Agentic AI promises a more responsive supply chain. Instead of waiting for a planner to interpret a forecast or exception alert, an agent can pursue an operational target by redirecting shipments, balancing transport capacity, anticipating inventory shortages and reorganising warehouse work.
That capability can strengthen resilience when a motorway closes or a carrier fails. It can also create a new single point of operational uncertainty if nobody knows why the system acted, whether the action was proportionate or how to recover manually. Digitalisation therefore has to include governance from the beginning.
Build a control architecture around decisions
Resilient deployment starts with a catalogue of decisions, not a general label for the AI. Reordering within a defined stock band or resequencing pick tasks may be suitable for automatic execution. Actions that increase transport spend beyond a threshold, alter a customer commitment, use a carrier outside an approved network or affect safety should be routed to a named human decision-maker.
This approach also makes automation easier to scale. Quantified limits can be tested, audited and adjusted as the operation learns. Vague instructions to approve “important” changes are difficult to apply during an overnight disruption.
Human-centred automation needs operating skills
The EU AI Act’s approach to human oversight is relevant even beyond systems formally classified as high risk. Supervisors should understand capabilities and weaknesses, recognise the danger of over-reliance, challenge recommendations and interrupt the system. A control room overloaded with approvals will not deliver that outcome.
Resilience requires a rehearsed manual mode. Teams should know who can switch off an agent, how exceptions are escalated and how the next shift continues without automated decisions. This is particularly important when AI allocates tasks or evaluates workers: systems using behaviour, personal characteristics or performance may enter the high-risk area, while employment and data protection duties remain in force.
Training should be role-specific. A warehouse supervisor needs to know the agent’s data sources and failure patterns; operators need a clear route for reporting an implausible instruction. The practical readiness test is simple: identify the decision owner, autonomy limits, review workload, shutdown authority and workforce impact. If any answer depends on improvisation, the deployment is not yet resilient.
