Powering supply chain with agentic AI

Supply chain organizations have invested heavily in technology over the past several decades, particularly in planning and insight generation. Yet turning those insights into action remains a persistent challenge for many companies. Advances in artificial intelligence, especially agentic AI, in which AI-based agents can orchestrate decisions across planning, logistics, and operations, have the potential to help close that execution gap. But how can companies maximize the value of their investments in agentic AI?

In this episode of McKinsey Talks Operations, host Christian Johnson explores the opportunities for applying agentic AI across the supply chain. He is joined by McKinsey Partner Rahul Shahani, who leads AI and technology enablement for manufacturing and supply chain and collaborates with the World Economic Forum on the Global Lighthouse Network of advanced manufacturers. Also joining the conversation is Kapil Dev Bansal, an associate partner and McKinsey’s global lead for agentic AI in manufacturing and supply chain. His research has developed the task goal orchestrator framework for autonomous supply chain planning.

The following conversation has been edited for length and clarity.

Christian Johnson: Based on what we’ve seen over the past few years, supply chain was an early adopter of some forms of what we now call AI. What are some of the lessons learned so far on what has worked and what hasn’t?

Rahul Shahani: Supply chain was an early adopter, investing heavily in foundational tech, such as ERP [enterprise resource planning] tools, analytics capabilities, control towers, and more. A lot of those tools are focused on keeping and creating visibility across complex global supply chains. So we’ve had visibility and used that insight to drive complex decisions.

But execution of the supply chain remains manual, fragmented, slow, and quite frankly, bureaucratic. Decisions still rely on emails, spreadsheets, phone calls, and handoffs, slowing the path from knowing what to do to actually doing it. Organizations today are starting to experiment with execution AI or agentic AI, as we will talk about. And what we’re seeing is that while everybody is excited about it, they are starting to get stuck in pilot purgatory when it comes to agentic AI.

Christian Johnson: Pilot purgatory, so that’s a phrase I’ve certainly heard before, but could you explain it for our listeners?

Rahul Shahani: Absolutely, pilot purgatory comes out of our research with the World Economic Forum on the Global Lighthouse Network. The findings that we’ve had—which have been true for over the last decade—is that every time we see new tech come out, many institutions gravitate toward it, and very quickly deploy the tech, but then struggle to see meaningful impact from the tech and scale it.

That’s really what we’re seeing here as well. Many organizations have started agentic AI pilots and have very quickly shown proof of concepts and proof that it’s working, but they have not been able to change their core business processes to bake it into the ways of working and drive value.

Christian Johnson: We’ve heard a lot about the promises of AI in the various forms that we’ve seen over the years. How can AI-based agents help overcome some of the issues we’ve seen?

Kapil Dev Bansal: Agentic AI is a technology focused heavily on enabling the execution. AI stacks are now a lot more modular, reusable, and repeatable, and that can help scale up in a way that was not possible with the prior technologies.

For example, we break down the agentic AI ecosystem into three different types of agents. The first is a task agent, which is doing a very simple task. The second is a goal agent, which is doing a very well-defined sequence of tasks for executing the supply chain. And third is the orchestration agent, where the sequence of steps is not very well defined, but the agent has to think and define what it needs to do, and then it goes and does it. So that’s a whole system of doing, thinking, and repeating it over and over. That’s something that was not possible in the prior technologies, and that’s why agents are a lot more impactful.

Once fully implemented and adopted, agentic AI has the potential to fundamentally reshape the supply chain. The key takeaway is that the value of agentic AI isn’t theoretical, but it isn’t automatic either—there has to be a lot of purpose and thinking on how we shape the journey.

Christian Johnson: I think that’s one of the things we really want to help our listeners understand—that agentic isn’t an automatic “set and forget” success mode. There are things you need to do to make it work properly. So what are the sort of failure modes that you see right now with some of the agentic efforts?

Rahul Shahani: What we’ve seen so far with traditional AI, and with emerging agentic AI efforts, is that they’ve traditionally been siloed. That means they’ve taken a very narrow focus and lean into use case thinking. That limits our ability to focus only on the task-based agents, as Kapil said. We’re not able to orchestrate and drive toward common goals, which is what the full potential of our efforts really unlocks.

As we see organizations take a more holistic and thorough approach to deploying AI into their core business processes, we start to see a significant uplift in AI. The conversation shifts from “What is possible with AI?” to “How do I use AI and trust it to take over core capabilities within my workflows?”

Christian Johnson: What are some of the differences that you see between organizations that are still at a very early stage and the ones that are really starting to see some results from their AI investments?

