Table 2 Revealed SPs and management...

Table 2

Revealed SPs and management strategies for AI adoption in S&OP

Revealed SP	Meaning of SP	Exemplary scenario	Exemplary quotes by S&OP informants	Validated in
Performing paradox	Short-term performance vs long-term development
SP1: Attaining short-term efficiency of using existing tools vs. awaiting fully developed AI maturity	The tension between relying on quicker and readily available S&OP tools and methods (e.g. Excel for demand planning, ERP for inventory management, PowerPoint for executive presentations) and allowing sufficient time for a fully mature and integrated AI system that enhances predictive capabilities for demand forecasting, inventory optimization and production planning	An S&OP team continues using Excel to manually consolidate demand forecasts for the next quarter, while simultaneously investing in machine learning algorithms that can analyze years of data to produce more accurate long-term forecasts	I2: “I think AI’s support for S&OP professionals will take time to show good results because, as you know, machine learning algorithms require time to mature and generate accurate data. This is a problem [for S&OP professionals], and I guess it is hard to ask them to wait and put a lot of time.” I3: “We are always running out of time to achieve our sales targets, so we have to use existing methods such as analyzing current RFQs and entering their data into the system … So, waiting for AI to be developed and ready to create a market analysis will take time and will not help in achieving our current targets, especially that it [AI] doesn’t collect data specific to our projects.” I8: “I can’t really see a difference between [traditional] statistical forecasting and AI-driven forecasting. If you add market intelligence into the AI model, like data showing the automotive industry is booming from external sources, it could be beneficial for us. Once AI matures to integrate both internal and external data, it will be very useful. But as long as it relies solely on internal inputs, I’m skeptical there will be a significant improvement over our current methods.”	D1, D2, D3, D4, D11, D17
AI experts’ reflections	E1: “Preparing AI for forecasting takes a really long time. You have to consider at least 3–5 months ahead to be able to provide the customer with what they want with the AI model.” E2: “I fully recognize this paradox. AI maturity takes time, and at the start, the models are often not as good as experienced humans because there’s so much data and knowledge that need to be fed into the system. It doesn’t happen overnight … But the AI itself is mature enough to make its own decisions in a well-defined environment where all the data and constraints are clear.” E3: “I don’t think the issue is with AI itself—I believe AI is mature enough. It’s more about taking small steps and not making the process too big initially.”
AI experts’ suggestions	E1: “As long as they [S&OP professionals] are well ahead in their planning efforts, AI will be useful. But if they are much of ad-hoc executors who change their plans rapidly based on highly fluctuating customer demands, then maybe AI would not be that much of use.” E2: “AI doesn’t get better if you wait; it gets better by using it. Do it on a very small scale. Do it in shadow runs. Make your first analysis, but don’t immediately act on them—just gain experience. Without that, it [AI] could immediately disrupt your whole business. Start using it with small impact, then slowly build the accuracy. Build the trust, and then do a cutover, one by one, on different components.” E2: “I know a semiconductor company in a very dynamic market that uses market performance indicators in its models to predict demand fluctuations over time. These aren’t visible in immediate demand forecasts, but they’re factored into overall forecasting. This isn’t unique to semiconductors; many sectors also rely on external data. Technically, it’s no harder than using internal data—external sources are often already curated.” E3: “You should take small steps and not make it too big at the start. Focus on solving short-term efficiency first, and then build up to address more long-term problems.”
SP2: Relying on gut-feeling decisions vs. indisputably following AI-generated insights	The tension between relying on experienced planners’ intuition for S&OP decisions (e.g. adjusting forecasts based on market knowledge, setting safety stock levels based on experience) and indisputably following AI-generated insights over the long term (e.g. accepting AI-generated demand forecasts automatically, implementing AI-suggested inventory policies over the long term without periodic review of their effectiveness)	During an S&OP meeting, a senior demand planner suggests increasing the forecast for a product based on their market intuition, while an AI system recommends decreasing it based on its analysis of point-of-sale data, social media sentiment, and economic indicators	I3: “If you are expecting formal procedures according to well-planned criteria, which is expected by AI analytics, this will not work in our current situation as decision makers usually make their decisions not on well-planned bases, rather intuition.” I6: “At the moment everything is based on the gut feeling of some people, and if there are some machine learning tools that look at data piles and say OK, your gut feeling is right, then it [AI integration] would be a good idea.” I8: “I feel AI could be riskier than gut feelings for forecasting because gut feelings involve human consciousness, while AI lacks that.”	D1, D3, D5, D6
AI experts’ reflections	E1: “That’s a known paradox. I’ve seen it and experienced it when using AI in my personal life. It comes down to that AI is a black box—you don’t know what knowledge, experience, or data it’s drawing on to give you answers, or if those answers are even correct. So if I’m an S&OP planner and those AI recommendations constantly go against my gut feeling, I’ll start to see the model as, so to speak, useless—right?” E2: “You need the tribal knowledge. S&OP is a hard field to fully automate with AI because a lot of knowledge is in people’s heads. That’s why, at the start, AI models are often worse; they only get better once you start bringing the gut feelings into the model.” E3: “I think this is a spectrum. Solely relying on gut feeling is wrong, but solely following everything AI throws at you is also wrong.”
AI experts’ suggestions	E1: “The people with the gut feeling are the ones best able to adjust the model. Typically, I say, rely on your gut feeling because you have the experience, and let AI provide extra insights. AI can either support your intuition or suggest, for example, ordering new capacity two months earlier—but this only works if the model can explain why.” E1: “If you need predictions on future demand and supply, you have to test and ensure those predictions are accurate. We run ‘what-if’ scenarios to see which choices the model makes to identify the best ones. This helps us determine if the supply chain process or capacity is more efficient than relying on gut feelings. We validate machine learning outcomes before production by using historical test data to make and verify predictions. This is standard practice for AI.” E2: “Always have someone review the results to check if any recent gut-feelings are missing from the model. It’s a genuine problem and challenge for these models, but it is solvable.” E3: “Experienced supply chain professionals need to use their intuition to check if something feels off with AI’s recommendations.”
SP3: Reactive (firefighting) vs. proactive (AI-driven) demand-supply balancing	The tension between reactively addressing S&OP issues as they arise (e.g. expediting orders to meet unexpected demand, adjusting production schedules to address material shortages) and using AI to develop proactive, forward-looking strategies (e.g. analyzing social media trends and point-of-sale data using AI algorithms to anticipate product demand, utilizing predictive maintenance algorithms to prevent equipment failures)	An S&OP team traditionally adjusts production plans monthly in response to actual demand. They are now implementing an AI-driven demand sensing tool that can detect demand shifts in near-real-time and automatically suggest production plan adjustments daily	I2: “We are now trying to predict safety time on the inbound material – trying to predict that based on historical data to set the safety time more correctly.” I3: “Usually, once our customers request certain materials, they want them now with immediate delivery. AI’s algorithms will not be able to build a clear process based on customer sudden demands.” I6: “You may have a supplier in a particular region that’s more likely to experience those interruptions like wars, and we might not spot it, whereas machine learning could predict these interruptions for you and say ‘hey, don’t you know this?’ – this will be fun.”	D1, D7, D8, D9
AI experts’ reflections	E1: “I agree with this one because in a reactive approach, the demand party often doesn’t recognize their needs until it’s too late. They usually realize they need more resources, like extra capacity, only when it becomes critical. But that’s not what S&OP is about, right? That’s more of a tactical issue … Quick reactions to unexpected changes fall outside the typical role of AI in S&OP.” E3: “Here’s the situation: companies often don’t have the luxury to consider new solutions because they’re constantly dealing with last-minute demands. They’re always just about managing to keep up with customer needs.”
