The End of Planner Heroics: How AI and Decision Engineering Are Reshaping Supply Chain Planning

Speaker 1:

Hi everybody. I’m Bob Bowman, editor-in-chief of Supply Chain Brain. Welcome to this special presentation, The End of Planner Heroics: How AI and Decision Engineering are reshaping supply chain planning presented by GAINS. Quick reminder, there will be an audience question and answer session at the end of this presentation. Audience members are encouraged to submit their questions at any time during the presentation by clicking on that Q&A icon at the bottom of your screen. So for decades, supply chain planning systems have generated forecasts, plans, and exceptions, but they’ve never made any actual decisions. That has been the planner’s job. Experienced planners knew which forecast to trust, how suppliers behave under pressure, and which trade-offs matter most during disruptions. Today though, that model is reaching its limits. Supply chains are more volatile and interconnected than ever before. Planners are managing more products, more signals, more channels, and more disruptions with the same tools.

Now AI technologies are creating a real opportunity to rethink how supply chain decisions get made and executed. So today we’re going to discuss how AI, machine learning and decision engineering can help planners work faster, focus on higher value exceptions and make more consistent decisions under pressure. With that, I’d like to introduce our speakers for today. Knut Alicke is a partner emeritus and independent senior advisor, supply chain executive, and thought leader with more than 25 years of global consulting experience at McKinsey & Company, including 21 years as partner and master expert. He has deep expertise in end-to-end supply chain transformation, digital supply chain, sales and operations planning, inventory optimization, warehouse automation and AI and Gen AI integration. Jeff Metersky is senior vice president of strategy and innovation at GAINS with over 40 years of experience supplying advanced analytics to improve supply chain performance. Jeff today drives solution strategies and leads GAINS’ applied research group advancing capabilities in sochastic optimization, simulation, and network design.

Jeff has helped hundreds of organizations across industries and geographies adopt supply chain planning principles and technologies. So with that, I’d like to turn it over to Knut for a presentation before our panel discussion comes a little bit later, but Knut, for now, take it away. It’s all yours.

Speaker 2:

Thank you very much, Bob, for the nice introduction. So I want to first start with a couple of thoughts on what I call the planner was the system. So as Bob said, I spent my life in supply chain logistics. I started in a small startup where we did planning software and that was the late ’90s and then joined McKinsey and build and scaled their supply chain practice. And what I saw over the years is a lot of companies apparently did not get their planning right. And it feels like technology does evolve, but still there’s some gaps to be closed. And this is what I want to talk about today. So looking into the reasons and then also looking into what will change with having now AI having gen AI, agentic AI and so on. So if we think about the first paradox, we have companies investing a lot of money in planning systems.

So in McKinsey, we did a survey and we asked, “Hey, what planning systems do you use? What APS system do you use? And how do you use those?” And the result was a little bit kind of expected, but also frustrating. We got back that 80% of the companies use Excel as their planning system. So why is that? So if we think about the different steps that we have in planning systems, then everything started with spreadsheets. So it’s kind of the generation one, so to say. Then we had the APS system, the advanced planning systems. And here we had, let’s say, knowledge graphs, complex systems to be used. We then moved over because we realized that the world is not deterministic, the world is stohastic. So we moved over to probabilistic planning. We even used AI machine learning to, for example, do predictive demand planning or predictive maintenance.

But it feels like the knowledge of the planner was not really used. And that is still what we often see that if you talk to experienced planners, they say, yes, the planning system, they can help us. So there’s more data I need to handle and so on. But at the end, there is still some gaps and that gap, I need to close myself. And that is where I need to put on my hands where I need to change the numbers, where I need to update the numbers. And unfortunately, very often they are right to do so. So if you think about what happens with supply chains, there’s a couple of things that happen and all of us know that. We have more complexity. We have a lot of disruptions these days. We have complex global supply chains to manage and this is getting harder and harder and harder.

We have more data. Everyone talks about real-time data available, but that basically means also your desk is flooded with data and you still then need to decide. We have exception management. That sounds good in theory. Reality basically tells us that if you open your SAP system, you have 500 exceptions every morning and you cannot really walk through that. So we lack the priorities and more and more, honestly, we also lack the experience of the planner. So what we see here on the right-hand side, the paradox is that we have more complexity, so we would need to have more experience to plan that, but the experience planners, they retire. They leave. So the knowledge, the knowledge to plan, the knowledge to apply this knowledge does leave our companies. And that is really a big challenge. If you think about what we are missing, I call it the experiential ontology or the experiential knowledge.

So we have the system set up to know the data. We know the parameters, but the planner knows exactly what’s going on. So here on that illustration on the right-hand side, to give you an example, the planner knows exactly that in France, France is closed in August because they’re all on summer vacation. So if we want to get a delivery in August, we need to make sure that this is already sent in July. So the planner might know that, “Hey, my customer is always optimizing their revenue towards the end of the quarter.” So there is some behavior and this behavior is not put into the system. The behavior sits with the planner. The planner basically then overrides the proposals and the system does not really help because the system gives the deterministic answer, but does not include the knowledge and the experience of the planner.