Rahul Shahani: Ninety percent of organizations report using AI, but only 7 percent have scaled. And the ones that have scaled are really looking to AI to take over accountability for certain processes. Or rather, it’s individuals in those organizations who take over accountability for work done by AI. To do that, they require the AI to be accurate, explainable, and auditable, as other employees in their teams and organizations are. That becomes the unlock for how you drive real value from agentic AI: It’s not just about a point solution or a technology; it’s about an evolution of our business processes to support AI at its center.

Christian Johnson: I’d love to give our listeners an example of the from–to here: going from this fragmented use case mentality to something that allows AI systems to take over accountability for certain types of workflows. What are some good examples of that in supply chain?

Kapil Dev Bansal: Let’s talk about order management. Traditional order management can be one of the slowest, most manual-intensive, and error-prone processes. Why? Because it must be done across different systems. You need one system to check your inventory, a second system to check the customer order, and a third to check if you have a logistics supplier available. This ecosystem is spread across different systems and is very manual. Even though systems are efficient in their own environment, they don’t talk to one another.

Christian Johnson: If you say these systems aren’t talking to one another, how does a company get real end-to-end value from their investments in AI?

Kapil Dev Bansal: In agentic order management, ATP [available to promise] and CTP [capable to promise] checks are done by agents across different systems. Agents can also check the transportation feasibility and cost-to-serve validation. In some cases, even if you don’t have a TMS [transportation management system] to calculate your cost to serve, agents can do it via spreadsheets, or they can build custom calculations. Then you need to have trade-off conversations around inventory cost and service, which agents can run—not just based on feelings but based on real value. And then they can do the autonomous execution within defined thresholds.

When we did the agentic order management for one of our clients, we saw the cycle time drop from 20 to 120 minutes to one or two minutes. Obviously, the improved customer experience made the customer happy, and then there was a lot of operational efficiency because of the reduced manual work. Value is demonstrable when agentic AI is orchestrated across systems and stitches together the systems and information.

Christian Johnson: That’s a remarkable change because you’re going from a process that previously took 20 minutes to maybe two hours, and you’re taking it down to minutes, maybe even seconds in some instances, correct?

Kapil Dev Bansal: Absolutely.

Christian Johnson: What are some of the things that supply chain leaders need to do to enable this shift from focusing just on the AI technology to real transformation across their systems?

Rahul Shahani: I think this largely harkens back to many of the traditional levers in transformation. At the core of this is the “what”—what does the business-led reimagination of the operation look like? These are clear aspirations tied to business outcomes. And what do I need to change to change the trajectory of my performance?

Then ask what are the domains that you’re prioritizing? A domain-level transformation that focuses on end-to-end workflow orchestration versus individual pilots. You’re taking a capability, putting in the infrastructure, and then having a series of agents tackling the tasks, the goals, and the orchestration of them to transform the work. And last, you need leaders who own the entire end-to-end learning journey of this process. That’s the “what.”

In parallel to this, you also need the “how,” which is where you start getting a little bit more agentic-specific, because here you need the core execution enablers. This is the end-to-end process redesign built around the capabilities of agentic AI. That includes the data products, the underlying tools and infrastructure, and the idea of atomic agents, which are agents that are built to perform tasks that are modular and repeatable so you can rapidly scale them up.

You need a governance model to monitor and performance manage these agents and continuously train and evolve them. The workforce, the human workers who are playing new roles, become agent supervisors, or as I like to call them, agentic tamers. They are overseeing a whole new set of capabilities and skills that we’ve never had to manage before.

And last, it also requires responsible adoption of AI with guardrails and human-in-the-loop controls as we build the accuracy of AI, the explainability of AI, and the auditability of AI in our system.

Christian Johnson: I love the phrase “agentic tamer.” What does an agentic tamer mean for somebody like me working in supply chain?

Rahul Shahani: Agentic tamer refers to a new skill set, a new task that many of our workers will need to upskill themselves on. It’s about asking, “How do I manage a group of agents that are performing a task in my business?” “How do I become the leader of a series of agents?”

That means overseeing the agents to make sure they’re performing their work accurately, and that they are not seeing bias in the execution of the tasks that agents are doing—and if they are, how do you train or tune the agent to be able to get back to a nominal correct outcome or a nominal expected outcome within the guardrails of execution.

An agentic tamer becomes a new muscle that we all need to build to help drive the right outcomes from the fleet of agents that we’re taking accountability for.

Christian Johnson: I think that’s a great summary of the “what” and the “how” for companies in thinking about AI across their supply chains. What’s the best way for companies to make the adoption feel achievable for them? Because of what you’ve described so far, this is a lot to undertake. How can this feel like something that companies can achieve realistically and get lasting value from, as opposed to a bunch of pilots that don’t really end up meaning very much?