AI experts’ suggestions	E1: “We are currently introducing an AI-based solution for a client, but the first step is to mature the S&OP process. Right now, their process is still in a ‘firefighting’ mode. That’s why we’re starting with a solution to bring transparency and observability. It allows users to see their current consumption, making them aware of the costs they incur and the wait times they face when they don’t plan or forecast far enough ahead.” E2: “You might be able to reduce firefighting by looking at how much firefighting you typically have and incorporating these into your [AI] scenarios.” E3: “You need to start by identifying common disruptions and preparing for those. AI can help by analyzing potential issues before they become critical, which doesn’t mean all reactive work will disappear, but it can significantly lessen the impact of unexpected problems.”
Organizing paradox	Strict structures vs. flexibility to adapt
SP4: Centralized vs. decentralized AI implementation	The tension between implementing AI for S&OP through a centralized, company-wide approach (e.g. a single AI-driven demand forecasting system for several product categories and regions) and enabling individual units to develop their own most suitable/specific AI solutions (e.g. region-specific demand forecasting models, product-specific inventory optimization algorithms)	The corporate S&OP team develops a centralized AI-driven demand forecasting system for all product categories. However, the regional teams argue that they need to develop their own AI models that account for local market nuances and unique product behaviors in their regions	I3: “Overall, in developed countries, we have specific plans that are working, and this will help AI to build its algorithms. However, in developing countries we don’t have well-developed plans … The inconsistencies in S&OP plan maturity between developed and developing countries will not allow AI to be implemented company-wide, especially for international corporations like us.” I7: “There’s a central department that is working on AI who has offered some support to us, but we’re still not sure what we need.”	D1, D3, D10, D11, D12
AI experts’ reflections	E1: “From a platform perspective, AI would be centralized because it needs a lot of compute power and data—and data access must be governed centrally. A major challenge for AI is ensuring access is limited to authorized users, and this is crucial for cybersecurity. For instance, we need to make sure the AI model doesn’t disclose sensitive data, like sales figures or HR information, to unauthorized users. That’s why centralized AI governance is important.” E2: “You have two ways to start, right? Either you break the process up into regional pieces and go all the way, or you keep the whole thing but keep it superficial. If you’re not taking regional factors into account, then you end up with an inaccurate model that can’t work effectively across all cases.”
AI experts’ suggestions	E1: “My conclusion is that we should definitely take into account regional settings and influences, while the model itself runs centrally with central access to a lot of enterprise data sources that are being governed.” E2: “If you immediately take the big central approach, it’s going to take a very long time before you get things accurate, so adoption will go very slow. Do it on a smaller scale, it will not be 100% accurate because you only include the regional part and not the bigger scale, but at least you can achieve something.” E3: “It depends on what you want to achieve with the model. For example, the headquarters of a big company with 20 factories worldwide can create an overall plan. They might say ‘this particular factory in this part of the world needs to produce this many products per month.’ That becomes the demand target for that factory. Then the factory can have its own S&OP AI program that takes these headquarters figures as input and determines ‘Hey, how do we plan to meet this target?’”
SP5: Maintaining traditional S&OP processes vs. redesigning toward AI-based S&OP frameworks	The tension between adhering rigidly to traditional S&OP processes (e.g. monthly planning cycles, static Excel-based reports) and redesigning processes entirely to fully leverage AI capabilities (e.g. continuous planning with real-time updates, dynamic dashboards with AI-generated insights and recommendations)	A company maintains its traditional monthly S&OP cycle with set meetings and reports. They are considering moving to a continuous planning process enabled by AI, with real-time updates and exception-based meetings, but this would require a complete redesign of their S&OP process	I2: “If you think of how AI can totally transform the way we run S&OP, you can imagine a complete redesign over time into continuous planning processes with real-time updates and exception-based meetings.” I3: “Although redesigning for AI will make our work easier, I am concerned that AI adjustments might complicate things in some cases, especially when trying to find a specific file or material that was manually entered into Excel.” I4: “It obviously depends on how it [the AI system] is set up and how many parameters you put into it and how.”	D1, D13, D14, D16
AI experts’ reflections	E1: “I don’t see AI completely taking over the orchestration of the entire process anytime soon, where people say I trust AI blindly to order $20 million worth of new equipment because it’s needed for a forecast that AI has seen for the next two months. Somebody still needs to be the brains to check.” E2: “I’ve not seen a single company yet that’s fully moved away from traditional S&OP and built everything around AI. I know how you could do it, but I’ve not seen it in practice yet.” E3: “Yeah, that’s maybe just their lack of understanding about what AI does. They think AI hides their Excel sheets or takes them away, like suddenly you don’t need that Excel file anymore, so it’s gone. But AI will never just delete your Excel file.”
AI experts’ suggestions	E1: “Maintain the traditional process because the AI model must learn to work with and support the existing setup. That’s where all the data is available. If you change the process too soon, there will be no data for the new process. So make sure that the traditional process keeps working but works better.” E2: “For the AI model, you don’t start with 100% accuracy—you can get close, maybe 90% or so, but it requires ongoing training, testing, and validation. It’s similar to people; most decisions people make aren’t 100% accurate either. That’s something important to keep in mind.” E3: “AI can function in parallel. I don’t have to delete the system to implement the solution I have. It operates through a different portal. What we often do is set up a database because we don’t want people having their own Excel files going off on their own journey, which can’t be tracked.”
SP6: Relying on data silos vs. utilizing integrated AI-driven data systems	The tension between maintaining separate data sources for different S&OP functions (e.g. sales data in CRM, inventory data in WMS, production data in MES) and transitioning to integrated, AI-driven data systems (e.g. a unified data lake with machine learning models that analyze cross-functional data for holistic S&OP insights)	Currently, a company’s demand data is stored in the sales system, inventory data in the warehouse management system, and production data in the ERP system. They are considering creating a unified data lake to enable AI-driven S&OP, but this would require substantial changes to data management processes	I1: “The main challenge with AI is data consistency. We need to ensure no false signals are presented due to biased reporting … Often, data appears to have changed because of how it’s reported, not because of actual changes.” I4: “There are differences in each dataset that you can handle with machine learning. For load and capacity planning, there is a big mountain of data that you could utilize. But at the moment, we don’t use this big mountain of data, we only use a small part of it to perform our calculations and planning.” I8: “I think AI might be a bit biased towards one function over another. If the data is collected from sales information, the algorithm will be biased towards sales, not towards production or production management.”	D1, D10, D12
AI experts’ reflections	E1: “You say it’s biased, I say it’s not. It just has too little information across the end-to-end process. Therefore, the model cannot do anything other than learn from the sales or forecasting processes, but not from the backend manufacturing process, because the model has no other data to learn from.” E2: “But now you get organizational politics into the mix. The AI itself doesn’t influence or change anything. It’s people who do that. So you have to agree on the basis of your data—whether you put it in a silo or on a platform, that’s your choice.” E2: “The source of truth should remain the same, right? So even if you bring data to a data lake, there’s a difference between just copy-pasting data from your ERP vs. changing it into reports and using that as your basis.” E3: “Taking the AI side, we always say, ‘garbage in, garbage out.’ So if what goes into the AI isn’t controlled well, the output will be unreliable.”