And this is clearly something that we need to work on. So if we think about how that might then turn out, it’s shown here. If we think about how GenAI will change the way we operate on the left-hand side, we basically see the system that I just described. So we have our ERP backbone, we have our planning system, our APS has some optimization modules, and so on and so on. And then we have the user or the planner on top. And the planner talks to the system and does not understand the result. So for example, the system has a great machine learning algorithm and tells the planner, “Hey, in three months for this SKU, the expected demand is 54.” And the planner is, “Hey, why is it 54? Why not 50? Why not 60?” And the planning system cannot really explain. So there is basically it’s giving back the answer, but as we don’t understand the algorithms, we need to trust.

And as there’s lack of trust, we see a lot of planners then manually correcting or even doing some site calculation in Excel and with this fixing it. Now think about what AI, what GenAI helps us to do on the right-hand side. We still have the same, let’s say, basic architecture. We have the ERP system, we have the planning system, and we have the planner. Now the planner asks the same question, ask the same question in natural language. And the question goes, “Hey, why is it 54 in three months?” And then the GenAI layer translates this into the planning system’s algorithms and then basically gets as a feedback, oh yes, we had different influencing factors. Look, last year we had this demand, but this year the vacation period is different. We have different weather, we have different sports events. And this is where we now can explain why we see in three months why we see 54.

So this is one thing that we can explain. And then the planning system can also provide much wider context. So what we will see is that with all of these new systems, with all of these new approaches, we will now finally be able to solve our supply chain problems. Think about your typical setup. You probably have something like 60, 70% of your problems is solved by a standard system. But then there is always 30, 40, sometimes even more of the cases that are super specific where it even doesn’t make sense to have a standard system solving that. And that is now something that you can solve with your GenAI layer. And with this, you basically go to the users and the planner’s knowledge and then you assess this and then you trace that and then you track the user’s knowledge and then you basically are able to put this into your planning system and finally have this experiential ontology or experiential knowledge captured in your planning layer.

And with this, enhance the system. And if you think about what that means for the planner, it basically means that the planner is augmented because now you can really use the data that you have and now you can really make decisions that bring you forward and with this increase availability, increase service level, and with this increase the performance of the system. So with this, let me hand over to Jeff to take us from here and looking forward to your questions.

Speaker 3:

Great. Thank you so much, Knut. As Knut has shown us why the planner has always been the most important planning system in the room. The software generated the forecast, the planner made a decision and they ended with something important. But now AI can act as that interface between the planner and the black box translating questions into system logic and system logic back to something a human can act on. That is a real and meaningful step forward. But I want to push us one level deeper because opening the black box is not the same as deciding what comes out of it. The interface tells you what the system recommends. It does not tell you which recommendations become a decision. Who is authorized to make that call or what happens when two recommendations conflict? That gap between a recommendation and an actual decision is where most AI initiatives fall out and closing it requires something most organizations have not yet built.

Let me forward this. So let me make this concrete. Let’s talk about two orders that arrive at the same time in emergency order. A customer needs product immediately and a service commitment is at risk. And the second one is a standard replenishment order. Routine scheduled part of the normal cycle, both valid, both competing for the same inventory. Without a governing policy, what actually happens? Well, someone decides what to do. Maybe the right person, maybe not. It could be inconsistent, undocumented, and impossible to scale. With decision governance in place, the organization has already answered three questions before the conflict ever happens. What cannot fail? Which policy governs this situation and what trade-off is acceptable? That structure is what allows AI to act, not because AI is guessing, but because the rules have been defined and AI is executing them consistently. And governance does not mean governing everything equally.

The simplest test for whether a decision, excuse me, as a candidate for AI is candidate for AI automation is reversibility. Can this decision be undone if it’s wrong?

So this is GAINS’ rapid framework, the risk autonomy principle for intelligent deployment of agents. I know it’s a mouthful, but the governance statement in the concept is basically what is the appropriate level of agent autonomy is determined by the reversibility of the decision it serves. So think about reversible decisions permit automation. Irreversible decisions require human authority. Everything in between is an active governance zone where one organization navigates deliberately as confidence, systems, trust and deployment maturity develops. So let’s give an example, make it a little bit more concrete. Think about a fast-moving low value skew that has a very short lead time is probably a really good candidate for automation. Why is that? Well, if I get that replenishment decision wrong this week, I can correct it next week with minimal impact. It’s basically reversible, but compare that the same type of decision that I’m making for replenishment to a slow moving high value skew with a very long lead time, maybe that I only order once or twice a year.

