AI in the Chain

After the summer holidays, I am returning to my work with renewed energy and a long list of ideas I want to explore.

A break creates useful distance. It allows us to step away from the daily stream of meetings, exceptions, dashboards and urgent requests and ask a more fundamental question:

Are we improving the way the supply chain makes decisions, or are we simply adding more technology around the same old processes?

This question feels particularly relevant now. The conversation about artificial intelligence in supply chain has moved quickly. A year ago, much of the attention was on copilots: tools that could summarize a report, draft an email or answer a question.

Today, the language has shifted toward agents, multi-agent systems, autonomous workflows and even the self-driving supply chain.

Some of this progress is real. Some of it is vendor language moving faster than operational reality.

The most valuable opportunity may be less dramatic than a fully autonomous supply chain, but far more useful: building a faster and more reliable decision system.

The real problem is not a lack of information

Most companies already have more supply-chain information than their teams can absorb.

They have ERP transactions, demand plans, inventory reports, supplier confirmations, transport milestones, customer forecasts, CRM opportunities, risk alerts and market intelligence. Many have also invested in control towers and dashboards intended to make the network more visible.

Yet visibility does not automatically produce action.

A supplier delay may be identified on Monday. The affected orders are analyzed on Tuesday. Alternative supply is discussed on Wednesday. The commercial impact is reviewed on Thursday. Approval arrives on Friday, by which time the alternatives have changed or disappeared.

The organization saw the problem. It simply could not respond fast enough.

This is decision latency: the time between a meaningful signal appearing and an effective action being taken.

It includes more than analysis time. It also includes the waiting, reconciliation, clarification, escalation and approval that happen between functions.

A faster dashboard can reduce the time required to detect a problem. It does not necessarily reduce the time required to resolve it.

If the workflow remains fragmented, greater visibility may only allow people to watch the delay unfold in higher resolution.

Outside-in planning changes where decisions begin

Traditional planning tends to start inside the company.

The organization creates a forecast, converts it into a supply plan, positions inventory and then measures how closely reality followed the plan.

That approach becomes weaker when volatility is structural rather than exceptional.

Customer priorities change. Projects move. Suppliers face capacity or material constraints. Transport routes are disrupted. Trade controls evolve. Geopolitical events alter lead times and availability.

Outside-in planning begins with changes in the market and the network. It connects customer behavior, channel conditions, supplier constraints, logistics signals and external risks to internal decisions.

But outside-in planning does not mean continuously adding data feeds.

A company can ingest thousands of signals and still make poor decisions. The important question is whether a signal changes a decision.

Which customer orders should be protected? Should scarce material be allocated differently? Is an expedite justified? Should the forecast, inventory policy or sourcing assumption change? Who has the authority to make that trade-off?

More sensing without better decision rules creates more noise.

The value comes from connecting the right external signals to a defined decision and an executable workflow.

Where AI agents can genuinely help

This is where agentic AI becomes interesting.

The strongest near-term use cases are not agents independently running an entire supply chain. They are agents handling bounded, structured parts of exception management.

An agent could monitor shipment milestones, identify a delay, retrieve the affected orders, compare the delay with customer requirements and prepare a recommended response.

Another could check whether order data matches the purchase order, product structure or contractual requirements. A third could collect the evidence needed for a planner to compare allocation alternatives.

These workflows are valuable because they contain work that is necessary but slow: searching, matching, checking, summarizing, routing and preparing options.

AI can shorten the distance between signal and decision by doing this work continuously and consistently. It can allow specialists to spend less time assembling the situation and more time judging the trade-offs.

This is decision support with operational depth.

It goes beyond writing emails, but it does not pretend that every decision should be delegated.

A collection of agents is not an operating model

The arrival of multi-agent systems introduces another challenge.

Imagine four agents examining the same customer order.

The planning agent wants to preserve a scarce component for the demand with the highest probability. The customer-delivery agent wants to protect the contractual date. The logistics agent recommends an expensive premium route. The finance agent identifies credit exposure and recommends blocking the transaction.

Each agent may be correct within its own objective.

Together, they may recreate the functional conflicts that companies already experience between planning, sales, logistics and finance.

This is one of the most important lessons for supply-chain leaders: AI agents can reproduce organizational silos digitally.

Adding an orchestration layer is not only a technical integration task. The company must define shared objectives, priorities and decision rights.

It must decide how service, inventory, cost, revenue, cash and risk should be balanced. It must establish which agent provides evidence, which recommends an action, which can initiate a workflow and which situations require human arbitration.

Technology can make these rules executable. It cannot decide what the enterprise values most.