AI experts’ suggestions	E1: “Successful AI solutions should be able to quickly integrate and learn from new datasets that are essential for the overall process.” E2: “The silos just make it impractical. You rather need a single layer where everything is combined without processing—no transformations, no cleaning, just a basic layer. This works because pulling data from different silos is unmanageable otherwise.” E3: “For AI to really make a difference, you need multiple data sources, and they need to be connected. You need a centralized database for the entire company—one that people can access based on their permissions. This becomes your single source of truth. That’s really the only way.”
Belonging paradox	Unified organizational identity vs. individual autonomy
SP7: Collective AI-driven S&OP identity vs. individualistic S&OP identities	The tension between establishing a collective, forward-looking identity via standardized AI-driven S&OP processes (e.g. deploying uniform AI-based models for demand forecasting that adjust inventory levels automatically across all regions) and preserving diverse, traditional planning identities (e.g. continuing to use local market knowledge and manual demand assessments in regions where relationships and/or individuals dictate business dynamics)	A global company implements a standardized AI-driven S&OP platform to be used by all regions. However, the European team argues that their unique market requires a more relationship-based planning approach, while the Asian team prefers a more hierarchical decision-making process that the AI system does not accommodate	I3: “If we have a fully integrated AI system, we will not know who is in charge for what or from where this data has been collected.” I3: “For optimal results, you need function-wise AI integration, otherwise errors will arise.” I7: “There could be some cultural, technical, or organizational factors at play if you want to implement AI for more than one location. Different sites might have unique attitudes about how to input data in a way that allows a central department to evaluate it effectively and provide benefits. Although we are all part of ‘X’ [the company], individuals at different locations might prefer to handle things in their own way. This often depends on the mindset of the people involved.”	D1, D2, D4, D15
AI experts’ reflections	E1: “We often see equipment data used differently across regions even when producing the same product—it carries different meanings. So, looking at those identities—regional vs. central—is one thing, but also the identity of each function. Sales thinks differently from manufacturing, and this way of thinking is really hard to capture in an AI model.” E2: “It’s about the AI model being a black box and needing transparency on how it reaches its conclusions … If you remove AI and run a full S&OP process by combining demand with production, don’t you still have the same issue? It all depends on who provides the data and whether that data is trustworthy.”
AI experts’ suggestions	E1: “You might train AI to ‘think like a sales expert’ with knowledge of products and upcoming releases, but the model needs all the data and solid training in sales processes. It’s like telling it to think like a five-year-old describing their favorite food—teaching AI to think like sales, forecasting, or production experts takes focused training and data.” E2: “To build trust, you’ll need to make the process transparent and involve end users and organizational leaders to agree on how it’s done. The same applies even without AI; if a controversial analysis is presented, the CEO will still ask, ‘Who provided that number?’” E3: “I think it’s about finding a balance. You really need to set all your priorities straight and keep company goals in mind. Based on those goals, you have priorities, and you should be able to provide that input.”
SP8: Streamlining AI-driven collaboration vs. preserving specialized S&OP expertise	The tension between promoting AI-enabled cross-functional collaboration in S&OP (e.g. using AI to generate a single consensus forecast) and maintaining distinct functional roles and specialized expertise (e.g. sales maintaining their own forecast based on customer relationships, finance keeping separate financial projections)	A new AI-driven S&OP system provides a single, data-driven forecast that it suggests should be used by all functions. However, the sales team insists on maintaining their own forecast based on customer insights, while the finance team wants to keep their separate financial forecasting process	I2: “I believe that building such a platform [AI-assisted system that combines all forecasts from sales, finance, and production] is very possible. But again, it is not the system that is the issue here. Sales want their forecasts to be dictating because they are closer to the market and customers. We know from experience that they [sales and marketing] prefer to resolve their issues in isolation. They have their own informal meetings for market assessments and changing that takes away this cultural aspect from them.” I8: “Production planning might be more efficient with AI due to its predictable nature and reliance on internal data and parameters. But this is not the case with forecasting, who by nature rely on future forecasts with external data sources that are hard to feed into the AI algorithm.”	D1, D3, D12
AI experts’ reflections	E1: “I think specialized expertise will always be needed. You can’t just capture all this expertise in datasets and feed it into a model that has to learn—it’s unlikely to happen anytime soon.” E2: “It really depends on the industry. In some industries, sales are easy to predict, and you can achieve high accuracy. In others, it’s much harder. The same goes for production—if the process is simple, you can get high accuracy, but in high-tech industries, it may vary a lot.” E3: “It depends on what you want from AI. You might say, ‘Hey, I don’t want AI to handle forecasting or predict future demand—I’ll use it solely for production planning.’ Or, ‘I don’t want it to consider production factors; I just want it to look at specific variables and give me a forecast or multiple forecast scenarios.’”
AI experts’ suggestions	E1: “But I do think AI models can help improve communication and collaboration, making sure everyone in the end-to-end process is better informed about product status, past releases, and what went wrong. You can simply ask the AI, and it will provide answers based on available data for everyone involved.” E3: “Ideally, AI should combine everything—it should evolve to bring together all these different perspectives. For example, if I’m using AI for forecasting, it might tell me, ‘Hey, this intel comes from production—do you need to collaborate more closely with the production team?’”
SP9: Resisting vs. embracing AI-driven S&OP	The tension between embracing AI in S&OP to streamline tasks (e.g. automating routine forecasting and inventory management tasks) and resisting the adoption of AI in S&OP in fear of losing jobs or reduced perceived value within the organization (e.g. demand planners concerned about AI replacing their role in forecast generation)	In an S&OP initiative, a firm integrates AI to automate forecasting and inventory planning, raising concerns among demand and supply planners about AI diminishing their roles and threatening their job security	I2: “We talked to several people [concerning AI in S&OP], and yes, we saw some variations in opinions. I would not say a specific department is more negative about it than the other … Some [S&OP professionals] seem to have the wrong idea about what AI is, and that it will make them do little and look less important. That is a possibility for resistance I guess.” I3: “Back-office employees might risk their positions with AI integration but front office workers who meet clients frequently would be very happy to utilize AI as it will result in conducting more meetings and reducing paperwork.”	D1, D10, D14, D16
AI experts’ reflections	E1: “If you look at how much manual work is happening in the S&OP process—both on demand forecast and supply side—which can be automated? Yeah, then some people will lose their jobs.” E2: “Yes. Currently we have a project where one guy is responsible for most of the demand forecasting process and he’s putting up a lot of resistance because he sees that ‘before I was the guy who did all the demand forecasts, and now 80% of that is going to happen through the tool—or maybe the tool will even show that my own forecasts were quite bad sometimes.’” E3: “In most cases AI won’t take your job. It will change your job. I think that’s what people are really scared of. Also, if you automate everything and something goes wrong, you can’t blame anyone anymore.”
