AI is already standard equipment in supply chain software. It drafts forecasts, flags exceptions and answers planners' questions in plain language. The open question, though, is whether it's improving their decisions.
The early evidence is mixed. Gartner predicted last year that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Those are execution failures, and supply chain is where they get expensive fastest.
Supply chain teams decide under real constraints. Plants have finite capacity. Inventory is limited. Lead times matter. Customer commitments compete with production realities, and cost, service, labor and transportation objectives interact in ways unique to each business. AI makes that complexity easier to work with. But easier to work with doesn't automatically mean better decisions, and closing that gap is the industry's challenge right now.
Language models need a decision engine underneath
Generative AI excels at interaction. It helps people prepare data, ask questions in natural language and explore scenarios without becoming optimization specialists. However, many supply chain questions need something more: a system that evaluates possible actions against operational constraints and determines what is actually feasible.
Take a railcar, for example. In one operation it simply moves product. In another, it doubles as temporary storage when tanks are full. That single difference reshapes the entire distribution problem. A language model can parse the sentence describing it. A decision model has to work out what it means for capacity, inventory and timing.
Consistency matters too. Language models generate answers probabilistically, so the same question asked twice can return two different responses. That's tolerable when drafting an email and a serious problem when deciding which plant absorbs a lost production line's volume. Mathematical optimization is deterministic: the same inputs produce the same answer every time, traceable to the logic behind it.
That's why generative AI and optimization work best as partners. The companies I see getting real value from supply chain AI are building on both.
Disruption shortens the shelf life of every answer
A production line goes down, a tariff changes overnight or a carrier pulls capacity, and the information planners are working from starts losing value immediately. The goal is to shrink the time between recognizing what changed and making a sound decision.
A planner should be able to ask, "Can we still hit our service targets if Line 3 goes down next week?" and get more than a plausible-sounding explanation. The answer should reflect available capacity, inventory, transportation options and everything else that determines what the business can actually do.
Speed only pays off when people trust the answer enough to act on it. That's where we need far more rigor.
Guardrails belong in the architecture from day one
As AI moves closer to consequential decisions, trust can't be bolted on during implementation. Organizations need to see the assumptions behind a recommendation, trace how it was produced and know which data and constraints shaped it. Models should be validated before anyone relies on them, changes should be visible, actions reversible where it makes sense, and accountability for decisions affecting customers, employees and capital clearly assigned.
AI should widen the options planners and executives can explore. Ownership of the decision should stay exactly where it was.
Regulators are pushing the same way. The EU AI Act centers transparency and human oversight for higher-risk AI systems, even as its timeline shifts. Supply chain leaders don't need to wait for a regulator to hold their own systems to that standard.
Here's a useful test for any supply chain AI investment: ask the planner who will rely on it whether they'd stake next week's service targets on its recommendation. If the honest response is "not until I check it in a spreadsheet," the technology isn't finished yet.
