00:00:00 Generative AI impact in supply chain
00:02:06 The reality of promotions in AI
00:03:35 Comparing gen AI’s potential and hype
00:04:56 Positive outlook on AI’s potential
00:06:54 Risks of premature tech adoption
00:08:21 Leveraging AI for curiosity-driven growth
00:10:14 Generative AI as modern smart Wikipedia
00:11:34 Language models offer process insights
00:13:39 Broadening client feedback via LLMs
00:15:37 Evolving user interfaces with natural language
00:17:38 Document generation via LLM analysis
00:19:43 LLMs seen as colleagues shift impact
00:21:37 Future decision-making with digital colleagues
00:26:46 Delegating tasks to AI agents wisely
00:29:02 Efficiency gains in demand planning
00:30:58 Insights on forecast-pricing dynamics
00:32:51 Demand review transcripts enhance maturity
00:35:08 Joannes explores AI’s potential future
00:37:14 Linking meeting discussions challenging for LLMs
00:40:26 Demand review meetings boost performance
00:44:34 AI tools change meeting participation
00:48:49 Communication evolution highlights privacy issues
00:52:39 Tech impacts business secrecy practices
00:55:43 Generative AI explores perception shifts
00:57:00 AI’s complex reality in algorithms vs. GenAI
00:58:10 Generative AI as blockchain successor
00:59:45 Cultural dysfunction causing money wastage
01:00:38 Intuition prevents financial waste from GenAI
01:03:00 Practical supply chain teachings at board level
01:05:39 Intellectual task mechanization in modern era
01:09:02 Perspectives on productivity via automation
01:11:02 Concluding interview with a farewell
Summary
GenAI’s buzz exceeds balance sheets. Short-term gains are modest—clerical acceleration, smarter triage, meeting discipline—with humans in the loop. Long run could rival containerization, if incentives align. Bans merely push usage to phones; guardrails beat prohibitions. The “value gap” indicts procurement theater, not the tech; leaders need mechanical sympathy and proofs, not 600-question RFPs. LLMs don’t learn; context/RAG remain bottlenecks, so curation matters. Board case: mechanize intellectual work or be outpaced. Shop-floor case: less drudgery, better defaults. Optimists say five years; pessimists, twenty. Either way, today’s spreadsheet theater lives on borrowed time.
Extended Summary
Generative AI has generated more buzz than balance sheets. The panel agrees that short-term effects in supply chains are modest but real: clerical acceleration, better document triage, and relief for repetitive work. Over the long run, the change could be as consequential as containerization—if organizations align incentives and expectations. Hype promises “solutions”; reality offers trade-offs.
Two forces pull in opposite directions. On one side, practical gains: language models act as “smart reference” tools, stabilize fragile RPA use cases, extract signal from free-form customer and supplier feedback, and enable conversational dashboarding. They can also serve as a coached, “digital colleague” that nudges meetings toward action and maturity—provided a human remains in the loop. On the other side, technical limits matter: today’s LLMs don’t truly learn; they operate within static parameters and ever-growing context windows that can become a distraction. The so-called “knowledge flywheel” remains an organizational problem disguised as a technical one. RAG helps, but scale and relevance filtering still impose costs.
Policy choices have consequences. Firms that try to ban LLMs will discover workarounds on personal devices, because the time savings are too large to ignore. Sensible guardrails—privacy, data handling, and spend—beat blanket prohibitions that incentivize shadow IT. Transparency, used judiciously, can improve end-to-end performance; secrecy as a reflex often preserves dysfunction more than advantage.
The “GenAI value gap” says less about AI than about procurement rituals. When executives lack “mechanical sympathy” for a technology, they green-light pilots designed to “solve world hunger” and then declare the field a disappointment. The cure is not another buzzword but better governance: proofs-of-concept with real data, measurable outcomes, and a narrative that connects availability to revenue and margin—rather than treating supply chain as a cost silo.
At the board level, the case is simple: the 21st century mechanizes intellectual work as the 20th mechanized physical work; competitors who automate clerical armies will move faster with fewer errors. On the shop floor, the pitch must be practical: tools that remove drudgery, surface better defaults, and make planners more effective—without turning systems into black boxes.
Timelines divide optimists and pessimists. If models improve context handling and organizations learn to curate knowledge, meaningful gains could arrive within five years; if culture and process lag, twenty is more realistic. Either way, the status quo—walls of dashboards, spreadsheet drudgery, and theater-grade selection processes—will not survive contact with compounding efficiency.
