The self-driving supply chain needs a map, not just an engine.

What if the most important question about AI in the supply chain isn't how fast we can automate. But whether we truly understand the system we're asking machines to run?

Alex Kruzel April 2026

Sources: McKinsey Global Survey on AI (March 2025, n=1,491); Fictiv 2026 State of Manufacturing; Reshoring Initiative 2025.

A few months ago I sat across from a COO who had just come from a board meeting where the company's AI supply chain "transformation" had been presented in confident terms, autonomous demand sensing, real-time inventory optimization, predictive sourcing risk. The slides were polished. The roadmap covered three years. And yet, when I asked him how many of those capabilities were live and running at scale across the network, he paused. Three of seventeen.

This is not a story about a company that is behind. This is, in my experience, a story about where most companies actually are, and the distance between that reality and the confident narrative being told to boards, investors, and press is one of the more consequential governance gaps I encounter in my advisory work.

The emergence of agentic AI, systems capable of making decisions and executing multi-step tasks with minimal human intervention, has renewed the ambition of what supply chains might become. The phrase "self-driving supply chain" has moved from theoretical to aspirational to strategic imperative, seemingly overnight (And the context accelerating that ambition is significant: American companies have made unprecedented commitments to bring manufacturing home) triggered by supply chain disruptions, shifting trade policy, and a genuine reckoning with single-source dependencies exposed by the pandemic and subsequent geopolitical turbulence.

My concern is not with the ambition. It is with the assumptions embedded in it. The companies getting this right are not the ones with the most sophisticated AI pilots (They are the ones that understood, early, that you cannot automate a supply chain you do not fully understand, and that the reshoring wave, rather than simplifying that problem) has dramatically complicated it.

In March 2026, Anthropic published a landmark labor market study, "Labor Market Impacts of AI: A New Measure and Early Evidence", introducing a concept they call "observed exposure": distinguishing between what AI systems are theoretically capable of doing in a given occupation and what they are actually doing in professional workflows.

The gap is not uniform. Click any occupation below to explore the data:

When a management team presents an "AI-enabled supply chain," they are almost always describing progress in the planning and analytics layer. These are real gains. But they are gains in the cognitive overhead of supply chain management, not in the physical execution of it. The factory floor, the warehouse, the last mile. These remain stubbornly human and mechanical.

The domestic manufacturing commitment announced by American companies over the past two years represents a genuine structural shift. As of mid-2025, private sector commitments to revitalize US chipmaking alone exceeded $500 billion, with the potential to triple domestic semiconductor capacity by 2032. Across manufacturing broadly, more than 24% of companies are actively reshoring or nearshoring, nearly double the rate from 2024. And 80% of manufacturing executives plan to invest 20% or more of their improvement budgets in smart manufacturing in 2026.

What is less visible in the announcements is what they require to become real operations. US labor costs run $25–30 per hour versus $6–7 in China. A gap that only closes with aggressive automation. Boston Consulting Group estimates reshoring adds 10–30% in costs versus offshoring. That cost premium only narrows with the AI-enabled operational infrastructure that most reshoring commitments have not yet built (the companies I find most compelling are those building the operational and data infrastructure in parallel with the physical commitment) treating the reshoring not as a destination but as a forcing function for supply chain redesign.

The phrase "self-driving supply chain" has real meaning, but it is precise in a way that most usage of it is not. PepsiCo describes it explicitly as deploying AI agents across all stages of the supply chain. From procurement and manufacturing to logistics and retail. With the goal of making autonomous, data-driven decisions continuously. BMW's "Virtual Factory" digital twins allow it to simulate production scenarios and optimize robotics before physical changes are made. These are genuinely different from companies running demand forecasting models and calling it transformation.

What distinguishes the leaders is not the sophistication of any individual AI application. It is the existence of a unified operational data layer: a real-time, high-quality, integrated data infrastructure that feeds every decision point in the supply chain simultaneously. Without it, agentic AI has no reliable surface to operate on. You get islands of optimization, not a self-driving system.

Not all supply chains face the same AI transformation opportunity or the same operational risk from mismanaging it. The sectors most materially affected, industrials, manufacturing, agriculture, retail, and logistics, face different structural constraints, different data architectures, and different governance questions. But they share a common dynamic: the distance between the AI opportunity available to them and the AI reality they have built is, in every case, wider than publicly acknowledged.

The questions I am most often asked in board settings are about whether management has the right roadmap (the more productive question, in my view) is whether the board has the right information to evaluate that claim. Supply chain AI transformation is one of the areas where the distance between the management narrative and the operational reality can be largest, and where the consequences of that gap materializing can be most material.

For the C-suite, the strategic imperative is demanding: AI investment in supply chain must be anchored to specific, finance-approved P&L outcomes, not capability milestones. McKinsey finds that fewer than 20% of organizations track KPIs for their gen AI solutions. Without measurement, you cannot manage the gap between the announced roadmap and the live reality.

For the board, the governance question is different: who owns this, and how will we know if it is working? Supply chain AI sits at the intersection of operations, technology, finance, and procurement in ways that most organizational structures have not resolved. Boards should be asking for baseline-and-outcome reporting, specific use cases, specific baselines, and specific targets. Rather than capability inventories. McKinsey finds that only 23% of organizations centralize AI adoption decisions, despite 57% centralizing risk and compliance. That asymmetry is a governance gap.

The privately held company context adds a further dimension. McKinsey finds that 62% of larger organizations ($500 million+ revenue) have a dedicated team to drive gen AI adoption, compared to 23% of smaller organizations. The technology is theoretically accessible. The organizational capacity is not, in most cases. I think the market is not being honest about that gap, particularly in the context of PE portfolio companies.

Alex Kruzel

CEO & Founder, Telesto Strategy

Alex advises corporate boards, management teams, and private equity sponsors on geopolitical risk, operational resilience, and sustainability strategy across the US and globally. She is the author of The Courage to Continue. She can be reached at www.alex-kruzel.com.