If that decision is wrong, the consequences sit on the balance sheet for months. Same decision category, fundamentally different governance requirements. Network design takes this one step further. Think about that I could be using agents to help automate individual steps within that process, data validation, model build, scenario setup, but I would never automate the entire network recommendation itself. That distinction between automating steps inside a process and automating the decision the process produces. Rapid governs the decision, not the workflow. And that differentiation is worth noting. Most platforms govern agent autonomy by the workflow position where that agent sits in a hierarchy of the decision-making process. Rapid governs by decision consequence what the decision costs of it’s wrong. That fundamentally a different governance philosophy.

So let’s talk about a process here of the path of moving from AI-assisted planning to AI-driven decisions in a deliberate progression through three stages. And each one has to be built before the next one is possible, at least in our perspective. So step one is what can I do with this technology in improving planning data? And you might reasonably ask, haven’t we always needed accurate data? And the answer is yes, but the stakes have changed when AI starts to act on it. In traditional planning systems, a bad parameter produces a bad recommendation. And thankfully we had the planner available to catch it, to override it, and then we move on and the error is somewhat contained. In an AI system, the executing those decisions autonomously should we move to that level, a bad parameter doesn’t produce one bad recommendation. It produces the same bad decision consistently and at scale before anyone notices.

The feedback loop a planner used to provide the correction, the judgment called, the override may no longer be there. So data quality is not just providing more accurate recommendations anymore. It’s a governance prerequisite. You cannot safely hand authority to a system you cannot trust. And if we move into step two, it’s a big part of what Newton had been talking about with the GenAI, is how we assist planners. And once we have better inputs, AI can compress the time from a recommendation to a confident decision, surfacing the exceptions that matter, explaining the trade-offs in plain language. The planner is still in control. AI is doing the cognitive heavy lifting, and this is where that Gen AI interface lives in practice, but we’re still not at the agentic part. And so step three is we now think about agents. So step three is autonomous decisions with a human in the loop.

AI begins to act executing routine decisions within guardrails the organization has defined, not independently, not without oversight, thereby reducing the burden on the planner to review, modify, and aprove. The sequence is not optional. Governance has to become, excuse me, governance has to come before automation. Organizations that jump directly to step three before building the foundation at steps one and two end up with automation the organization doesn’t trust and that’s worse than no automation at all. So when we think about things at GAINS and how we think about everything we build, if an agent does the move the needle on one of these outcomes, it shouldn’t be built. And just because you can deploy AI doesn’t mean that you should. The question we ask about every agent in the framework that we’ll discuss in a moment is the same question we ask about everything we build.

How does that make supply chain performance measurably better? So these seven outcomes aren’t a marketing list, they’re a filter. And many of these are probably familiar to you. And for the sake of time, I won’t go around the wheel and talk about all of them. But at the end of the day, some of these are moving the business needle on working capital, service levels, and operating costs. But as important when we start putting Agentic technology in place is are we increasing planner productivity? Are we increasing decision velocity? And how have we improved the trust and adoption of the recommendations that then actually turn into decisions? So as we think about the world today, we’ve kind of made this classification scheme of seven different areas where we believe agents can actually be applied. And I’m not going to dive deep into these today for the sake of time, but I wanted just to provide a very high level overview where data intelligence makes sure the inputs can be trusted before any decision gets made.

Analytics and insights is a category identifies what’s driving performance and where the gaps are. Decision support improves the quality of the recommendations itself. Performance management is a crucial area or the closing the loop and providing feedback on tracking those outcomes, measuring plan versus actual variances and driving continuous improvement. And then many others that are in here. But at the end of the day, what’s really important is that every one of these connects back to the outcome that we had talked about on the previous slide. If an agent in any of these domains can’t point to a measurable performance improvement, it really doesn’t belong here. So let me talk a little bit about if you take this three-stage journey, what we’ve seen our customers achieve and some of those benefits that they’ve actually come up with. So selectively in terms of improving inputs, talking about a machine learning application to lead time prediction.

And while many people have focused on demand prediction and improving the quality of forecast accuracy on the demand side, GAINS started its journey on improving the accuracy of lead time prediction, probably one of the most undervalued inputs in supply chain planning. So instead of using static averages, AI learns from supplier history, delivery patterns, order size effects, and real-time signals to produce a probabilistic range of likely outcomes. This connects directly back to what Knut described. Without having this capability, what used to happen is the planner would know that they should be adding a week every time because they know that the system didn’t take care of it. And when this tribal knowledge gets into the system and stays there, then we have a situation and a problem that makes it very difficult for us to automate. By doing this lead time prediction capability that was machine learning back at one of our customers, the results are a 65% lead time accuracy improvement that turned into that in and of itself is not the value.