Full Transcript
Conor Doherty: Supply chain has been buzzing with talk of generative AI for at least two years. However, in September 2025, the tone has shifted just a little bit. Now people are asking a slightly different question: what difference has it made? And a subsequent question: has that difference been positive or negative?
Now today’s guest, Knut Alicke, joins Joannes and me in studio in Paris to discuss that very issue. Knut is partner emeritus at McKinsey. He’s been teaching supply chain management for over 25 years and, most impressively, he’s really good at the saxophone.
Before we get into the discussion, you know the drill: if you like what we do at Lokad and you want to support us, follow us on LinkedIn and subscribe to the YouTube channel. And with that, I give you today’s conversation with Knut Alicke.
Knut, thank you for joining us. It’s great to have you in studio. I think this is your third appearance on LokadTV.
Knut Alicke: Yes, I think so too, and the first time to be here in person, so really great to be here. And you’re actually the first person to sit on the new Lokad couch— you and Joannes breaking it in, professionally of course.
To get started, Knut, I’m sure a lot of people already know you. You’re a partner emeritus at McKinsey; you’ve taught supply chain science and supply chain management for 25 years. So my first question to set the table: where do you also find time to play the saxophone?
I kind of, going back, really enjoyed playing the saxophone, and I spent my army time in the band. That’s where I got excited to practice much more than the normal students. I try to keep my level; this is practicing during the night or over the weekend and making sure that you always have a band that keeps you busy.
Conor Doherty: Just writing down: you played music in the army. So you’ve lived quite a few lives, in fact. Does that actually influence the way you approach business and supply chain?
Knut Alicke: I don’t think so. I was too young to be influenced by that back then. I was 19, and I just enjoyed one year of practicing from the morning to the night and then enjoying very nice concerts in Hamburg. That’s where I enjoyed going to a lot of jazz concerts and improving.
Conor Doherty: I think we’re all fans of jazz here, but we’re also fans of generative AI, which I think is what brought us together in the studio. Joannes, I will come to you in a moment, but first, Knut, set the table. I’ll pitch you a statement and then a question, and then you can go.
The statement is: the supply chain— and I’d say the world overall— has been buzzing with generative AI for at least, let’s say, two years. It has changed the supply chain landscape. That’s the statement. And the question is: do you agree with that? And if yes, has it made it better or worse?
Knut Alicke: I would say it has not significantly changed yet. What I see is that GenAI will significantly change the way we manage and operate supply chains. I always compare it to the invention of the shipping container: something that changed global flows of goods, made it more efficient, easier, standardized. I see that now with GenAI; we are at a similar point in time.
At the same time, it’s clearly hyped these days. People always overestimate the impact of a new technology in the short term and underestimate it in the long term. If you look into what is possible these days with GenAI— all of us use ChatGPT— and think about what is possible in supply chain to help improve performance and make life of planners much easier, there’s a lot we will see in the next couple of years. But we also need to say it’s early days; it has only been two years, and the models are improving so much. It will change, and I’m sure we will go more into detailed examples later on.
Conor Doherty: Joannes, do you agree with Knut? It’s been for the better or for the worse?
Joannes Vermorel: I would say very clearly for the better, but it is still very small. The small pieces are very good; those tools are extremely useful. Right now we are talking about people using GPT on the side to speed-run something that would have been a very clerical task, and that is very good.
Models have been progressing enormously. The fact that you have bigger context means that you can upload fairly long documents and say, “Please find in this agreement if there is a term specific to this and that.” In the past, scanning a 50-page document would have taken you an hour to find if this part is discussed; here you can have it in half a minute. So it is very much something that is good.
Then we shall see for the future— I agree with your statement about the impact short-term versus long-term. Right now, as we have been teaching, one of the things that are going to be very tricky with this generative AI is that nowadays students can literally completely cheat all their homework. There is no such thing as student homework anymore; ChatGPT will just be doing it. For a professor, it’s almost impossible to sort out, except saying, “I’m going to give a bad grade to all the students who give me a copy without spelling mistakes,” because if there are no spelling mistakes, I’m going to assume it was written by ChatGPT.
