Supply chain leaders aren’t waiting to see whether agentic artificial intelligence works before putting it to work. Most say that AI agents already embedded in daily operations are speeding up decisions, and a strong majority of chief supply chain officers expect these agents to lift process efficiency.

The urgency makes sense. Geopolitical risk and trade tensions now top the list of what keeps supply chain executives up at night, and an agent that can reroute a shipment or requalify a supplier in minutes rather than days looks like a competitive necessity.

The trouble is that adoption and oversight aren’t moving at the same speed anywhere in the enterprise. Recent research across industries found that nearly three quarters of organizations expect to deploy agentic AI within two years, while only about one in five currently have a governance model mature enough to manage systems that act on their own. Supply chain functions, where agents are already writing purchase orders and rerouting freight, sit squarely inside that gap.

An agent that misreads a demand signal is a forecasting error. One that autonomously places an order with the wrong supplier, reroutes freight through a restricted corridor, or clears a shipment that should have triggered a compliance hold is an operational and legal event. Procurement agents already generate structured purchase orders from replenishment signals and supplier data with no person in the loop for routine categories. That works well until the agent optimizes for price or lead time in a way that introduces a sanctioned party or a customs misclassification into the chain.

Regulators aren’t treating this as hypothetical. In the European Union, AI systems used as safety components in road traffic and other critical infrastructure are classified as high risk, pulling a meaningful share of autonomous routing and dispatch systems into a documented, lifecycle-long compliance obligation rather than leaving them as an internal efficiency tool.

In the U.S., transportation regulators have moved in narrow, deliberate steps rather than one sweeping rule. For years, an obscure federal safety requirement that a human physically place warning triangles around a disabled truck functioned as a de facto ban on driverless freight, regardless of how capable the underlying system was. Only this past October did regulators grant a limited waiver allowing cab-mounted warning beacons to substitute for that human task, and only for the company that requested it. That’s a regulator drawing a line around one specific physical action, not granting blanket autonomy.

Singapore has chosen a different posture, building its national AI governance framework to evolve in short cycles, publishing a version, learning from live deployments, and revising within months. For a country whose ports and freight corridors move a disproportionate share of global trade, that iterative model is itself a bet about how fast supply chain autonomy needs to move.

Three regulatory philosophies, one set of freight lanes crossing all of them.

A workable response is to stop asking whether an agent is trusted and start asking which category a given decision falls into. The first tier covers decisions an agent can make unsupervised — typically high-frequency, low-consequence calls such as reordering a commodity SKU from an already qualified supplier. The second tier covers decisions that always require a human to confirm before execution, such as switching suppliers mid disruption or rerouting a shipment through an unfamiliar corridor. The third tier covers decisions that stay off limits to AI entirely, including sanctioned party screening and country of origin determinations, where the legal exposure of a wrong call is too severe to delegate.

Underneath the governance question sits an infrastructure one. Most transportation and enterprise resource planning systems were built to report status upward, not to execute instructions and confirm them back down the chain. An agent that reroutes a shipment needs to instruct a carrier system, receive confirmation that the change was accepted, and log that exchange for audit. Very few supply chain stacks have that closed, two-way loop built and tested, which is the reason so many pilots stall before reaching tier-one decisions at scale.

Liability hasn’t moved nearly as fast as capability. When an autonomous procurement or routing agent makes a call that turns out to be wrong, the company that issued the order or the shipment carries the legal and financial exposure, not the software provider and not the model. That unresolved accountability question is doing more to slow the move from tier two to tier one than any remaining technical limitation.

There’s a candid problem worth naming directly. Exception queues in supply chain operations are already overloaded, and flooding planners with every borderline agent decision creates fatigue and rubber-stamped approvals, not safety. The more durable approach is narrowing what requires review gradually, using validated outcome data to move specific, well-understood decision types out of tier two only after they’ve proved themselves in production, not in a demo.

That’s a slower, less dramatic story than most agentic AI coverage tells, but it’s the one supply chain leaders will actually need to execute.

Dijam Panigrahi is co-founder and chief operating officer of GridRaster.