The Bosch RL Dispatch Result Is Real. So Is Its Limit.

A new Bosch case study shows reinforcement learning cutting truck dispatch costs by up to 23%. Read that as "OR-based dispatch is obsolete" and you'd be wrong.

The result

In August, researchers at the Bosch Center for Artificial Intelligence in Shanghai tested a transformer-based RL policy on three real milk-run delivery networks feeding manufacturing plants. Against standard less-than-truckload freight pricing, it cut costs 18-23% and kept empty capacity under 25%, on real order data (171 to 504 orders per network) — not a simulation.

That's a genuinely strong result.

The catch

But look at what the authors themselves flag: the policy only works reliably when new orders resemble the data it trained on, and generalizing it to new demand patterns is still an open problem. In other words, it's excellent at the exact shape of problem it was shown, and untested outside that shape.

Why OR isn't going anywhere

Compare that to classic OR-based dispatch: mixed-integer programming and routing heuristics don't need retraining when a new depot opens or order volumes shift. They come with provable guarantees, they're auditable by a human planner, and decades of engineering have made them fast enough to run in production at scale. What they don't do is adapt on their own or squeeze out that last double-digit percentage of savings in a narrow, stable use case.

The open question

So the honest read isn't "RL beats OR," it's that RL earns its complexity in a specific, well-defined slice of the problem, while OR remains the backbone for everything else. The question worth arguing about: how do you decide when a network is stable and well-characterized enough to justify a learned policy, versus when you're better off with a solver you can actually explain to an auditor?

Our take

This is exactly the call we make with clients building out dispatch automation — knowing when a learned model starts providing real value and is sustainable, and when a well-tuned OR solver is still the right answer.

Link to the paper: arxiv.org/abs/2608.06668