Having people who have been potentially miseducated because of this gap may create problems. But for companies, in terms of generative AI, the people who are really taking risks with these technologies are more like the companies who go all in into vibe coding. That’s the sort of thing where it hasn’t even started yet in supply chain.
People are still, my understanding, in very superficial use. The few uses where it’s used, it is really massive low-hanging fruits, quick wins, no question. The more dangerous, elaborate, impacting uses, as far as I understand, have not even started yet.
Knut Alicke: Let me comment on this— a couple of ideas. If we compare 25 years ago when the internet started, there were a lot of companies that would basically block the internet. Their people were not allowed to use the internet because they were worried they would read the news and not do their work or do their private stuff. Nowadays it’s normal and it’s adding a lot of value.
A lot of big companies now also block ChatGPT or other large language models. That is the worst thing you can do. You need to make sure that you educate your people how to use these models. You need to provide an environment where they can use it without uploading secrets and so on. At the end, you want a curious organization that is learning and exploring the opportunities.
Same with students: when I teach, I ask my students, “Please solve this question also with ChatGPT, and then find out where it’s going the right way and where it’s going the wrong way,” to enable and even encourage them to use it. Otherwise, we will never learn what is possible.
Joannes Vermorel: I think companies blocking LLMs are going to rediscover what happened in the early 2000s, where people would just connect to the internet through their mobile device and EDGE connection at the time— just because at some point they would circumvent the limitation because it’s so annoying. People would just do that with their mobile phone. If you block it at the corporate level, people will use their personal account on their smartphone and do it, because for clerical tasks the productivity gains are so great that resisting it is extremely difficult.
It’s very hard to re-justify to someone that this person should spend three hours doing something hypertedious when there is a tool that can do it in five minutes. As soon as this person feels morally justified for this one thing, the temptation is very great to use it for all the other things.
Conor Doherty: This ties back to the short-term/long-term perspective. In the short term, Knut, what do you see as the primary positive impacts that generative AI has had in the supply chain context?
Knut Alicke: What we see a lot— and that is already working— is if you don’t understand something, you use it as a very smart Wikipedia. You look things up and you learn. We see cases where very administrative, repetitive tasks can be addressed with what I’d call a smart robotic process automation. RPA struggled when processes changed a bit; you had to reprogram the process. Here you see early wins.
If I go a bit ahead, the vision I would have for a GenAI application is: typical processes in supply chain are well defined— from an algorithmic point of view clearly defined— but then reality tells a different story. People find ways around it: manual changes of numbers, lack of trust, whatever. The result is not as expected.
Here comes the power of language models: you don’t see in the data the reasons why processes don’t work. You just see forecast accuracy going down. Why is that? Because this person changed the numbers— and you don’t understand why. Why not have a “supply chain avatar,” a digital Joannes, in discussion with this person who changed the number, and then slowly find out what’s going on? The real reason may be they don’t trust the planner; they want to serve their customers; in the past they ran out of stock. Then you fix this by building trust or increasing inventory. This is where the models can add a lot of value.
Conor Doherty: That’s almost like a tool to improve a process indirectly— a discursive interaction: “Why did you do that?” Joannes, are there other examples of plugging LLMs in so they change a process largely unattended?
Joannes Vermorel: Yes. For example, if you think about quality of service: many companies have various flavors of Net Promoter Score. They survey their clientele once a week, once a month; B2C might take a sample. The traditional way is multiple-choice questions— very low resolution. Why do you do that? Because if you do it any other way, you end up with 200–500 free-form answers and it’s hard to do anything with it.
With LLMs, suddenly you don’t have to force your clientele into giving you feedback according to your preconceived boxes. Maybe the client will complain about something you didn’t even realize was a problem: “I received a device with an American plug instead of a British one. I solved it, but it was annoying.” Your checklist “Was the product damaged? Yes/No?” says no— but it’s still a problem.
Traditionally, anything free-form was a pain. Same for suppliers. With LLMs, you can imagine systems where partners give free-form input and LLMs crunch that into statistics without hard assumptions that narrow the problem into neat boxes. That lets you rethink a process quite deeply.
Knut Alicke: To build on this, everything related to creating dashboards. When you implement a new system, a big chunk of time is defining what you want to see, and then it’s hard-coded. Everyone has new ideas. Imagine a world where you talk to your system— your LLM— and say, “I want to see this and this. Please highlight this on the x-axis, this on the y-axis,” and then you see it. If you like it, it goes into standard; if not, you refine.