AI experts’ suggestions	E1: “Right now, firms are spending a lot more money on additional work, like having people build dashboards or gather data in one place, just to ensure there are some insights for better planning in S&OP meetings. But these are extra investments they don’t want to make. So I want to turn it around—it’s not about people losing their jobs, but about saving money on additional hires needed for this work.” E2: “You have people who are enthusiastic from the start because they like technology, and you have those concerned for valid reasons. Then you have people who are concerned because they’re just worried for their position. The last are irrational views, at least from a company perspective.”
Learning paradox	Exploiting existing knowledge vs. exploring new skills
SP10: Developing traditional S&OP skills vs. acquiring AI-related competencies	The tension between developing traditional S&OP skills (e.g. cross-functional communication, business acumen, Excel-based analysis) and acquiring new competencies more focused on AI (e.g. data science skills, machine learning model development and interpretation, AI ethics understanding)	An S&OP team with years of experience in consensus building and cross-functional communication is now being asked to learn data science skills to develop and interpret machine learning models for demand forecasting, requiring a momentous shift in their skill set	I2: “The goal of our two AI projects in S&OP is to be data-driven and reduce reliance on human error. For example, a person might interpret market dynamics as indicating more customer orders are coming, while the data suggests the opposite. We should either train our salespeople on how to assess trends accurately, or we should let the data speak for itself through a trained model. Perhaps we need some kind of combination.” I3: “I need our new employees to understand the history of our work as it is today. At the same time, I want them to learn how to use AI to optimize performance and efficiency.” I8: “Not every supply chain person knows how to use Python [for developing AI algorithms], as its typically made for developers. However, if AI interfaces are made user-friendly for supply chain purposes, it could work. Currently, I’m not aware of any AI models that are specifically designed for S&OP functions.”	D1, D10, D11, D16
AI experts’ reflections	E2: “You have people with AI skills to prepare the models, run simulations, handle the data, and understand what’s happening in the back end. And then you still need people with traditional S&OP skills who do the balancing and negotiations.” E3: “When you hire a new S&OP professional, you want, on one end, to teach them how to do things traditionally. On the other, you need to teach them the latest, including how to work with AI and how to inform their decisions using AI.”
AI experts’ suggestions	E1: “S&OP people using AI don’t have to be AI experts. They don’t need to understand the model’s inner workings or how data is accessed. It’s more about being good users—knowing how to prompt the AI effectively and adjust the guardrails so it provides the best answers for their process.” E2: “You need to train them [S&OP professional] on how to work with the AI models, but they don’t have to become data scientists. I don’t think you need S&OP people to use Python a lot. If you have good models with good interfaces, you get IT people who handle that in the backend, and you get S&OP people who are savvy enough to work with those tools and understand what the model does and how to interpret it.” E3: “Traditional S&OP is very important for decision-making, and AI should really focus on having the data ready, giving you the correct scenarios, and putting you in the driver’s seat. It shouldn’t be about making S&OP people experts in programming.”
SP11: Relying on human learning vs. leveraging machine learning	The tension between relying on acquired human learning from S&OP experience in isolation (e.g. understanding of market dynamics, knowledge of product lifecycles) and leveraging easily accessed machine learning algorithms for continuous improvement (e.g. AI systems that automatically detect and adapt to changing demand patterns or supply chain disruptions)	Inventory planners have traditionally adjusted safety stock levels based on their experience with stockouts and excess inventory. The company is now implementing a reinforcement learning algorithm that continually adjusts safety stock levels based on various factors, potentially outperforming human decision-making in this area	I3: “Human learning will always result in creative proposals. At the same time, machine learning will produce well-organized proposals that might not be as creative as those generated through human learning.” I5: “Machine learning eventually requires accurate data input from us, humans. If you fill in data that shows what you want rather than the reality, then you end up with useless calculations from these machines.” I8: “AI can give superpowers to normal people … However, whether it’s AI-based forecasting or [traditional] statistical forecasting, you still need human expertise and interaction, especially for negotiations with the production unit. This is something AI will never be able to replace.”	D1, D3, D4, D5
AI experts’ reflections	E1: “I don’t think AI will ever have gut feelings—it will always react to the data it’s been given. If certain data hasn’t been fed into the model, AI will never be able to produce anything that depends on that missing data.” E3: “Data always has to come from somewhere. So, in the end, it should always come from a person—or at least be supervised by a human. Otherwise, it’s just random, generated, or not real.”
AI experts’ suggestions	E2: “You will still rely on human expertise for some part of it. And you can put that into a model, but new human expertise will pop up that you don’t have in the model, and there’s a cycle of continuously adding it, even though you’ll never reach 100% accuracy.” E2: “AI will always have boundaries; it performs the assignment you give it. At the same time, humans also have biases. If you look at a game of chess, you’ll see models doing things no human has ever thought of, so there’s definitely creativity there. But it’s always within the scope you set. I’ve never seen a model think outside the rules you give it.”
SP12: Making incremental improvements vs. embracing disruptive AI innovation	The tension between making small, continuous improvements to existing S&OP processes (e.g. refining statistical forecasting methods, gradually improving inventory policies) and embracing disruptive AI innovations (e.g. implementing autonomous planning systems that dynamically adjust forecasts, inventory levels, and production schedules based on real-time data and complex algorithms)	An S&OP team has been gradually refining their ABC classification for inventory management over years. They are now considering implementing an AI system that dynamically classifies SKUs based on multiple factors including demand volatility, supply risk, and strategic importance, fundamentally changing how inventory policies are set	I2: “We had the idea and option to choose between going all in with AI integration, but we preferred to narrow it down to our project only. We know that only smaller steps are realistic and can be approved, if you know what I mean.” I3: “Well-established companies like us cannot fully adopt disruptive technologies as this will affect their annual figures, but slow changes will be smooth and will not affect annual turnover.” I6: “Currently, a lot of our processes are just informal rules noted in Excel sheets or kept in mind. It seems these could be partially automated. My colleague, who is responsible for the new [AI] tool, mentioned that we might implement AI features slowly later on.”	D1, D9, D11, D13
AI experts’ reflections	E1: “That [incremental improvement] is really the most important part for getting customers to adopt AI solutions, especially given it’s a black box. Looking at machine learning examples, you can see accuracy gradually improving from 80% to 95%. And this improvement depends on how much time and effort you invest in learning and retraining the models.” E1: “If you change the process too early, then you won’t have any data for the new process—there’s nothing for the AI model to learn from.” E2: “If you’re discussing using AI for preparing the S&OP process, like predicting your demands, forecasting supply, measuring them, or running simulations, understand that these aren’t quick, short-term actions. They’re part of a longer-term strategy.”
AI experts’ suggestions	E2: “Yeah, start with incremental, smaller steps. There’s a lower risk if things going wrong. It’s a safer way to begin, especially when there’s uncertainty about whether it will work or how much budget it will require. It’s a clear, easy way to get started.” E2: “We put all your knowledge into the system, and then, if we’re lucky, we can gradually automate 40% of the easier decisions, leaving you with the harder ones. Over time, this might reach 50% or 60%, and we’ll see how far we can go. We don’t force it on anyone; instead, we let them decide if the model can take over their repetitive tasks or not.” E3: “Baby steps is the only way to do it—like optimizing the recent approach a little bit, or take one small part of it, optimize that, make it a bit more efficient, and then move to the next process. And then, with those steps, you combine maybe a few of them, and slowly you work towards it.”