The value is the 18% reduction in lost sales and $21 million in inventory reduction. So these are not technology metrics, but they’re business outcomes. So there’s some examples in here. I know that we’re running a little short on time, but I do want to touch a little bit on this final thing on autonomous decision-making. So this is the application of where we’re applying supply decision automation to that replenishment planning situation. So here we’re talking about machine learning have the ability to study patterns and how your most experienced buyers are actually making decisions, which orders they approve, which they modify, which signals lead them to an override recommendation. It learns that decision logic including judgment calls on complex orders and executes it automatically within defined governance guardrails. And as you can see, there’s a significant improvement of what rules-based kind of things that can get stale that are programmatically coded in, that if we move to a machine learning environment, this customer moved to a situation where they had 95% of their order lines never being touched or reviewed any longer and over 80% of their orders.

So for the sake of time, I’m not going to go through my summary and I’m going to move forward and I’m going to give this and hand this back over to Bob.

Speaker 1:

Thank you very much, Jeff. And thank you, Knut, both of you for that great presentation helping to understand how AI can really play a role in planning and where it’s good, where it’s not. That’s some really good information. I appreciate that. So time for me to have the privilege of asking you guys some more questions in our panel discussion. And when we’re done with that, we will turn to the audience and field your questions. And even as we are discussing now, audience, you should continue to be submitting your questions by clicking on that Q&A icon at the btom of your screen and we’ll get to you as soon as we can. All right, let’s start out with a question for new. You opened up with what you call the planning paradox. 30 years of investment in planning systems and experienced planners are still overriding them.

Why do these overrides happen? And why are planners often right to make them?

Speaker 2:

That’s a very good question. And if we think about the practical application, we always say, why do you do that? Why don’t you just stick with a system? Because the system was expensive, it was implemented over years and so on, and it’s meant to take or propose the right decisions. If you think about what the systems were designed to, it was basically you have these complex algorithms. You have basically a deterministic system and that does the netting and so on and so on. And then it recommends something to do. What is not in here is the context or knowledge. So if you think about, and this is what I also mentioned in my presentation, there’s a lot of experience. There’s a lot of experience of the planner that does not find its way into the planning system. So if there’s some lead time, for example, that is not reliable.

So Jeff mentioned the lead time, learning the lead time, but there might still be something where the planner knows, “Hey, in Q4, I’m always ordering more with this supplier and then I reach capacity. And if I reach capacity, that means that they need longer. So if I know that, then I order earlier. So basically to make it short, it’s basically the knowledge of the planner that models this much more complex system. And with this, I would say, enhances the planning system that is only looking at deterministic numbers.

Speaker 1:

Okay, thank you. Jeff, you titled your section decision engineering. Now, what do you think is the difference between a planning system that generates recommendations and one that actually makes decisions?

Speaker 3:

Well, a recommendation is an output. And to differentiate a decision is an output with authority behind it. So if we think about the reality that we’re asking our systems now to minimize if we’re moving towards more autonomous thinking, then it’s not just a recommendation I have to take that action. And so a decision is an output with authority behind it. Someone or something that is accountable for the action operating under a defined policy with clear rules for when to act and when to escalate. And so planning systems have always been good at generating the output. What most have never built is the governance layer that converts it into a decision. Who is authorized to act on it? Which policies apply when recommendations conflict? What oversight is required before that action is executed in this gap exactly where most AI initiatives stall. You can deploy a more sophisticated recommendation engine and the organization is still manually converting every output into a decision.

And so we talked about early on, Knut talked about things started from spreadsheets, then we went to these more advanced systems, and we’re still seeing lots of overrides. And if we’re going to really leverage and take advantage of AI, we’ve got to get into this situation where we’re incorporating in that governance that we have expected planners to take care of that have been outside the system to put it inside the system. And that’s what we mean when we talk about decision engineering.

Speaker 1:

That’s quite a challenge. Thank you. Knut, you were describing three forces driving the breakdown of traditional planning, more complexity, more volatility, fewer experienced planners. Which one should supply chain leaders be most focused on right now?

Speaker 2:

So I would say this is the next paradox because if you talk to supply chain leaders, they say, oh, volatility is increasing, complexity’s increasing. Not a lot mentioned the drain of experience planners. And for me, this is the most important one because here that is something where you really, really lose the experience, you lose basically the understanding of your business, of your suppliers and your customers. If you think about complexity and volatility, that is something that you need to work on. You can solve it with increasing inventory, being more agile. Here you can really implement technology to be faster. But if you don’t have the experience in your planning team, then you have the technology at hand, but you cannot use it in the right way. So this is something that is really frightening. And if we think about the next generation, that is even more frightening because what we see is that there’s a lot of argument that how AI will now help to automate all the basic tasks, the basic tasks of the planner.

But if you think about how a junior planner is onboarded, it’s not so much on these basic tasks. It’s much more on building the network, meeting the colleagues, talking to the colleagues, understanding what to do in exceptional situations, and then learn and build this knowledge. If you don’t build this, then this is a real challenge. I just recently had a conversation with a COO of a consumer goods company here in Europe producing a lot of natural products. And she called it a gherkin king who basically knows everything about production capacity and so on. They tried to replace him or put a data scientist at his site and it was just impossible to replace him. So this is something that companies really need to work on to capture the knowledge of their experienced planners.