The user interface will be a natural-language interface where you find what you need. In addition, the system should provide things you missed. You ask for one KPI and another, but you miss service level— super important. Then this digital Joannes could say, “Interesting that you look at these two, but did you also look into service? Did you look into how service correlates with inventory? Is there something cooking?”
Joannes Vermorel: At Lokad, we are looking at the same problem but in a very different way. The typical problem with dashboards is that, very quickly in enterprise settings, we have walls of metrics— tons of numbers. The problem becomes: what am I looking at exactly?
Take lead time measured in days. Is it business days or calendar days? Do we remove outliers? If something was never delivered, does it count as infinite? As a thousand? There are tons of conventions. Our approach is not dynamic composition of a dashboard, but generating extremely detailed documentation on the fly by having the LLM look at all the code that led to this number and compile in English the stuff that matters. What is the scope? What did we filter? How far in the past?
We drown in dashboards and numbers, and the semantics are tough. That’s the battle we’re fighting.
Knut Alicke: Let me throw in another idea I explored with a client: from lean manufacturing we know the Five Whys— or everyone raising kids knows the Five Whys. You ask why something is happening, then why again, until the root cause. That’s super powerful. Building on your walls of KPIs: if something goes wrong, use the LLM to go deeper, deeper, deeper until you really find the reason and where to change a parameter— inventory, for example— to improve performance.
Joannes Vermorel: Absolutely. Again, the way people often perceive any new technology, certainly AI, is as a tool to be used. But the way you talk about it, Knut, frames it like a colleague you interact with— a digital member of the team.
Conor Doherty: Would that be a fair description?
Knut Alicke: Yes, it is. Let’s think about hiring a new colleague from university. He or she arrives and is trained; we have a mentor, a coach. The new colleague first does simple tasks, then more complicated tasks. In the beginning they might decide to buy things for 10 euros; over years it’s 100,000 euros. We develop this colleague. No one expects a new joiner to know everything.
Interestingly, when we implement a planning tool, the planner expects the tool to do magic and know everything. Why not have a GenAI bot as a digital colleague? We also need to train it: business context, specifics about a certain customer that always complains, that we don’t follow “loudest shout first serve,” and so on. We train the model to understand our specific context.
The model comes with enormous speed— the capability of handling and analyzing data not possible for humans. If we combine that with specific knowledge, we have a real digital colleague. I foresee a future where experienced business people talk to this model like a colleague and improve the quality of their decisions significantly. They don’t need to do the boring stuff— copying in and out of Excel— and they’re presented with, “Did you think about this? Did you look into that?” Then, combining it with their experience, they come to a much better decision.
Conor Doherty: I like the metaphor— or simile, rather. If you take the example of a mentor: you don’t pick just anybody. There’s a skill set required to effectively teach. What are the important skills for that mentor to train an LLM? Do you have to be an expert coder or computer scientist?
Knut Alicke: You need to be open, curious, transparent. A good mentor is not only training but also open for feedback— a give and take. I started to train a digital copy of myself— a digital Knut. At some point I was super frustrated because I felt, “This digital Knut does not know anything. It doesn’t know me.” Then I realized: if that were a real colleague, I would also be frustrated but continue to coach and develop. With a digital colleague, my digital self also needs to be developed. Same thing.
Joannes Vermorel: Here we are touching a deep limitation of GenAI: right now LLMs are not learning anything. Technically, what you have is a pre-trained model— you dump a sizable fraction of the internet, Wikipedia and more, into training, and you get a static model. The parameters do not change. ChatGPT has no memory whatsoever; it’s stateless. The only thing you can adjust is the context.
Fortunately, over the last year the contexts have grown enormously. The latest model— for example GPT-5 at the API level— we are talking about a context window of 400,000 tokens. It’s enormous. You can’t use all of it for inputs; you can use, top of my head, something like 270,000 tokens as inputs; the rest is used for reasoning because you need space for the reasoning.
But the tricky thing about the current paradigm is we have models that have a kind of crystallized intelligence, but it’s static. You can contextualize more, but the thing cannot really become more intelligent; it remains as intelligent as day one. You can enrich the context.