Revealed SP	Meaning of SP	Exemplary scenario	Exemplary quotes by S&OP informants	Validated in
Performing paradox	Short-term performance vs long-term development
SP1: Attaining short-term efficiency of using existing tools vs. awaiting fully developed AI maturity	The tension between relying on quicker and readily available S&OP tools and methods (e.g. Excel for demand planning, ERP for inventory management, PowerPoint for executive presentations) and allowing sufficient time for a fully mature and integrated AI system that enhances predictive capabilities for demand forecasting, inventory optimization and production planning	An S&OP team continues using Excel to manually consolidate demand forecasts for the next quarter, while simultaneously investing in machine learning algorithms that can analyze years of data to produce more accurate long-term forecasts	I2: “I think AI’s support for S&OP professionals will take time to show good results because, as you know, machine learning algorithms require time to mature and generate accurate data. This is a problem [for S&OP professionals], and I guess it is hard to ask them to wait and put a lot of time.”I3: “We are always running out of time to achieve our sales targets, so we have to use existing methods such as analyzing current RFQs and entering their data into the system … So, waiting for AI to be developed and ready to create a market analysis will take time and will not help in achieving our current targets, especially that it [AI] doesn’t collect data specific to our projects.”I8: “I can’t really see a difference between [traditional] statistical forecasting and AI-driven forecasting. If you add market intelligence into the AI model, like data showing the automotive industry is booming from external sources, it could be beneficial for us. Once AI matures to integrate both internal and external data, it will be very useful. But as long as it relies solely on internal inputs, I’m skeptical there will be a significant improvement over our current methods.”	D1, D2, D3, D4, D11, D17
AI experts’ reflections	E1: “Preparing AI for forecasting takes a really long time. You have to consider at least 3–5 months ahead to be able to provide the customer with what they want with the AI model.”E2: “I fully recognize this paradox. AI maturity takes time, and at the start, the models are often not as good as experienced humans because there’s so much data and knowledge that need to be fed into the system. It doesn’t happen overnight … But the AI itself is mature enough to make its own decisions in a well-defined environment where all the data and constraints are clear.”E3: “I don’t think the issue is with AI itself—I believe AI is mature enough. It’s more about taking small steps and not making the process too big initially.”
AI experts’ suggestions	E1: “As long as they [S&OP professionals] are well ahead in their planning efforts, AI will be useful. But if they are much of ad-hoc executors who change their plans rapidly based on highly fluctuating customer demands, then maybe AI would not be that much of use.”E2: “AI doesn’t get better if you wait; it gets better by using it. Do it on a very small scale. Do it in shadow runs. Make your first analysis, but don’t immediately act on them—just gain experience. Without that, it [AI] could immediately disrupt your whole business. Start using it with small impact, then slowly build the accuracy. Build the trust, and then do a cutover, one by one, on different components.”E2: “I know a semiconductor company in a very dynamic market that uses market performance indicators in its models to predict demand fluctuations over time. These aren’t visible in immediate demand forecasts, but they’re factored into overall forecasting. This isn’t unique to semiconductors; many sectors also rely on external data. Technically, it’s no harder than using internal data—external sources are often already curated.”E3: “You should take small steps and not make it too big at the start. Focus on solving short-term efficiency first, and then build up to address more long-term problems.”
SP2: Relying on gut-feeling decisions vs. indisputably following AI-generated insights	The tension between relying on experienced planners’ intuition for S&OP decisions (e.g. adjusting forecasts based on market knowledge, setting safety stock levels based on experience) and indisputably following AI-generated insights over the long term (e.g. accepting AI-generated demand forecasts automatically, implementing AI-suggested inventory policies over the long term without periodic review of their effectiveness)	During an S&OP meeting, a senior demand planner suggests increasing the forecast for a product based on their market intuition, while an AI system recommends decreasing it based on its analysis of point-of-sale data, social media sentiment, and economic indicators	I3: “If you are expecting formal procedures according to well-planned criteria, which is expected by AI analytics, this will not work in our current situation as decision makers usually make their decisions not on well-planned bases, rather intuition.”I6: “At the moment everything is based on the gut feeling of some people, and if there are some machine learning tools that look at data piles and say OK, your gut feeling is right, then it [AI integration] would be a good idea.”I8: “I feel AI could be riskier than gut feelings for forecasting because gut feelings involve human consciousness, while AI lacks that.”	D1, D3, D5, D6
AI experts’ reflections	E1: “That’s a known paradox. I’ve seen it and experienced it when using AI in my personal life. It comes down to that AI is a black box—you don’t know what knowledge, experience, or data it’s drawing on to give you answers, or if those answers are even correct. So if I’m an S&OP planner and those AI recommendations constantly go against my gut feeling, I’ll start to see the model as, so to speak, useless—right?”E2: “You need the tribal knowledge. S&OP is a hard field to fully automate with AI because a lot of knowledge is in people’s heads. That’s why, at the start, AI models are often worse; they only get better once you start bringing the gut feelings into the model.”E3: “I think this is a spectrum. Solely relying on gut feeling is wrong, but solely following everything AI throws at you is also wrong.”
AI experts’ suggestions	E1: “The people with the gut feeling are the ones best able to adjust the model. Typically, I say, rely on your gut feeling because you have the experience, and let AI provide extra insights. AI can either support your intuition or suggest, for example, ordering new capacity two months earlier—but this only works if the model can explain why.”E1: “If you need predictions on future demand and supply, you have to test and ensure those predictions are accurate. We run ‘what-if’ scenarios to see which choices the model makes to identify the best ones. This helps us determine if the supply chain process or capacity is more efficient than relying on gut feelings. We validate machine learning outcomes before production by using historical test data to make and verify predictions. This is standard practice for AI.”E2: “Always have someone review the results to check if any recent gut-feelings are missing from the model. It’s a genuine problem and challenge for these models, but it is solvable.”E3: “Experienced supply chain professionals need to use their intuition to check if something feels off with AI’s recommendations.”
SP3: Reactive (firefighting) vs. proactive (AI-driven) demand-supply balancing	The tension between reactively addressing S&OP issues as they arise (e.g. expediting orders to meet unexpected demand, adjusting production schedules to address material shortages) and using AI to develop proactive, forward-looking strategies (e.g. analyzing social media trends and point-of-sale data using AI algorithms to anticipate product demand, utilizing predictive maintenance algorithms to prevent equipment failures)	An S&OP team traditionally adjusts production plans monthly in response to actual demand. They are now implementing an AI-driven demand sensing tool that can detect demand shifts in near-real-time and automatically suggest production plan adjustments daily	I2: “We are now trying to predict safety time on the inbound material – trying to predict that based on historical data to set the safety time more correctly.”I3: “Usually, once our customers request certain materials, they want them now with immediate delivery. AI’s algorithms will not be able to build a clear process based on customer sudden demands.”I6: “You may have a supplier in a particular region that’s more likely to experience those interruptions like wars, and we might not spot it, whereas machine learning could predict these interruptions for you and say ‘hey, don’t you know this?’ – this will be fun.”	D1, D7, D8, D9
AI experts’ reflections	E1: “I agree with this one because in a reactive approach, the demand party often doesn’t recognize their needs until it’s too late. They usually realize they need more resources, like extra capacity, only when it becomes critical. But that’s not what S&OP is about, right? That’s more of a tactical issue … Quick reactions to unexpected changes fall outside the typical role of AI in S&OP.”E3: “Here’s the situation: companies often don’t have the luxury to consider new solutions because they’re constantly dealing with last-minute demands. They’re always just about managing to keep up with customer needs.”