Speaker 1:

AI has been trying to do it for decades Decades. I mean, one of the original manifestations of it was called expert systems, an attempt to replicate that and we

Speaker 2:

Can

Speaker 1:

See that they have not yet reached that ideal in many ways. Okay, Jeff, you were offering us a direct test for AI autonomy and decision making. The test was, can this decision be undone and what is the financial impact? So with that in mind, how should supply chain leaders app that test in practice?

Speaker 3:

Yeah, I think from my perspective, the test is somewhat straightforward to app. So for any decision you’re considering to handle and hand off to AI, we should ask two questions. First is if it turns out to be wrong, can I correct it in my next planning cycle, whatever that next planning cycle is without significant disruption? Second, where does the exposure land if I’m wrong? Is it financial? Is it service? Is it inventory positioned capacity or some combination? And the more of those impacts that come out, the less it’s a candidate for optimization. So if the answer to the first question is yes and the exposure is contained, then that decision is a candidate for automation today. And I gave the example of the fast-moving skew versus the slow-moving skew because there’s impacts that are associated if I get those answers wrong. But other examples where you would think about where do I not apply it?

So of a supplier contract, a network restructuring, a major capacity commitment. Again, it’s this notion of how wrong can I be? How irreversible is that decision? And so if I’m wrong on any of those, the exposure can run deep across all those dimensions across service, cost capacity. And I may have to live with that decision that was made by me automatically by the system for months, if not years. And so those are not great candidates. And so I think at the end of the day, what this approach and this methodology we’ve created of Rapid, it’s not saying that AI can’t touch complex decisions. That’s not the output of this. It says that AI cannot own the irreversible outcome-based decisions. It doesn’t mean in my decision-making process that I can’t automate steps inside of it. So when we think about part of what GAINS does in network design, I mean AI agents are being built right now to help with tasks like data validation and scenarios set up and model building.

But the actual decision of how should I reconfigure my network, I’m not going to automate that. I’m still going to have a significant amount of human interaction that’s involved with that process. So it sounds simple. We’ve been doing some work with some of our customers to create approaches and methodologies that will ultimately be codified in the system. But I think the core question is how wrong can I be and what’s the impact of being wrong? And the smaller that impact is of being wrong, the easier it is to think about automating some piece of the process or the ultimate decision itself that may come out of it.

Speaker 1:

And do you expect that guidance to remain going forward for the foreseeable future or do you expect AI to get more accurate and better at making decisions to the point we can trust it more with these so – called irreversible decisions?

Speaker 3:

No, I don’t think it will. I mean, it will become more of a partner and assistant in doing parts of the task as we go along, but I mean the whole concept is because it is irreversible, I’m not going to turn it over. So we have ascribed, I mean I believe and I’m sitting here representing GAINS, but we don’t believe that the ultimate destination is the self-autonomous supply chain where everything will be automated and running. There’s an appropriate place to apply it and there’s some places that aren’t in certain decisions are irreversible. They have an impact if I get that wrong. If I put my manufacturing plant in the wrong place, I’m stuck for a very long period of time. And I don’t think that AI will ever help us make that decision, but it’ll help support us making that decision.

Speaker 1:

Humans always in the loop. Almost always in the loop. That’s great. Thank you. Knut, your slide, it drew a clear distinction between what planning systems know and what experienced planners know. Can you give the audience one concrete example of that gap in practice?

Speaker 2:

Happy to do so, but let me first support Jeff in the statement of the human in the loop or Bob, also you. I feel this is super important and I hear a lot of presentations at conferences where people talk about autonomous supply chains. And it goes back like 10 years, 15 years where this idea came up. And honestly, I still don’t see a single autonomous supply chain. It’s very, very clear that we can automate as Jeff described, but there are so many exceptions. And especially now in the time where we have so much tensions in global supply chains that we need to make sure that we always have the human in the loop. And this is also kind of the human in the loop, that is what we can learn from. So to answer your question on an example, we talked already about the lead time example with French being off and we talked about the capacity example where I order in Q4.

But let me give you another example which basically is in both directions. So what we often have, we all do our statistical forecast and then to prepare our demand review meeting, we ask sales and marketing to provide input. And in an ideal case, they have this feeling for the customer. So they know exactly, oh, here it’s increased by 10%, here it’s decreased by 5% and so on. And then we go into a healthy discussion. And this feeling, if you then check what was the forecast value ad, if you then understand what was the pattern recognition of these sales colleague to learn this is super, super important. At the same time, it might also be that the sales colleague always goes 30% up and the reason is not feeling. The reason is that the sales colleague does not trust supply chain to deliver enough or to put enough on allocation.

So this is also something that we need to learn. And here comes the other element that I described in my future architecture that you have GenAI, so to say, as a coach so that you understand what’s going on, that you also understand the soft factors. And with this, you can work on the soft factors and improve the capabilities of people and with this be, so to say, or operate a better supply chain.