Who is going to maintain that context? In technical terms people speak about a knowledge flywheel. Who maintains it? Is it the LLM itself that maintains the flywheel— adding or removing nuggets from its own bank of information? ChatGPT does that if you let it— it will record nuggets about you and re-inject them. But adding too many things in the context becomes a distraction. You can add hundreds of pages of context, but the LLM, to answer any question, has to load this context, and that can worsen performance with irrelevant trivia.
So, to have an LLM agent as a true colleague, you have a super-intelligence in some ways and in others it is extremely stupid because it cannot learn anything ever— at least in the current paradigm.
Conor Doherty: Even taking both statements at face value, there are still tasks you could entrust to this digital colleague and some you keep in human hands. Knut, first: what tasks would you feel comfortable delegating quickly to a digital colleague, and which do you keep people in charge of?
Knut Alicke: I’m not sure I would delegate anything 100%. I would still have the human in the loop— that is super important. To give an example: when we do diagnostics, you collect data, do analysis; in supply chain you then interview people to understand forecasting, demand review, S&OP, and so on. That does not necessarily tell the full truth. It’s like a Gemba walk in manufacturing: in a meeting room, they explain the nice and shiny process; on the shop floor, it looks different.
Here I see GenAI add a lot of value by observing the process. Imagine a demand review meeting. Often the deck is not well prepared, the agenda is not followed, no actions defined. Many people connect with video off, mic off, not contributing; only a few talk. If you have your GenAI bot listening in and providing feedback to the facilitator— “Do this, here something is wrong”— you can also massively parallelize this. If you have 200 demand planners, you can coach all 200: “Look, here you could do better or different.” That’s how you coach people.
Where we see applications today: procurement. Agents can already add value in long-tail spend where you have many small categories or products that are never checked properly due to time and people. Having an agent do the analysis, do comparisons, and trigger a renegotiation of prices— this can add value as a low-hanging fruit today. It’s possible in a relatively unsupervised way with a low budget for the long tail; you wouldn’t do that for high-value items, but for long tail it starts, and then you move into other categories.
Conor Doherty: The example you described— your AI bot supervising meetings— off the air, you explained an experiment simulating exactly that. You said you’d done it like 20 times with different exchanges. Can you explain that again?
Knut Alicke: I created a synthetic dataset— a synthetic company— to experiment with real-like data. I went through analysis: “What’s going on with this forecast? Can you do a forecast as well? What’s going on with pricing? Correlation with promotions?” Then I also created manual inputs: I took two demand planners. One was increasing the forecast; the other was improving the forecast. The classical KPI, forecast value add, for the first was very bad; for the second was good.
I then talked to ChatGPT in voice mode and claimed I was demand planner one because I wanted to understand how the model would react. The model had all the context on the company, SKUs, customers, fluctuations, the positive bias of the first planner. I complained about supply chain people not understanding me, production people never delivering what I want, and that’s why I need to increase the forecast. I overdid it a bit but reflected reality.
What happened was interesting: the bot listened and slowly started to recommend what to do differently. It didn’t immediately say, “Your forecast is stupid; you’re always increasing by 30%.” It started slowly and carefully with hints, tips, and tricks.
Another experiment: I created 20 transcripts of a demand review meeting— made-up, with problems like mic off, video off, no contribution. I trained the model by explaining how I would see a best-in-class demand review meeting, what should be in, what should not, typical issues. Then I asked the model to evaluate what’s going wrong and defined a maturity model from very basic to best-in-class. I asked the model to rate the transcript’s maturity. The rating was very good; the findings were consistent— probably the same findings I would have, because I trained the model.
If you now imagine a bot listening to real meetings and reflecting on what it heard, that creates a big efficiency boost. Imagine 20 people connected and 15 with video off, mic off— they don’t need to connect. What can you do with the time saved? There’s a lot to improve. You’re not only doing a diagnostic; as you already have the context, you can immediately go into continuous improvement.
Joannes Vermorel: I believe that in some form a future like this will come to pass. Whether it’s with the current paradigm of LLM, the main problem is the data or knowledge flywheel: the LLM cannot learn, so the LLM must decide how to split bits of knowledge and store them for later use. That problem is not properly solved. Tomorrow we could consider a descendant or an alternative theory to LLM where learning is baked in.
Back to the case: having passive listening through the company, auto-archiving and organizing— ideas being exchanged, automation updating your library of insights and understanding for the current mind map of the company— that would have gigantic value.