Rahul Shahani: That’s a great question. The most important place is knowing where to start. And this goes back to understanding where the business value sits. Trapped business value is in domains that are ripe for tackling through agentic AI. This largely means business domains that are information-flow heavy, which is the sweet spot for agentic AI.

I would start with a high-value, information-flow-heavy business domain. This could be quality, it could be procurement, it could be planning—it’s somewhere we drive a lot of information flow. Pair that domain with clear metrics for cost, service, resilience, and agility. Ask what are the outcomes you want to drive toward?

Invest in data readiness. And interestingly, data readiness in an agentic world is a lot less rigid than data readiness we’ve seen in the past because agentic AI is capable of ingesting fuzzy data as well as monitoring, maintaining, and improving data quality and performance.

Last, it’s about workforce enablement embedded from day one—how do I train and upskill my workforce to become agent handlers out of the gate.

Christian Johnson: I’d like to be able to give the listeners an example of what it means to go from the use case mentality through to the full orchestration. Are there any examples where clients were really surprised by where the value lay? They initially thought it was going to be X, and that was essentially thinking on a use case basis, before realizing where the real value lay across the end-to-end process.

Rahul Shahani: Let me paint the picture of an order coming in. Traditionally, when an order comes in, many times clients think the value sits in saying, “Do I know what I have in stock?” “Where should I dispatch this from?” “Where should I route this?”

If I have this in the control tower error system, I’m able to answer that question very quickly and process the order. The reality is that this process tends to break down in the handovers between organizations and functions.

When the order comes in, it is the person looking at the dashboard saying, “I have 200 units in stock in warehouse A. I’m going to call up the supervisor at warehouse A to confirm that they have that in stock. Then once the supervisor confirms, I’m going to see whether I can ship it. I’m not sure if the truck can come in and pick it up, so I’m going to call vendor X or send an email and wait for 24 hours for the response on whether I can fulfill this order effectively or not.”

The idea here is that agentic is built around the master data that allows us to answer a lot of these questions autonomously or automatically trigger those questions to the right stakeholders. Through that orchestration, we can speed up these cycles from hours or days to minutes. That unlock happens because what we’re really tackling here is the simplification of the organizational friction that happens from day to day, and that speeds up everyday processes and cycles.

Kapil Dev Bansal: Just to echo that point: A lot of clients think the value is in doing a lot of predictions or a lot of analysis. But clients soon realize that with all the BI [business intelligence] and all their technology, they already have a lot of insights. They just don’t know which insights to work on, which ones to prioritize, and which ones to act on. The surprise is that agents help to prioritize the available insights and then help to execute on them.

That’s where the real unlock is. It is less about what more can you know about your supply chain, because your existing tools, the BI tools, already provide enough insights. Instead, agentic becomes the execution layer, which is often a pleasant surprise for clients doing this work.

Christian Johnson: What would you like to say to people out there who may be concerned about their employees? Or some of the employees listening who may be concerned about their own futures in a world in which AI is now taking over some of the tasks that have taken up a lot of their day-to-day work?

Rahul Shahani: When we think about agentic AI, the role of the individual is evolving from a doer or controller of the task to a conductor or an overseer of the task. That means AI allows us to move a lot faster as individuals because we have the agents doing the execution activity. That instead allows us to oversee the agents and ask, “Are we making the right decisions?” and “Is the agent performing reliably and repeatably in its activities?” We effectively move ourselves into performance management of these small, atomic microagents that are driving the execution of the tasks.

That does mean that, as end users, or overseers of agents, we need to build new skills. We need to learn how to program agents, how to tune agents, and how to manage agents’ performance, which are relatively new skills in the workforce that we will rapidly have to pick up.

Last, agents don’t remove human judgment—they relocate it. Because in many cases humans are still in the loop, auditing many processes, especially the more complex processes. The agents are allowing us to spend more of our time being the expert versus being the executor. And that is how we tend to upskill our teams to focus on the more complex problems.

Christian Johnson: What is some of the real value we are seeing agentic AI create?

Kapil Dev Bansal: We have seen agentic unlocking real value for the clients where we have supported transformations. It can reduce the cost of goods sold by 4 to 7 percent. It can improve productivity by 20 to 50 percent. And it can shrink decision cycles from weeks to hours and in some cases, even to seconds.

Once fully implemented and adopted, agentic AI has the potential to fundamentally reshape the supply chain. In fact, it is the most disruptive technology that has been discovered in the past two decades for the supply chain. But the value is not automatic. There is serious work that must be done to get to the value. The value is not theoretical—it requires a lot of purposeful thinking about how we go and get it.

Christian Johnson: What we’ve been hearing today is that with agentic AI, the supply chain won’t just plan better—it will act continuously, autonomously, and with human judgment where it matters most.