This kind of AI has an old name: operations research. It uses mathematics to find the best arrangement: how to plan routes, how to assign tasks, what order to do things in. It has been around for decades. Nothing about it is new.

Yet it has the most solid payback. SF Express has used operations research to plan its whole network since 2018. According to award materials published by INFORMS (the Institute for Operations Research and the Management Sciences) in 2025, it has saved the company more than USD 1 billion in total. UPS spent more than ten years and about USD 250 million building its route optimization system ORION, which is expected to save USD 300–400 million a year once fully deployed.

Warehouses are the same. The robot fleet model Amazon launched in 2025 changes how a robot fleet is assigned tasks and routed, and Amazon says it will improve robot travel efficiency by 10%.

My 20% follows the same logic. A wave is a batch of orders grouped together to be picked and shipped at the same time; it decides what work the equipment has in hand. Scheduling decides which machine does which job first. If waves are messy, scheduling can only choose among messy tasks. That is why, on site, these two were always tuned together.

2. What is overrated: digital twins, LLM dispatching, humanoid robots

Digital twins are 3D warehouses on a computer, synchronized with the site in real time. In demos, little vehicles move across the screen and it looks great. But it needs real-time, accurate data coming back from the site, and many warehouses have not even finished filling in their location master. Without that data, the warehouse on screen is just an animation. In my projects, it made it to display, not to use.

LLM dispatching. Academic surveys show that in operations research, large language models mainly help people build models and write solver code; the dispatching plans themselves are still computed by solvers. When JD Logistics describes its “Super Brain”, what it emphasizes is solving problems with tens of millions of variables, which is still operations research at heart. My view: LLM dispatching is still rather hollow today.

Humanoid robots. In January this year, Gartner predicted that by 2028, fewer than 20 companies will have taken humanoid robots to production use in manufacturing and supply chain. The robots actually working in warehouses are wheeled mobile robots and purpose-built robotic arms.

“Using AI” and “earning from AI” also need to be separated. A China Federation of Logistics & Purchasing report claims AI penetration in logistics and supply chain exceeds 37%, but does not disclose how that was calculated. BCG surveyed more than 180 logistics companies in January this year: only about one in ten had brought AI into core operations, and 13% reported measurable value.

3. So what comes first: get the site running smoothly

Gartner predicts that through 2026, 60% of AI projects without “AI-ready” data will be abandoned. In in-plant logistics, “AI-ready” means something very plain:

① Master data and the location master are correct. Part numbers, pack sizes, equipment parameters for each location: miss one column and the algorithm has nothing to compute with.

② You can change the scheduling and wave rules yourself. As the previous post explained, in a black-box solution both sit with the integrator. That 20% depended on being able to change the rules.

③ Someone reads the data every day. When efficiency drops, you can find which step caused it the same day, not wait for the month-end report. Start with hourly outbound volume, equipment waiting time and shortage counts.

Get these three right and operations research pays off right away. Get them wrong and no new buzzword will help.

---

Has anyone on your site put a digital twin to real use? In which step?