AI experts’ suggestions	E1: “We are currently introducing an AI-based solution for a client, but the first step is to mature the S&OP process. Right now, their process is still in a ‘firefighting’ mode. That’s why we’re starting with a solution to bring transparency and observability. It allows users to see their current consumption, making them aware of the costs they incur and the wait times they face when they don’t plan or forecast far enough ahead.”E2: “You might be able to reduce firefighting by looking at how much firefighting you typically have and incorporating these into your [AI] scenarios.”E3: “You need to start by identifying common disruptions and preparing for those. AI can help by analyzing potential issues before they become critical, which doesn’t mean all reactive work will disappear, but it can significantly lessen the impact of unexpected problems.”
Organizing paradox	Strict structures vs. flexibility to adapt
SP4: Centralized vs. decentralized AI implementation	The tension between implementing AI for S&OP through a centralized, company-wide approach (e.g. a single AI-driven demand forecasting system for several product categories and regions) and enabling individual units to develop their own most suitable/specific AI solutions (e.g. region-specific demand forecasting models, product-specific inventory optimization algorithms)	The corporate S&OP team develops a centralized AI-driven demand forecasting system for all product categories. However, the regional teams argue that they need to develop their own AI models that account for local market nuances and unique product behaviors in their regions	I3: “Overall, in developed countries, we have specific plans that are working, and this will help AI to build its algorithms. However, in developing countries we don’t have well-developed plans … The inconsistencies in S&OP plan maturity between developed and developing countries will not allow AI to be implemented company-wide, especially for international corporations like us.”I7: “There’s a central department that is working on AI who has offered some support to us, but we’re still not sure what we need.”	D1, D3, D10, D11, D12
AI experts’ reflections	E1: “From a platform perspective, AI would be centralized because it needs a lot of compute power and data—and data access must be governed centrally. A major challenge for AI is ensuring access is limited to authorized users, and this is crucial for cybersecurity. For instance, we need to make sure the AI model doesn’t disclose sensitive data, like sales figures or HR information, to unauthorized users. That’s why centralized AI governance is important.”E2: “You have two ways to start, right? Either you break the process up into regional pieces and go all the way, or you keep the whole thing but keep it superficial. If you’re not taking regional factors into account, then you end up with an inaccurate model that can’t work effectively across all cases.”
AI experts’ suggestions	E1: “My conclusion is that we should definitely take into account regional settings and influences, while the model itself runs centrally with central access to a lot of enterprise data sources that are being governed.”E2: “If you immediately take the big central approach, it’s going to take a very long time before you get things accurate, so adoption will go very slow. Do it on a smaller scale, it will not be 100% accurate because you only include the regional part and not the bigger scale, but at least you can achieve something.”E3: “It depends on what you want to achieve with the model. For example, the headquarters of a big company with 20 factories worldwide can create an overall plan. They might say ‘this particular factory in this part of the world needs to produce this many products per month.’ That becomes the demand target for that factory. Then the factory can have its own S&OP AI program that takes these headquarters figures as input and determines ‘Hey, how do we plan to meet this target?’”
SP5: Maintaining traditional S&OP processes vs. redesigning toward AI-based S&OP frameworks	The tension between adhering rigidly to traditional S&OP processes (e.g. monthly planning cycles, static Excel-based reports) and redesigning processes entirely to fully leverage AI capabilities (e.g. continuous planning with real-time updates, dynamic dashboards with AI-generated insights and recommendations)	A company maintains its traditional monthly S&OP cycle with set meetings and reports. They are considering moving to a continuous planning process enabled by AI, with real-time updates and exception-based meetings, but this would require a complete redesign of their S&OP process	I2: “If you think of how AI can totally transform the way we run S&OP, you can imagine a complete redesign over time into continuous planning processes with real-time updates and exception-based meetings.”I3: “Although redesigning for AI will make our work easier, I am concerned that AI adjustments might complicate things in some cases, especially when trying to find a specific file or material that was manually entered into Excel.”I4: “It obviously depends on how it [the AI system] is set up and how many parameters you put into it and how.”	D1, D13, D14, D16
AI experts’ reflections	E1: “I don’t see AI completely taking over the orchestration of the entire process anytime soon, where people say I trust AI blindly to order $20 million worth of new equipment because it’s needed for a forecast that AI has seen for the next two months. Somebody still needs to be the brains to check.”E2: “I’ve not seen a single company yet that’s fully moved away from traditional S&OP and built everything around AI. I know how you could do it, but I’ve not seen it in practice yet.”E3: “Yeah, that’s maybe just their lack of understanding about what AI does. They think AI hides their Excel sheets or takes them away, like suddenly you don’t need that Excel file anymore, so it’s gone. But AI will never just delete your Excel file.”
AI experts’ suggestions	E1: “Maintain the traditional process because the AI model must learn to work with and support the existing setup. That’s where all the data is available. If you change the process too soon, there will be no data for the new process. So make sure that the traditional process keeps working but works better.”E2: “For the AI model, you don’t start with 100% accuracy—you can get close, maybe 90% or so, but it requires ongoing training, testing, and validation. It’s similar to people; most decisions people make aren’t 100% accurate either. That’s something important to keep in mind.”E3: “AI can function in parallel. I don’t have to delete the system to implement the solution I have. It operates through a different portal. What we often do is set up a database because we don’t want people having their own Excel files going off on their own journey, which can’t be tracked.”
SP6: Relying on data silos vs. utilizing integrated AI-driven data systems	The tension between maintaining separate data sources for different S&OP functions (e.g. sales data in CRM, inventory data in WMS, production data in MES) and transitioning to integrated, AI-driven data systems (e.g. a unified data lake with machine learning models that analyze cross-functional data for holistic S&OP insights)	Currently, a company’s demand data is stored in the sales system, inventory data in the warehouse management system, and production data in the ERP system. They are considering creating a unified data lake to enable AI-driven S&OP, but this would require substantial changes to data management processes	I1: “The main challenge with AI is data consistency. We need to ensure no false signals are presented due to biased reporting … Often, data appears to have changed because of how it’s reported, not because of actual changes.”I4: “There are differences in each dataset that you can handle with machine learning. For load and capacity planning, there is a big mountain of data that you could utilize. But at the moment, we don’t use this big mountain of data, we only use a small part of it to perform our calculations and planning.”I8: “I think AI might be a bit biased towards one function over another. If the data is collected from sales information, the algorithm will be biased towards sales, not towards production or production management.”	D1, D10, D12
AI experts’ reflections	E1: “You say it’s biased, I say it’s not. It just has too little information across the end-to-end process. Therefore, the model cannot do anything other than learn from the sales or forecasting processes, but not from the backend manufacturing process, because the model has no other data to learn from.”E2: “But now you get organizational politics into the mix. The AI itself doesn’t influence or change anything. It’s people who do that. So you have to agree on the basis of your data—whether you put it in a silo or on a platform, that’s your choice.”E2: “The source of truth should remain the same, right? So even if you bring data to a data lake, there’s a difference between just copy-pasting data from your ERP vs. changing it into reports and using that as your basis.”E3: “Taking the AI side, we always say, ‘garbage in, garbage out.’ So if what goes into the AI isn’t controlled well, the output will be unreliable.”
