Resilient Supply Chain Podcast: Agentic AI Shifts the Bottleneck to Human Decisions
Artificial intelligence promises to accelerate supply chain operations, but automating execution does not automatically improve decision-making. In a recent episode of the Resilient Supply Chain Podcast, host Tom Raftery speaks with Elon Salfati, CEO of Salfati Group, about the organisational changes, governance requirements and operational constraints that determine whether AI agents deliver measurable value. Their conversation examines why successful adoption depends less on deploying additional technology than on redesigning workflows, preserving human expertise and establishing clear accountability. The full episode is available at www.resilientsupplychainpodcast.com
From Task Automation to Operational Redesign
Many organisations approach AI as an additional layer of technology, deploying enterprise chatbots or individual agents while leaving existing processes largely unchanged. Salfati argues that this approach overlooks the more consequential opportunity: reconsidering how work is organised when routine execution no longer requires continuous human involvement. He describes a model built around "an army of agents with a human", reversing the conventional assumption that each employee should individually supervise multiple AI tools. Under this approach, agents handle defined operational tasks continuously, escalating decisions that require human judgement. Salfati illustrates the concept with his company's operating model, where he describes agents executing work overnight and presenting a small number of decisions for human review. The strategic implication is a shift in management attention from supervising individual activities to governing processes and resolving exceptions.
The Knowledge That Systems Cannot Capture
One of the most significant obstacles to AI adoption is information that never enters corporate systems. Supply chains frequently depend on experienced employees who understand informal procedures, supplier relationships and workarounds developed over years. These capabilities can be critical when established processes fail, yet remain largely invisible to automation systems. Salfati identifies a further complication: employees may perceive sharing their expertise as surrendering the knowledge that makes them valuable. This makes knowledge capture an organisational challenge rather than simply a technical exercise. Recording conversations and making their content accessible to AI may help preserve expertise, but introduces additional questions about confidentiality, permissions and commercially sensitive information.
Human Accountability Remains Essential
As agents take responsibility for more operational execution, deciding where human intervention belongs becomes increasingly consequential. Asked who should be accountable when an AI agent takes action, Salfati responds: "Hundred percent a human." The distinction matters because automated systems can complete tasks or generate recommendations without possessing the organisational context necessary to determine whether an outcome is appropriate. Salfati advocates retaining human responsibility for strategic direction while allowing AI to support analysis and execution. However, human oversight creates its own operational constraints. If every consequential decision requires approval, organisations must determine how exceptions are prioritised, which information decision-makers receive and how quickly unresolved issues can be escalated.
Verification Becomes an Operational Requirement
Greater automation also changes the work required to maintain control. Salfati acknowledges that AI systems can produce confident but incorrect outputs. Consequently, eliminating manual data entry does not eliminate the need for human involvement; some of that effort must shift into verification. Audit logs, decision records and approval histories become essential for establishing what an agent did, what information informed its actions and who authorised the outcome. Security introduces another constraint. Pressure to accelerate implementation can encourage organisations to bypass established safeguards, potentially expanding their exposure through systems granted access to operational information.
Measuring Outcomes Rather Than Adoption
The conversation also challenges organisations to distinguish AI deployment from operational improvement. Salfati argues that successful projects should begin with a measurable business outcome rather than a general instruction to adopt AI. Cross-functional handovers provide a particularly relevant example. Processes can suffer repeated delays when decisions are postponed or errors discovered late require earlier work to be repeated. Reducing end-to-end processing time and execution costs therefore offers a more meaningful measure of value than the number of agents deployed. Salfati also discusses emerging outcome-based commercial models, while acknowledging that their practical effectiveness remains to be demonstrated.
The Next Test for Operational Leadership
For supply chain executives, the central question is no longer simply which activities AI can automate. It is how organisations redesign responsibility, information flows and decision rights as automated execution expands. The organisations that benefit most will need to combine measurable operational outcomes with effective verification, knowledge governance and clear human accountability. Faster execution creates an opportunity for more resilient operations, but only when the decisions governing that execution remain reliable.