Human oversight needs to be designed carefully

“Human in the loop” has become the standard answer to concerns about AI risk.

It is necessary, but it is not a complete operating model.

If every agent recommendation enters the same approval queue, the human becomes the new bottleneck. The system may generate answers faster while the organization makes decisions at exactly the same speed as before.

Oversight should be based on risk and familiarity.

Low-risk, reversible actions can eventually be executed within clearly defined limits.

Medium-risk actions can be prepared by the agent and approved by an accountable person.

High-risk actions should be supported with evidence and escalated.

When the situation is unfamiliar, the data is incomplete or the recommendation falls outside established parameters, the agent should stop rather than improvise.

Write access should be earned.

An agent can begin with read-only access, move to recommendations, then initiate approved workflows and only later execute narrowly defined actions. This progression makes trust measurable instead of rhetorical.

Human expertise remains essential, but its role changes.

People will spend less time collecting information and more time defining objectives, examining assumptions, managing exceptions and judging consequences across the end-to-end system.

Technical accuracy is not the same as a good decision

AI teams often evaluate whether a system extracted the correct information, used an approved source and followed its instructions.

These tests are essential because large language model outputs can vary.

Supply-chain leaders must add a second layer of evaluation: did the system help the business make a better decision?

An agent may correctly identify that premium freight will recover a delivery date. That does not mean premium freight is the right decision.

The order value, contractual exposure, customer priority, inventory position and alternative uses of the material may change the answer.

Agent evaluation therefore needs both technical and operational measures.

Technical measures can include extraction accuracy, source traceability, workflow compliance and unsupported statements.

Operational measures can include decision latency, exception closure time, recommendation acceptance, human override frequency, avoidable expedites, service impact and inventory consequences.

This distinction matters.

A technically impressive agent can still optimize the wrong outcome.

How to get started

The best starting point is not to automate an end-to-end process.

It is to select one recurring exception that consumes time and has a measurable consequence.

Start with an exception such as delayed shipment investigation, missing order information, supplier confirmation mismatch or customer-priority review.

Map the current process from the first signal to the final action. Measure how long the complete cycle takes and identify where people wait, search, reconcile or request clarification.

Then define the agent’s role narrowly.

Specify the sources it may use, the evidence it must provide, the actions it may recommend and the situations it must escalate.

Create a representative set of test cases, including normal situations, incomplete data and difficult exceptions. Evaluate both technical correctness and business usefulness.

Keep the first implementation read-only or recommendation-based.

Review results weekly with the people who perform and depend on the process. Capture not only whether the agent was right, but also why people overrode it.

Those overrides may reveal missing context, unclear decision rules or disagreements in the operating model.

Finally, measure whether the time from signal to action has decreased.

Hours saved are useful, but the stronger outcome is a faster, more consistent and better-documented decision.

The leadership question behind the technology

The future supply chain will probably contain more AI agents.

They will monitor conditions, assemble evidence, generate scenarios, coordinate tasks and sometimes act within controlled boundaries. Standards that allow agents to communicate across tools and platforms may accelerate this development.

But the defining capability will not be the number of agents deployed.

It will be the quality of the decision system around them.

Does the organization share the same assumptions? Are trade-offs explicit? Are decision rights clear? Can teams move from an external signal to an approved action without days of reconciliation? Does technology strengthen end-to-end accountability, or automate each function separately?

As I return from the summer break feeling engaged and energized, this is the question I want to keep exploring:

Are we building AI around how supply chains should make decisions, or merely placing AI on top of the processes and silos we already have?

We may eventually reach something that deserves to be called a self-driving supply chain.

For now, the more urgent—and achievable—goal is a supply chain that senses earlier, decides faster and learns systematically, while keeping accountability where it belongs.

That would already be a meaningful transformation.

References

How AI Agents Are Transforming Supply Chains — BCG
https://www.bcg.com/publications/2026/how-ai-agents-are-transforming-supply-chains

Resilient by Design: The Agentic Supply Chain — Deloitte
https://www.deloitte.com/us/en/insights/industry/manufacturing-industrial-products/agentic-supply-chain-artificial-intelligence-manufacturing.html

Gartner Identifies Top Supply Chain Technology Trends for 2026
https://www.gartner.com/en/newsroom/press-releases/2026-06-30-gartner-identifies-top-supply-chain-technology-trends-for-2026

Google-Backed Agentic A2A Protocol Gets a New Home — Axios
https://www.axios.com/2026/08/17/a2a-agentic-ai-foundation-open-ai-standards

Global Critical Minerals Outlook 2026 — International Energy Agency
https://www.iea.org/reports/global-critical-minerals-outlook-2026