Automating the Order Can Automate the Chaos

New evidence suggests that handing replenishment to AI agents does not just inherit the bullwhip effect but can generate a fresh one.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link:

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

The pitch for AI in the supply chain is seductive and, in fairness, often true at the level of a single company. Feed a capable model real-time sales data, and it forecasts demand more accurately, trims inventory and smooths the orders it sends upstream. The classic bullwhip effect, in which a small wobble in consumer demand swells into wild swings at the factory, is exactly the kind of problem these tools are sold to solve. New research points to a complication that few of those sales decks mention: when AI agents run the ordering themselves, they can introduce volatility of their own.

CONTEXT AND BACKGROUND

The promise is real. Consultants report that firms deploying AI across the supply chain have cut inventory levels and logistics costs substantially, and the academic literature agrees that AI can, in principle, smooth the bullwhip effect by producing more accurate and flexible forecasts. The classic bullwhip was made worse by human latency, the lag while planners deliberated and orders were batched. Agentic ordering removes that pause; the model commits in an instant. That shift, from advice to action and from lag to instant commitment, is where the new risk appears.

INSIGHT AND ANALYSIS

A recent study put autonomous AI agents in charge of a classic supply-chain simulation, the Beer Game, and found something the brochures do not describe. On top of the familiar bullwhip, the agents produced a second layer of amplification the researchers call the agent bullwhip: order swings generated by the models’ own decisions rather than by any change in customer demand. Running the identical scenario again could produce materially different orders, because these models decide probabilistically and the same inputs need not produce the same choice, and this instability amplified both across the tiers of the chain and over time, even when consumer demand was held perfectly flat. The finding worth sitting with is that averaging over repeated runs did not remove it, which means it is not noise you can smooth away by asking the model twice.

This is not an argument that AI ordering is simply worse. The same study found that a capable reasoning model could beat human teams and dampen the bullwhip, while weaker models amplified it into far higher costs. The lesson is about reliability rather than average performance: a system can look excellent on the headline number and still swing unpredictably from one run to the next. The concern also grows at industry scale. Financial regulators have warned that when many organisations rely on models trained on similar data, those models tend to make correlated decisions that can amplify volatility rather than dampen it. A supply chain in which every tier runs a similar autonomous agent is exposed to the same dynamic.

IMPLICATIONS

For a supply-chain or board leader, three things follow. The first is that the question is not whether to use AI but which model, and how it was tested: a tool that will place orders should be checked for reliability under repeated, identical conditions, not only for average accuracy. The second is that automated ordering needs a circuit-breaker, such as a cap on how far an order can move from one period to the next, or a threshold above which a person must approve, so that a correlated overreaction cannot propagate through the chain in minutes. The third is to ground the model in real consumer demand rather than the inflated tier-to-tier orders that created the bullwhip in the first place, so the system does not learn the distortion as if it were signal.

CLOSING TAKEAWAY

AI can genuinely tame the bullwhip effect, and for many firms it will. The mistake would be to assume that automating the decision removes the volatility, when it can also relocate it, from the demand the chain reacts to, into the behaviour of the machines doing the reacting. Handing over the order is not the same as handing over the problem. The organisations that benefit will be the ones that automate with oversight, test for stability as well as accuracy, and keep a human close enough to pull the brake.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net