Speaker 1:

Right, thank you. So Jeff, three-step roadmap, improve data, assist planners, autonomous decisions. All that’s clear in sequence. Where though do most organizations get stuck?

Speaker 3:

I think most organizations get stuck, excuse me, in the gap between step two and step three. And it’s really that it’s both a governance gap and a technology gap. And the two are connected. Improving planning data has a clear ROI, excuse me, ROI. I’m having trouble speaking now. That’s all right.

Speaker 1:

You’re human.

Speaker 3:

That’s easy to justify. And then assisting planners with better information, richer explanations like we’ve seen through our DEO AI assistant. That our gen AI that’s inside of the solution that allows planners to interact. They’re getting richer explanations of why the system is recommending what it’s recommending and better context to make confidence decisions is well understood. Although sometimes very hard to actually quantify the benefit. But really where it is, that’s extending automation. So when we try to get beyond rule-based, which we’ve been doing for years and years and years that become static and stale and we’re trying to be more dynamic and we’re trying to avoid overrides from planners and moving to that more and more autonomous decision, I think that’s where we’re getting stuck. And the reason we’re getting stuck is this governance layer. We’ve not established guardrails that make the automation safe to deploy.

We’re not answering the organizational design questions that see it underneath it. I’m going to go back to who approves what? Which policies govern when objectives conflict? What happens when AI and the planner disagrees? So can’t automate as much as we think we should be able to. Even as much as I’ve proposed how simple it is to do inside of this framework, at the end of the day, we still have to work on putting that in place and doing that is tough. I mean that gap of trying to understand many organizations that I’ve seen don’t even understand what their guardrails are. I mean, they don’t understand what are the decisions that they’re making and why they’re making them. And even if the planners are overriding them, it’s like tribal institutional knowledge. And so taking all of that knowledge, whether it be right or wrong, do I want to continue to maintain what all my planners are doing or some subset of them and translate that into governance that allows the system to then respond to how to act?

I think that’s where organizations are getting stuck. And for me, that’s one of the biggest challenges. I think step one and step two, valuable, important. And I think we know as a community how to get after those. I think the real challenge in front of us is how to actually focus on that governance layer and how to incorporate it into systems.

Speaker 1:

Great. So Knut, your final slide shows GenAI serving as the interface between the planner and the planning system so – called black box. What changes when that gap is closed?

Speaker 2:

So we will have a lot of happy planners around there. We will have our supply chain problems solved. Finally, finally, as I explained in the presentation, I feel that 30 to 40% of all the planners are so, sorry, the problems are so specific that they are solved by the planner or by Excel. So that is clearly something that we will see. Very important, what changes is that we will have value impact. And this is also what I feel a lot of companies still get wrong because they’re always looking for cost reduction. When I talk to my clients, it’s always the same question. If I implement all of these new cool things, this AI stuff, how much planners do I not need? And this will be part of the business case, but the bigger part of the business case is we create value. Why do we create value?

Because we augment the planner. They come to much better decisions. With this, we increase the availability, we increase the service level. And with this, we are able to increase revenue and we can increase margin. So we have happier customer or happier consumer. And with this, our supply chain is also much better. And that is for me, the core. So to summarize it, we finally solve our supply chain problems. We solve the 40% that is not yet solved with GenAI, with us vibe coding, and we have value impact.

Speaker 1:

Okay. Jeff, you showed us that rapid slide, risk autonomy principle for intelligent deployment maps AI autonomy to decision reversibility. So how do organizations use that framework to make practical deployment decisions? Jeff, you are muted.

Speaker 3:

Well, that’s not good. Okay. So we covered how the reversibility test worked earlier in the question you asked me. And what I wanted to ad is how rapid function as an organizational tool for making deployment decisions over time, not just at the point of the initial deployment of the solution technology. Most organizations don’t cross this governance zone of going from inform to automate in one pass. So how do we build in trust? They move through confidence steps, starting with humans reviewing every AI-driven action, then auditing the outcomes in agregate, and then ultimately extending autonomy as the track record builds. And so Rapid provides this concept of a progression that makes the distinction and makes the distance between where you are today to where you’re permitted to go explicit. But the more important point is what governs placement in the first place. We think about this and where we think it needs to move is many software providers are thinking about platforms assigning agent authority by a position in the workflow.

And that thinks about workflow automation. And we try to think about this as more of decision automation. And so this concept of under Rapid, the same agent can have a high autonomy ceiling when one context is applicable and one is actually too low. So let me just give you a really quick, fast example. Take a demand sensing agent when it’s flagging a short-term promotional spike that may be very contained, correctable in the next forecast cycle and high autonomy is appropriate. But if that same agent has detected a significant shift in your demand, there’s a functional fundamental uplift in demand that’s actually occurred. And if I actually want to accept that future forecast signal, if I’m wrong, the consequences could be major because now I’m going to have significant changes in inventory capacity, supplier commitments. So when we think about the framework to make practical decisions, we just keep on coming back to understand at the end of the day the impact that your decision that you’ve now automated because I don’t have a planner there to catch me anymore.