Right now we have approximations: record two hours of meetings and produce very clean meeting notes. That’s useful; it saves time. But it’s not something you can later invoke readily. The LLM would have to rescan everything to find whether a specific point was discussed. Unless you ask it, it will not on its own click connections between “this was discussed” and “this was also discussed in two different meetings; we have contradictions.”
We lack learning. It’s absent. There are no “aha” moments for the model. It processes the context window linearly with a static, rigid intelligence that is extremely capable in very inhuman ways— but learning goes very deep and presents limitations.
One thing that’s strange: for data management, I suspect models will get better automatically irrespective of the technology, because they will ingest more examples. If you ask ChatGPT now to compose a prompt, it’s much better than two years ago. Why? Not because the core technology evolved on that front, but because there are now tons of examples on the web of good prompts fed back into the model. Hundreds of thousands of people posted tips and tricks; ChatGPT re-ingests that as part of training.
So for knowledge management: if enough people post tips on what counts as good tidbits of knowledge, these tools will get better because they integrate many heuristics.
Knut Alicke: Building on your meeting-notes example— I wouldn’t expect the bot to know everything. That’s why, coming back to the idea of the digital crew— digital Joannes or digital Conor or digital Knut— you start as a human in the loop. You have the transcript; you mark, “This goes wrong, this goes wrong.” You build the context more and more. After you did this ten times, the model can already spot 80% of the stuff. That’s the classical 80/20. In my experience over 25 years, 80% is always the same; you can train that. Then you have special cases where you add additional context.
I would agree it does not spit out, for example, “Do we need this demand review meeting at all?” That wouldn’t be a result. But to improve structure and results— and with this, a better-performing supply chain— this will be possible soon.
Joannes Vermorel: At Lokad, part of the prompts we now include when we summarize planning meetings: we add hints such as, “Whenever a date or a price in dollars or euros is mentioned, isolate it and check if there is a call to action attached to the date.” We do it in two passes to create a high-quality memo: scan the discussion, pause at dates, capture calls to action; same for financial amounts— what are the stakes? We give the LLM as part of the prompt tips to identify the truly useful things.
That’s the Lokad recipe. Now imagine Lokad publishes that on the web, and hundreds of thousands of people also publish their tips. “For these meetings, here’s the list of stuff to get a very effective summary.” That’s why I say knowledge flywheels will progress because people post tips and tricks.
But the core problem not solved is how to manage knowledge at scale. The closest approximation is RAG (Retrieval-Augmented Generation), but it’s still crude and does not scale very well. In a large company you quickly exceed the LLM capacity. Even if you don’t exceed the token window— now very long— if you throw hundreds of pages of quasi-irrelevant stuff, you do not get very good performance. You need something better.
There are ways to duct-tape the situation: linear scans, multipass. First pass to remove irrelevant stuff— but all of that is duct tape around the fact that learning is a second-class citizen in the current paradigm.
Conor Doherty: Back to the human side: some people are lovely to talk to off camera, but if you put a camera and a microphone in front of them and they know they’re being recorded, they get shy. It changes willingness to buy in because there’s a permanent record. Apply that to a demand planning meeting where people know there’s an AI tool listening, recording, analyzing, archiving, and possibly determining performance metrics. Do you see that being a problem for buy-in and participation?
Knut Alicke: For the first meeting, yes. For the second meeting, half. Then it diminishes. In lean, with the Gemba walk, when you observe assembly, the first observation shows people trying to do everything at their best. Come back the next day, and the next— sticking to the process diminishes and they return to normal habits. The fifth day shows more issues.
As a consultant connecting to these meetings, the first meeting goes relatively well; then you reconnect and reconnect, and people realize, “Oh, that’s normal,” and you see what’s going on. What is not yet clear is how to convince people that this bot is not doing bad.
One way might be to create a supply chain avatar— you’re a good-looking man, so it would look like you— and then people build trust and start to talk to the avatar in a normal way. There will still be people who won’t agree to talk; that would be their loss.
Conor Doherty: About guardrails— incorporating technology but ensuring safety and security. A core example is client meetings, demand planners, diagnostics: lots of sensitive information discussed— numbers, dates, values. People might have concerns about security and guardrails.
Joannes Vermorel: Absolutely. One of the things I taught my students in computer science almost 20 years ago, when email was still a little new: treat every email as if it were going to be made forever public. Once you send an email, you have no control. It can be forwarded. I was saying, “Assume all your emails will be dumped onto Usenet”— at the time, the equivalent of Reddit— and thus write accordingly.