AI experts’ suggestions	E1: “Successful AI solutions should be able to quickly integrate and learn from new datasets that are essential for the overall process.”E2: “The silos just make it impractical. You rather need a single layer where everything is combined without processing—no transformations, no cleaning, just a basic layer. This works because pulling data from different silos is unmanageable otherwise.”E3: “For AI to really make a difference, you need multiple data sources, and they need to be connected. You need a centralized database for the entire company—one that people can access based on their permissions. This becomes your single source of truth. That’s really the only way.”
Belonging paradox	Unified organizational identity vs. individual autonomy
SP7: Collective AI-driven S&OP identity vs. individualistic S&OP identities	The tension between establishing a collective, forward-looking identity via standardized AI-driven S&OP processes (e.g. deploying uniform AI-based models for demand forecasting that adjust inventory levels automatically across all regions) and preserving diverse, traditional planning identities (e.g. continuing to use local market knowledge and manual demand assessments in regions where relationships and/or individuals dictate business dynamics)	A global company implements a standardized AI-driven S&OP platform to be used by all regions. However, the European team argues that their unique market requires a more relationship-based planning approach, while the Asian team prefers a more hierarchical decision-making process that the AI system does not accommodate	I3: “If we have a fully integrated AI system, we will not know who is in charge for what or from where this data has been collected.”I3: “For optimal results, you need function-wise AI integration, otherwise errors will arise.”I7: “There could be some cultural, technical, or organizational factors at play if you want to implement AI for more than one location. Different sites might have unique attitudes about how to input data in a way that allows a central department to evaluate it effectively and provide benefits. Although we are all part of ‘X’ [the company], individuals at different locations might prefer to handle things in their own way. This often depends on the mindset of the people involved.”	D1, D2, D4, D15
AI experts’ reflections	E1: “We often see equipment data used differently across regions even when producing the same product—it carries different meanings. So, looking at those identities—regional vs. central—is one thing, but also the identity of each function. Sales thinks differently from manufacturing, and this way of thinking is really hard to capture in an AI model.”E2: “It’s about the AI model being a black box and needing transparency on how it reaches its conclusions … If you remove AI and run a full S&OP process by combining demand with production, don’t you still have the same issue? It all depends on who provides the data and whether that data is trustworthy.”
AI experts’ suggestions	E1: “You might train AI to ‘think like a sales expert’ with knowledge of products and upcoming releases, but the model needs all the data and solid training in sales processes. It’s like telling it to think like a five-year-old describing their favorite food—teaching AI to think like sales, forecasting, or production experts takes focused training and data.”E2: “To build trust, you’ll need to make the process transparent and involve end users and organizational leaders to agree on how it’s done. The same applies even without AI; if a controversial analysis is presented, the CEO will still ask, ‘Who provided that number?’”E3: “I think it’s about finding a balance. You really need to set all your priorities straight and keep company goals in mind. Based on those goals, you have priorities, and you should be able to provide that input.”
SP8: Streamlining AI-driven collaboration vs. preserving specialized S&OP expertise	The tension between promoting AI-enabled cross-functional collaboration in S&OP (e.g. using AI to generate a single consensus forecast) and maintaining distinct functional roles and specialized expertise (e.g. sales maintaining their own forecast based on customer relationships, finance keeping separate financial projections)	A new AI-driven S&OP system provides a single, data-driven forecast that it suggests should be used by all functions. However, the sales team insists on maintaining their own forecast based on customer insights, while the finance team wants to keep their separate financial forecasting process	I2: “I believe that building such a platform [AI-assisted system that combines all forecasts from sales, finance, and production] is very possible. But again, it is not the system that is the issue here. Sales want their forecasts to be dictating because they are closer to the market and customers. We know from experience that they [sales and marketing] prefer to resolve their issues in isolation. They have their own informal meetings for market assessments and changing that takes away this cultural aspect from them.”I8: “Production planning might be more efficient with AI due to its predictable nature and reliance on internal data and parameters. But this is not the case with forecasting, who by nature rely on future forecasts with external data sources that are hard to feed into the AI algorithm.”	D1, D3, D12
AI experts’ reflections	E1: “I think specialized expertise will always be needed. You can’t just capture all this expertise in datasets and feed it into a model that has to learn—it’s unlikely to happen anytime soon.”E2: “It really depends on the industry. In some industries, sales are easy to predict, and you can achieve high accuracy. In others, it’s much harder. The same goes for production—if the process is simple, you can get high accuracy, but in high-tech industries, it may vary a lot.”E3: “It depends on what you want from AI. You might say, ‘Hey, I don’t want AI to handle forecasting or predict future demand—I’ll use it solely for production planning.’ Or, ‘I don’t want it to consider production factors; I just want it to look at specific variables and give me a forecast or multiple forecast scenarios.’”
AI experts’ suggestions	E1: “But I do think AI models can help improve communication and collaboration, making sure everyone in the end-to-end process is better informed about product status, past releases, and what went wrong. You can simply ask the AI, and it will provide answers based on available data for everyone involved.”E3: “Ideally, AI should combine everything—it should evolve to bring together all these different perspectives. For example, if I’m using AI for forecasting, it might tell me, ‘Hey, this intel comes from production—do you need to collaborate more closely with the production team?’”
SP9: Resisting vs. embracing AI-driven S&OP	The tension between embracing AI in S&OP to streamline tasks (e.g. automating routine forecasting and inventory management tasks) and resisting the adoption of AI in S&OP in fear of losing jobs or reduced perceived value within the organization (e.g. demand planners concerned about AI replacing their role in forecast generation)	In an S&OP initiative, a firm integrates AI to automate forecasting and inventory planning, raising concerns among demand and supply planners about AI diminishing their roles and threatening their job security	I2: “We talked to several people [concerning AI in S&OP], and yes, we saw some variations in opinions. I would not say a specific department is more negative about it than the other … Some [S&OP professionals] seem to have the wrong idea about what AI is, and that it will make them do little and look less important. That is a possibility for resistance I guess.”I3: “Back-office employees might risk their positions with AI integration but front office workers who meet clients frequently would be very happy to utilize AI as it will result in conducting more meetings and reducing paperwork.”	D1, D10, D14, D16
AI experts’ reflections	E1: “If you look at how much manual work is happening in the S&OP process—both on demand forecast and supply side—which can be automated? Yeah, then some people will lose their jobs.”E2: “Yes. Currently we have a project where one guy is responsible for most of the demand forecasting process and he’s putting up a lot of resistance because he sees that ‘before I was the guy who did all the demand forecasts, and now 80% of that is going to happen through the tool—or maybe the tool will even show that my own forecasts were quite bad sometimes.’”E3: “In most cases AI won’t take your job. It will change your job. I think that’s what people are really scared of. Also, if you automate everything and something goes wrong, you can’t blame anyone anymore.”
