A recent paper on AI agents in supply chains examines both their benefits and their risks.
The researchers study autonomous agents in the MIT Beer Game, a simulated supply chain. Capable models can make very good decisions and reduce costs substantially.
They can also create new instability. The authors call this the “agent bullwhip effect”: small uncertainties or variations in decisions can grow along the supply chain.
Several agents working with incomplete data, delayed information or inappropriate goals can make the chain perform worse very quickly.
Before adding AI to logistics, we need to clarify:
- Which data may it use?
- Which decision may it prepare?
- Where is human approval required?
- Which goals apply?
- How can we prevent a local optimisation from damaging the overall system?
These questions determine whether AI is useful in operations.
AI know-how for logistics professionals in 2 minutes
What are guardrails?
Guardrails define what an AI system may do, what it may not do and when it must stop.
In logistics, examples include:
- Do not send email without approval.
- Never ignore a permit condition because it rarely occurs.
- Flag uncertain data.
- Escalate conflicts between a permit and an order.
- Document decisions.
Guardrails provide operational safeguards. People’s role changes as AI gathers some of the information: they still need to define goals, boundaries and decision criteria.