I don’t have that expertise that’s in the organization, that tribal knowledge because I put it all in the system. We’ve got to make sure that that makes sense that I can leave that open for the system to truly automatically deploy a decision without someone doing interface.

Speaker 1:

Well, listen guys, I’m having a great time hogging the question role here, but I think we really do need to bring the audience in. We have some audience questions. Audience continue to submit your questions while we’re answering what has already come in and we’ll get to as many of them as we can time permitting. Let me start with this one. Questioner says you close, and I think they might be talking about Knut here. You close by saying the real opportunity isn’t just preserving planner knowledge, but using that knowledge to engineer how decisions get made at scale. So what does that shift look like in practice, Knut?

Speaker 2:

Yeah, that’s a very good question. Thanks a lot for bringing this up. So what we discussed is we need to have the AI to observe. We need to understand in the background, we need to observe what the planners decide. So there’s a lot of stuff that is solved deterministically as we discussed, but then there’s this overrides. And here we need to understand why. So my vision is basically that we have, so to say, kind of a supply chain avatar listening in the back, observing, and then always asking, “Hey planner, you now changed this number. Why did you do that? ” And what you then have with this is basically two things. One is that you learn. And on the other hand with the salesperson overriding the sales forecast I just gave is also that you can enter coaching because you can then see that, “Hey, here you did override, but that did not make sense.

Why did you do that? ” So you can immediately coach the planner, the salesperson, the procurement colleague. Now, if we do this, then we build, so to say, a digital twin of our organization. We then have really the experiential ontology in our GenAI system. And this we can then leverage for our young colleagues. We can have the young colleagues experiencing disruptions, so to say, in this model and reacting to that. And with this learning, learning very, very fast. And it’s also very clear that if we implement all of this, it’s intellectually very stimulating because we then don’t do the routine work. We always need to take decisions, we need to prepare decisions. And that might be hard for some people to be honest, but I think it’s also super interesting to be in supply chain. To do that, I feel the most important element is we need to be curious.

We need to stay curious. We need to experiment with these new opportunities and we need to stay open to apply those. So curiosity is key. Yeah.

Speaker 1:

Questioner says, is this planning paradox we talked about, is it unique to supply chain? Or do you see the same gap between expert judgment and systems in other industries? Knut, I think maybe you are also a good one for that one.

Speaker 2:

So I would say it is applicable in a lot of different functions. I would say supply chain is probably the most complex because if you think about what we do, we work end-to-end. So we talk to salespeople and marketing people. We have our finance colleagues talking to us. We have our operations colleagues talking to us. So we have everyone speaking to us. So it’s more in supply chain, but the basic idea also applies in other functions. So fully agree.

Speaker 1:

Right. This questioner says, what does a bad AI deployment in supply chain actually look like and how common is it? Jeff, you want to take that one?

Speaker 3:

Sure. I think the most common failure mode is automating a process rather than governing a decision. I think there’s been too many examples of where we have not looked back at what is our understanding of how our processes work today and how decisions are actually being made, and we’ve just decided to automate what we’re already doing. And so the organization takes that workflow. It could be a purchase order approval, a replenishment trigger, and automates the steps, but they never really define the decision workflow it’s supposed to serve. What the objective? Why were the constraints supplied? Sometimes we just go back to, I’ve always done it this way. Now I just want to take that planner out of the loop and I want to actually automate everything. But I think that’s a mistake and that’s where deployments and failures actually happen.

Speaker 1:

Good. Great. Questioner says, “What happens to a planning operation when that experience judgment walks out the door and nothing replaces it? ” Again, this whole idea of the tribal knowledge loss of the human expert. Knut, what do you think?

Speaker 2:

So there’s clearly a risk for a disaster because if you don’t have the experience, you need to stick to the system, to the deterministic system that does not capture everything. So that means that by design, there is a couple of decisions that are either not taken or wrong taken. So it’s a big risk. And honestly, I mentioned this consumer good example already earlier, that is what we see in practice, that people are retiring and a lot of companies do not necessarily have a plan to capture their knowledge and to codify their knowledge and to have something like a supply chain arbor trial listening in and recording what they do. So it’s a real risk and we will see a lot of companies going forward where the missing planner leads to panic. We will see a lot of war rooms being set up because that then leads to backlog and service issues.

And then we go back and solve it by hand. I can only encourage everyone to not do that and start early, to check who is your most experienced planners? What do they know? When do they retire? Also when they go on a sick leave, what happens? Are you still able to survive or not?

Speaker 1:

Yeah. And war rooms of course are reactive by their very nature and you don’t want to have to wait till the disaster happens before you start addressing it.