AI experts’ suggestions	E1: “Right now, firms are spending a lot more money on additional work, like having people build dashboards or gather data in one place, just to ensure there are some insights for better planning in S&OP meetings. But these are extra investments they don’t want to make. So I want to turn it around—it’s not about people losing their jobs, but about saving money on additional hires needed for this work.”E2: “You have people who are enthusiastic from the start because they like technology, and you have those concerned for valid reasons. Then you have people who are concerned because they’re just worried for their position. The last are irrational views, at least from a company perspective.”
Learning paradox	Exploiting existing knowledge vs. exploring new skills
SP10: Developing traditional S&OP skills vs. acquiring AI-related competencies	The tension between developing traditional S&OP skills (e.g. cross-functional communication, business acumen, Excel-based analysis) and acquiring new competencies more focused on AI (e.g. data science skills, machine learning model development and interpretation, AI ethics understanding)	An S&OP team with years of experience in consensus building and cross-functional communication is now being asked to learn data science skills to develop and interpret machine learning models for demand forecasting, requiring a momentous shift in their skill set	I2: “The goal of our two AI projects in S&OP is to be data-driven and reduce reliance on human error. For example, a person might interpret market dynamics as indicating more customer orders are coming, while the data suggests the opposite. We should either train our salespeople on how to assess trends accurately, or we should let the data speak for itself through a trained model. Perhaps we need some kind of combination.”I3: “I need our new employees to understand the history of our work as it is today. At the same time, I want them to learn how to use AI to optimize performance and efficiency.”I8: “Not every supply chain person knows how to use Python [for developing AI algorithms], as its typically made for developers. However, if AI interfaces are made user-friendly for supply chain purposes, it could work. Currently, I’m not aware of any AI models that are specifically designed for S&OP functions.”	D1, D10, D11, D16
AI experts’ reflections	E2: “You have people with AI skills to prepare the models, run simulations, handle the data, and understand what’s happening in the back end. And then you still need people with traditional S&OP skills who do the balancing and negotiations.”E3: “When you hire a new S&OP professional, you want, on one end, to teach them how to do things traditionally. On the other, you need to teach them the latest, including how to work with AI and how to inform their decisions using AI.”
AI experts’ suggestions	E1: “S&OP people using AI don’t have to be AI experts. They don’t need to understand the model’s inner workings or how data is accessed. It’s more about being good users—knowing how to prompt the AI effectively and adjust the guardrails so it provides the best answers for their process.”E2: “You need to train them [S&OP professional] on how to work with the AI models, but they don’t have to become data scientists. I don’t think you need S&OP people to use Python a lot. If you have good models with good interfaces, you get IT people who handle that in the backend, and you get S&OP people who are savvy enough to work with those tools and understand what the model does and how to interpret it.”E3: “Traditional S&OP is very important for decision-making, and AI should really focus on having the data ready, giving you the correct scenarios, and putting you in the driver’s seat. It shouldn’t be about making S&OP people experts in programming.”
SP11: Relying on human learning vs. leveraging machine learning	The tension between relying on acquired human learning from S&OP experience in isolation (e.g. understanding of market dynamics, knowledge of product lifecycles) and leveraging easily accessed machine learning algorithms for continuous improvement (e.g. AI systems that automatically detect and adapt to changing demand patterns or supply chain disruptions)	Inventory planners have traditionally adjusted safety stock levels based on their experience with stockouts and excess inventory. The company is now implementing a reinforcement learning algorithm that continually adjusts safety stock levels based on various factors, potentially outperforming human decision-making in this area	I3: “Human learning will always result in creative proposals. At the same time, machine learning will produce well-organized proposals that might not be as creative as those generated through human learning.”I5: “Machine learning eventually requires accurate data input from us, humans. If you fill in data that shows what you want rather than the reality, then you end up with useless calculations from these machines.”I8: “AI can give superpowers to normal people … However, whether it’s AI-based forecasting or [traditional] statistical forecasting, you still need human expertise and interaction, especially for negotiations with the production unit. This is something AI will never be able to replace.”	D1, D3, D4, D5
AI experts’ reflections	E1: “I don’t think AI will ever have gut feelings—it will always react to the data it’s been given. If certain data hasn’t been fed into the model, AI will never be able to produce anything that depends on that missing data.”E3: “Data always has to come from somewhere. So, in the end, it should always come from a person—or at least be supervised by a human. Otherwise, it’s just random, generated, or not real.”
AI experts’ suggestions	E2: “You will still rely on human expertise for some part of it. And you can put that into a model, but new human expertise will pop up that you don’t have in the model, and there’s a cycle of continuously adding it, even though you’ll never reach 100% accuracy.”E2: “AI will always have boundaries; it performs the assignment you give it. At the same time, humans also have biases. If you look at a game of chess, you’ll see models doing things no human has ever thought of, so there’s definitely creativity there. But it’s always within the scope you set. I’ve never seen a model think outside the rules you give it.”
SP12: Making incremental improvements vs. embracing disruptive AI innovation	The tension between making small, continuous improvements to existing S&OP processes (e.g. refining statistical forecasting methods, gradually improving inventory policies) and embracing disruptive AI innovations (e.g. implementing autonomous planning systems that dynamically adjust forecasts, inventory levels, and production schedules based on real-time data and complex algorithms)	An S&OP team has been gradually refining their ABC classification for inventory management over years. They are now considering implementing an AI system that dynamically classifies SKUs based on multiple factors including demand volatility, supply risk, and strategic importance, fundamentally changing how inventory policies are set	I2: “We had the idea and option to choose between going all in with AI integration, but we preferred to narrow it down to our project only. We know that only smaller steps are realistic and can be approved, if you know what I mean.”I3: “Well-established companies like us cannot fully adopt disruptive technologies as this will affect their annual figures, but slow changes will be smooth and will not affect annual turnover.”I6: “Currently, a lot of our processes are just informal rules noted in Excel sheets or kept in mind. It seems these could be partially automated. My colleague, who is responsible for the new [AI] tool, mentioned that we might implement AI features slowly later on.”	D1, D9, D11, D13
AI experts’ reflections	E1: “That [incremental improvement] is really the most important part for getting customers to adopt AI solutions, especially given it’s a black box. Looking at machine learning examples, you can see accuracy gradually improving from 80% to 95%. And this improvement depends on how much time and effort you invest in learning and retraining the models.”E1: “If you change the process too early, then you won’t have any data for the new process—there’s nothing for the AI model to learn from.”E2: “If you’re discussing using AI for preparing the S&OP process, like predicting your demands, forecasting supply, measuring them, or running simulations, understand that these aren’t quick, short-term actions. They’re part of a longer-term strategy.”
AI experts’ suggestions	E2: “Yeah, start with incremental, smaller steps. There’s a lower risk if things going wrong. It’s a safer way to begin, especially when there’s uncertainty about whether it will work or how much budget it will require. It’s a clear, easy way to get started.”E2: “We put all your knowledge into the system, and then, if we’re lucky, we can gradually automate 40% of the easier decisions, leaving you with the harder ones. Over time, this might reach 50% or 60%, and we’ll see how far we can go. We don’t force it on anyone; instead, we let them decide if the model can take over their repetitive tasks or not.”E3: “Baby steps is the only way to do it—like optimizing the recent approach a little bit, or take one small part of it, optimize that, make it a bit more efficient, and then move to the next process. And then, with those steps, you combine maybe a few of them, and slowly you work towards it.”

Source(s): Created by authors

[ViewLarge]

Sharing Unavailable