Agents Are Not Enough

A chess computer that can recognize the pieces but has never studied the game will move legally every time. It will never blunder off the board. It will also never win against anyone who knows what they are doing.

That is roughly where many supply chain organizations are headed with agentic AI.

The race to deploy agents is real. Gartner forecasts that SCM software with agentic AI capabilities will grow from under $2 billion today to $53 billion by 2030, and predicts that 60% of enterprises will have adopted agentic AI features within that same window. Every major platform vendor is embedding agents into their workflows. Every conference session this year has "agentic" somewhere in the title. The momentum is genuine.

This momentum risks generating expensive noise if it is solely concerned with speed and not directed toward generating understanding that was not possible before.

Are you asking "how many agents do we have?" or "what do our agents actually know?"

Here is the uncomfortable truth that the platform vendors are not rushing to advertise: a basic AI agent is an executor. It receives a task, accesses tools, takes action. It can automate a procurement workflow. It can trigger a replenishment order when inventory drops below a threshold. It can generate a disruption alert and route it to the right inbox. These are genuinely useful things.

They are also things that rule-based automation has been able to do for twenty years.

McKinsey put it plainly in their 2025 analysis of agentic AI adoption: off-the-shelf agents embedded in software suites may streamline routine workflows, but they rarely unlock strategic advantage. The firms capturing transformational value, McKinsey found, are building custom agents for high-impact processes, agents armed with contextual intelligence specific to their operations, their market, their risk profile.

An agent that triggers a reorder when stock hits a floor number is executing a rule. An agent that understands the probabilistic odds of a stockout given current supplier lead time variability, port congestion trends, demand signals from three external data sources, and a seasonal pattern unique to your customer base is generating understanding that was not possible before. The difference between those two things is the difference between automation and intelligence.

The supply chain case for intelligence is not theoretical. It is already being measured.

The January/February 2025 issue of Harvard Business Review published research from MIT's David Simchi-Levi and colleagues at Microsoft, documenting how large language model-based technology in supply chain settings reduced decision time from days to minutes, while dramatically increasing the quality and reach of what planners and executives could actually evaluate. The insight was not just speed. It was that the intelligence layer allowed planners to explore scenario space they previously could not access at all.

Gartner's 2025 Supply Chain Technology Trends report identified what they call Decision Intelligence as a distinct and necessary layer, combining decision modeling, AI, and advanced analytics to support, augment, and automate decision-making in ways that move the needle on business outcomes. They did not describe it as a feature. They described it as a trend because it represents a fundamentally different category of capability from process automation.

Capgemini's 2025 research found that AI adoption in supply chains reduced fulfillment costs by 23% on average and improved forecast accuracy by up to 85%. Early McKinsey adopter data showed 15% lower logistics costs and 35% reductions in inventory. These are not efficiency gains at the margin. These are structural shifts in how supply chains perform.

And critically, none of them came from simply adding agents to existing software.

What separates execution from intelligence is what the agent carries into each decision.

Think about what a supply chain planner with thirty years of experience actually does when something breaks. They do not follow a flowchart. They draw on pattern recognition built across hundreds of disruptions. They understand which suppliers have historically over-promised and under-delivered. They know that the Q3 demand spike in one product line always lags the Q2 promotional cycle by six weeks. They carry a mental model of the system, its failure modes, its rhythms, its quirks.

That accumulated knowledge is the intelligence. The decisions are the output.

An agent without that intelligence layer is a fast, tireless rule-follower. Useful, but bounded. When conditions fall outside the rules it was given, it can escalate to a human or it acts on incomplete understanding and inference, which can accelerate bad decisions rather than prevent them.

Gartner's own research from late 2025 found that 43% of organizations trust AI agents with only limited or routine operational tasks. Only 6% of companies fully trust agents to handle core business processes without close supervision. That trust gap reflects an accurate reading of the intelligence gap between what agents can execute and what decisions actually require.

This is where the Intelligence Estate becomes the real competitive question.

The companies that will extract transformational value from agentic AI in supply chain are not the ones who activated the most agents. They are the ones who built the most intelligence for those agents to carry.

That intelligence is specific. It lives in proprietary data about their suppliers, their customers, their demand patterns, their failure modes. It is built through deliberate investment in the models, the contextual knowledge, and the organizational capability to tend it. It cannot be purchased off a shelf or subscribed to monthly on the same terms your competitor holds.

When that intelligence is embedded in agents, the math changes entirely. An agent that can run probabilistic simulations across thousands of scenarios before recommending an inventory decision, in real time, as conditions change, is not automating the old process. It is enabling a process that was never possible before.

That is the boundary between incremental and transformational. And that boundary is not defined by how many agents you deploy. It is defined by how much intelligence you arm them with.

The question worth asking

Gartner predicts that by 2031, 60% of supply chain disruptions will be resolved without human intervention. That headline is striking. The question worth asking is: resolved how? Resolved by agents executing pre-set rules under predictable conditions? Or resolved by agents reasoning through novel disruptions with enough contextual intelligence to find the right answer without a human in the loop?

The path to the second outcome runs through the intelligence layer. There is no shortcut.

The supply chain leaders who are building that intelligence now, who are treating their proprietary data and domain knowledge as an asset to be systematically captured and made available to their agents, are not ahead of a trend. They are ahead of their competitors.

The ones waiting for a software vendor to package that intelligence into a subscription are building a very fast chess computer that has never studied the game.