AI agents mimic human supply-chain decisions more closely with behavioral cues, study finds
Seven models played a simulated supply-chain game. Describing human choices—and even the pressure players might feel—improved the match, pointing toward more realistic simulations rather than better ordering decisions.
Researchers at Northeastern University found that language describing human decision patterns—including mistakes—made AI agents behave more like people in a simulated supply chain. The result matters for teams using agents to rehearse decisions with real-world consequences: a more human-like simulation may better reflect how people respond, but it is not necessarily more efficient or reliably predictive. The study tested seven models in 231 sessions; its proposed policy and business applications remain possibilities, not demonstrated outcomes.
01
The beer game assigned four players—retailer, wholesaler, distributor and factory—to make linked ordering decisions with incomplete information and delivery delays.
02
Behavioral guidance described how people respond to shortages, excess stock and inventory in transit without prescribing order quantities.
03
In a separate test, experience-based prompts made agents from six of seven models respond more like people, though some resisted the guidance.
AI agents became more convincing substitutes for human supply-chain players when researchers explained how people make decisions, including their mistakes. A study published November 16 in Nature Computational Science found that adding human behavioral context brought agents closer to a human benchmark—not necessarily closer to running the most efficient supply chain.
Northeastern University professor Babak Heydari and graduate student Hongan Zhu tested agents from seven AI models across 231 sessions. Northeastern Global News described the work on November 19.
Four players, incomplete information
The researchers used the “beer game,” a simulated supply chain with four roles: retailer, wholesaler, distributor and factory. Customer demand is set by the game. Each player sees only part of the system and must decide how much beer to order or supply.
Deliveries take time, so players also need to track supplies already on the way. Ordering too little leaves customers without beer; ordering too much creates surplus and storage costs. One player’s response changes what others face, making the game a test of interacting decisions rather than one isolated forecast.
Without the added behavioral context, some agents continued ordering excessively even after demand leveled off, rather than adjusting as people might.
Explaining choices without prescribing orders
Zhu and Heydari then added what they call cognitive grounding: language explaining patterns in human decision-making. The information came from previous beer-game experiments. It described how people estimate demand, react to shortages or excess stock, and sometimes fail to fully account for supplies already in transit.
Zhu described the guidance as a paraphrase of a mathematical model. Across the sessions, the agents’ behavior moved closer to the human benchmark, with some nearly matching it.
Another round of testing took a different route. Instead of explaining people’s decisions, the researchers described experiences players might have: uncertainty, pressure to order more, or an urge to overcorrect. Agents from six of the seven models responded more like people, though some agents remained resistant to the guidance.
A rehearsal for decisions with consequences
Heydari sees more realistic simulations as a way for businesses and policymakers to test changes before imposing real-world consequences. Potential uses include checking whether evacuation instructions create gridlock, whether cheaper transit fares draw drivers onto buses, or whether relocating a pharmacy makes it harder for existing customers to reach.
Those are possible applications, not outcomes demonstrated by the beer game. Sangeet Khemlani, a senior cognitive scientist at the U.S. Naval Research Laboratory, told Northeastern Global News that simulations can explore scenarios too risky to test on people. But inaccurate predictions can backfire, he cautioned.
For this task, reproducing human mistakes was part of making the simulation more faithful. The researchers redirected agents through language and observed their choices, rather than examining their internal neural processes—a behavioral approach to making simulated decisions look more like the people they represent.
Sources
news.northeastern.eduCan teams of AI agents recreate the chain reaction of human decisions?
Reader comments
Newest comments first. Replies stay oldest first.