Speaker 2:

Yes, exactly.

Speaker 1:

Okay. So the questioner says, for a planner sitting in the middle of this transformation today, what does moving from step to step actually feel like? How do they know they’ve genuinely progressed? Jeff?

Speaker 3:

That’s a great question. If we go back to that three step that I talked about step one often feels like the technology is catching up possibly what the planners already knew. But when they see that data’s getting cleaner, lead times are getting more accurate, they can trust them more and forecasts stop being wrong in predictable ways. The planners actually notice. So they can have more trust and reliability in the system, but they’re still essentially doing the same job. The work hasn’t changed, it’s just more reliable at the end of the day. And step two is where something genuinely I think shifts. The first time a planner asks a question to the system in plain language and gets back a trustworthy explanation in plain language, it’s kind of eye-opening. I know personally from my interactions of just using things like ChatGPT and using Claude and what we’re putting in the system, it’s amazing how well explanations could be generated and content could be created almost on the fly.

And so again, we’re on this trust journey so that I can move to that last step. But step three is really harder to feel in real time. How do I know when I’m in it? I think you know when you’re in it when you start reviewing fewer and fewer exceptions and you’re actually spending your time really on the outliers. So instead of having your laundry list of things that you are aproving, reviewing, and overriding, when you see all that shrinking and you’re really in a smaller subset of items that you’ve got to focus on, that’s when you actually know that you’ve generally progressed through that final step or all three steps, if you will.

Speaker 1:

Okay. Maybe we can get a quick answer to this one. The questioner, is there a particular area of supply chain planning where the demand, inventory, procurement, logistics, et cetera, where AI autonomy is advancing fastest and where is it moving slowest? Jeff, you want to give us a brief answer? I try. Which could probably take about an hour.

Speaker 3:

It’s a little challenging to do, but I think what I’ve seen and what we’ve seen is an inventory replenishment is probably the fastest moving. And it’s that these decisions are frequent. They could be well-bounded, especially in my fast moving small lead, short lead time skews. Those are really ripe in areas as we’ve talked about in rapid to be correct. I’d say that demand planning is advancing quickly. I’m more on the assistance side. So I don’t know that we’re getting all the way to just accepting blindly automatically what are the results of a forecasting and demand planning system, but the ability for assistance to come out of that better than it ever was before. And I think the one that’s moving the slowest broadly is probably procurement and probably for a good reason. I mean, things with supplier selection, contract negotiations, strategic sourcing decisions, those go back to the concept that we proposed in this conversation about irreversibility.

And therefore we’re guiding them better, but we’re not to the point that we’re automating them.

Speaker 1:

Makes a lot of sense. Okay. Well, unfortunately, I am the bearer of bad news. We are just about out of time with this fascinating discussion. We do indeed have time for one final question. And Jeff, I’m going to pitch it to you. So Newt opened today for us by revealing something that seems obvious in hindsight. The most sophisticated planning system in most organizations wasn’t the software. It was the experienced planner absorbing complexity and making judgment calls the system could not see. Now you’ve laid out the engineering path to change that. And as we close here today, what do you want every supply chain leader in this audience to take away and act on?

Speaker 3:

Well, I think Knut did a great job of giving us the diagnosis. The planner has always been the real decision engine and absorbing complexity, compensating for what the system couldn’t see, converting recommendations into decisions through judgment that was never captured, never codified, never scaled. That’s a profound insight and it reframes what we’ve been building towards for 30 years at GAINS. The challenge is now turning that individual judgment into organizational capability. And that’s not a technology problem. The technology to some extent is ready, but the question every supply chain leader in this audience needs to answer is the one we’ve been circling around throughout this conversation. What decisions I’m willing to let AI make? What decisions require human authority? Always. How do I build a governance framework that scales the distinction consistently across my enterprise? The companies that will lead in the AI era are the ones with not the ones with the most algorithms.

They’re the ones that have done the governance work. They know which decisions are reversible and can be automated. They know which decisions are irreversible and require human judgment always. And they have built the decision rights, the escalation paths, and the oversight mechanisms that make it safe to give AI authority where it belongs and keep humans in authority or keep them in the loop where it matters. That’s what it means to us when we say engineer decisions rather than just generate recommendations. And I think that’s the work that’s in front of all of us.

Speaker 1:

Great guidelines, great guidance to make a success out of AI implementations and planning and the supply chain generally. Thank you so much both of you, Jeff and Knut, for your fantastic presentation. Audience also, thank you for your participation and your great questions. You have a chance to read a full article by Knut here. The planner was the system. You can scan the code, that QR code you’re looking at right now if you have time in the next couple of seconds to pull out your camera, but if you don’t, not a problem. That will be sent to all of you attendees upon the conclusion of this webinar. One more time. Thank you, Jeff. Thank you, Knut. Thank you audience. It’s been a great presentation to everybody. Have a great day. All