Generative AI beaten by basic rules?

To teach the principles of Demand Driven flow, we use the "Flow Sim Game"

The Beer Game

This simulation is based on the Beer Game, a classic game developed by MIT in the 1960s. If you work in supply chain, this is a highly recommended experience that allows you to get a hands-on understanding of the problem that is wreaking havoc in global supply chains: the bullwhip effect.

We run two iterations in this game:

1. Classic Beer Game: each participant works to the best of their ability, according to their forecasts and management rules. Inevitably, a superb bullwhip effect is generated. Here is an example generated by one of our clients during the training session
2. We insert a DDMRP buffer-driven flow logic. While demand variability in this second scenario is much higher, the bullwhip effect is greatly reduced—service rates, inventory turnover, and costs are greatly improved. Between the two iterations, supply chain costs are divided by a factor of 2 to 3.

We can therefore see that simple and pragmatic operating principles, known for decades, yield excellent results in effectively mitigating the bullwhip effect and the associated costs: you just need to implement decoupling and end-to-end pull flow.

In the age of AI, all this seems a little trivial though.

Can AI agents manage a supply chain?

This is what a team of researchers from Harvard, MIT, and Georgia Tech wanted to check by putting generative AI to the test in the Beer Game.

Can AI agents successfully manage a supply chain while avoiding the bullwhip effect?

Spoiler alert: no!

You can test this yourself on the Beer Game by AI Agents website and put the LLM of your choice to the test. However, even with demand sharing and orchestration options, the results are disappointing.

Despite massive investments in these AI models, despite the resources used to run them, the results are no better than those of humans who have not yet been taught a few common-sense rules.

The most worrying thing is that the Beer Game supply chain is extremely simple. A single product, four links. If AI models can't handle that, how can we reasonably imagine deploying them on a large scale in a real end-to-end supply chain?

At a time when large companies, software publishers, the Big Four, Gartner, and other pundits swear by artificial intelligence, perhaps we should refocus on the fundamentals that operational practice has taught us since at least the 1990s.

However, we do have an opportunity: to teach pull flow techniques to AI models...