The GenAI Beer Game 🍺
Can AI agents 🤖 manage a classic supply chain exercise?
Keep scrolling to explore
Best experienced on desktop
What is the Beer Game?
The Beer Distribution Game was developed at MIT in the 1960s and has been played by MBA and Executive students for over 60 years. The goal is simple: minimize supply chain costs. Four players take different roles (retailer, wholesaler, distributor, and factory) in a beer distribution network.
The original Beer Game board used in MBA classrooms
Customer
Retailer
(Player 1)
Wholesaler
(Player 2)
Distributor
(Player 3)
Factory
(Player 4)
How the Beer Game Works
The game works in rounds, each representing one week. Each player makes just ONE decision per round: how much to order from their supplier (or how much to produce as the factory).
Customer
Retailer
2?
Wholesaler
4?
Distributor
6?
Factory
1?
The Bullwhip Effect
Despite this simple decision, two things make the game challenging:
1. Limited Information: If you're not the retailer, you don't see what customers actually want.
2. Delivery Delays: Orders take time to arrive from suppliers.
Players must track both their on-hand inventory and their pipeline inventory (what they ordered but hasn't arrived yet). These complications create the famous Bullwhip Effect – small changes in customer demand create wild swings in orders upstream in the supply chain.
We will explore what happens when AI 🤖 plays the Beer Game.
Key Metrics
What happens when AI plays the Beer Game?
Customer
Retailer
Wholesaler
Distributor
Factory
- The LLM-powered Beer Distribution Game shares much with the classic version – but with a twist
- Instead of human players, this game is played among Large Language Models (LLMs) as autonomous decision makers
- Each LLM (powered by e.g., GPT-4o-mini) controls one component of the supply chain – retailer, wholesaler, distributor, or factory
- Identical to the traditional Beer Game, each role manages inventory, responds to downstream orders, and decides how many units to order upstream to minimize costs
- Users can choose from an extensive list of models which to deploy. For fast comparison of scenarios, GPT-4o-mini is recommended.
- In the default scenario, the prompt given to LLM agents informs them of their role, their current stock and incoming orders, the various costs they face (including holding cost and backorder costs) as well as order lead times. They are instructed to minimize total costs by placing orders strategically.
Orchestrator-Aided Decision-Making
The LLM-Powered Beer Game has additional options that model various situations. These differing options enable you to illustrate and compare concepts such as lead-time reduction and global information sharing. For information sharing, an orchestrator is introduced to share order or inventory information across the supply chain. For each scenario, additional information is added to the prompt to facilitate decision-making.
Scenarios
- 1. Information Sharing of Customer Demand: Consider a scenario exactly as described above, except that each supply chain member has full knowledge of the external demand of the customer in the past X weeks, with a given parameter X.
- 2. Information Sharing of Customer Demand Volatility Analysis: In addition to sharing customer demand information, the game allows sharing customer demand volatility analysis over the last X weeks to all supply chain members to aid with memory and decision-making.
- 3. Information Sharing of Inventory Position: To alleviate the memory bottleneck of AI agents (and humans) associated with tracking unfulfilled orders resulting from order lead-time delays, the game allows sharing to each player both pipeline inventory and their inventory position, the latter defined as the sum of on-hand inventory, backlog, and pipeline inventory.
- 4. Information Sharing of Downstream Inventory Position: Additionally, the game allows each agent to access the inventory position of its immediate downstream partner. This shared information enables upstream agents to anticipate demand more accurately and align their replenishment decisions with downstream needs.
- 5. Lead Time Impact: To illustrate the effect of lead time, order lead times can be reduced from the two weeks (default) described above to only one week. Finally, additional production lead time can be incurred to the factory for realistic simulation.
Live Simulation
Experience the Beer Game with AI agents for the first time
- STEP 1: Set your simulation parameters that determines the supply chain scenario
- STEP 2: Click "Run Simulation" to start – LLMs play the game in the background
- STEP 3: Visualize the outcome of the game, the bullwhip effect and its associated metrics
Lessons Learnt from the GenAI Beer Game
- #1 Decentralized supply chains expose GenAI Agent weaknesses.
In the default decentralized setting – where each GenAI agent makes its own decisions, as in most real-world supply chains – most models suffer from severe bullwhip effects. GPT-5-mini is a notable exception. - #2 Adding a centralized orchestrator can improve performance.
Incorporating an orchestrator (who makes no decision) with visibility across the supply chain and shares information to GenAI agents (who makes their own decisions) show mixed results: direct customer demand sharing isn't always helpful, but analyzing demand volatility can dramatically improve outcomes. - #3 Pipeline and downstream visibility help in theory – but aren't translating into better decisions.
In theory, giving agents visibility into pipeline inventory should ease memory bottlenecks, and sharing downstream inventory positions should improve order forecasts. In practice, however, GenAI agents showed little to no performance improvement from these additional signals. - #4 Advanced models signal a broader opportunity.
State-of-the-art GenAI models (e.g., GPT-5-mini) can approximate near-optimal strategies even with minimal information sharing. The challenge now is to harness GenAI to make supply chain management efficient, reliable, and accessible – not just in isolated cases, but at scale.
Authors
Credits
The discussion of the traditional beer game above is based on the Computerized Beer Game described in Designing and Managing the Supply Chain: Concepts, Strategies and Case Studies, McGraw-Hill, IL, Third Edition, July 2007, D. Simchi-Levi, P. Kaminsky and E. Simchi-Levi.
Get in Touch
We'd love to hear your thoughts about the LLM-powered Beer Game!
Click the main button to email the team, or choose a specific category above