We are entering a very strange world where it will be very difficult in the next 20 years to resist passive listening tools that record everything. I can see so many productivity gains; it will be difficult to resist. A company will embrace that and realize they are saving a lot of time. I’m talking of 20 years, not two.
Emails as we practice them today would feel strange to people in the 70s. The idea that a private conversation is always at risk of being sent to the national press would have sounded crazy. The idea that one email you wrote was forwarded to dozens of people happens all the time. People do not feel that you have to ask permission to forward an email. A postal letter was private; forwarding it to a third party was not okay.
Knut Alicke: Coming back to training: make sure people are aware of what should be shared and what should not. Another example from our book: one contributor told the story that in his S&OP process he wanted to invite the three most important suppliers. Everyone complained, “We can’t; they’ll know our production plan.” He said, “That’s exactly what I want to share, so they can prepare.” It’s the three most important, not thousands.
With transparency, you can be better. There’s often fear: “If they know what we do…” Yes, but if they know, they can prepare better and the overall supply chain is better. Same with these models— but clearly, if there’s a real secret, you don’t want it in the press the next morning.
Joannes Vermorel: My take is it changes the organization. For example, my parents’ generation working at Procter & Gamble: they had a quasi-military setup internally; the hierarchy was extremely strict; privacy and secrecy were paramount; information was distributed on a need-to-know basis. This has evolved enormously. Procter & Gamble nowadays is nothing like 50 years ago.
This technology will shift the market toward companies that say, “If everything is public, we have little to hide because our competitive edge does not come from that.” Yes, sometimes keeping cards close to your chest would be an edge, but you can have a business where secrets are unimportant. It will displace businesses more in this direction.
It will be difficult to truly secure these tools. For example, when I do audits of technological startups, I do it with no device whatsoever— just pen, paper, notebook— so I’m not going to accidentally leak anything. Everything is analog; leaking the data means stealing my notebook, and I have very illegible handwriting, so it’s semi-obfuscated on top.
Overall, with emails companies have embraced more openness; they can’t operate easily with secrets as 50 years ago. Now anybody can shoot a video and upload on TikTok; that creates complications. If there are things the public should not see, maybe we should not have them, because someone will upload a video: working conditions, a dirty kitchen. This technology will push it further, complicating what is your competitive edge in a world with very few secrets. That’s an ongoing conversation over 20 years, because these technologies will take time to deploy.
Conor Doherty: Guardrails are multifaceted: security and spend. Some argue generative AI has transformed supply chain for the worse, pointing to the “generative AI value gap”— enormous investments with little return; pilots purgatory. In both your perspectives, does that color your perception of GenAI’s transformative impact?
Knut Alicke: As discussed earlier, we are not yet there. We are still in the early phase. I heard a smart colleague say: when you do a pilot, ask the CEO or business unit leader whether he or she uses ChatGPT on a daily basis. With this, they know what is possible and what is not. I found that super intriguing.
If the boss has an idea of what’s possible, then either there is no pilot because it’s too early, or the pilot is set up with the right scope and expectations. If not, pilots are set up to solve world hunger and can only fail.
A lot of people still mess around and confuse GenAI and AI and digital and algorithms. It feels like it’s the new word to always use, and then everything you do is now GenAI— that’s not true. Bring it down to where it can add value and what is possible; then we will see impact in the next couple of years.
Another thought: very often the idea of impact is too local. In supply chain, the biggest impact of proper work is still availability. Availability is higher margin, higher revenue. Still a lot of people feel supply chain is only cost and inventory. Bring everything together end-to-end, look into your impact, and then you will see a boost in overall performance.
Joannes Vermorel: I think it’s a perfect blockchain replacement. More seriously: like all buzzwords, there is usually some technological tidbit that is genuine. But being part of the world of enterprise software vendors, the purchasing processes of large companies for enterprise software are dismal. A lot of money is wasted. GenAI happens to be what it’s wasted on. If we didn’t have GenAI, the money would be wasted on another buzzword.
The causality is not “GenAI creates the waste.” The causality is “the purchasing process is dysfunctional; thus money is wasted,” and the waste lands on the buzzword of the day. Two or three years ago it was blockchain; five years ago big data; 10 years ago data mining.
Your point about the boss playing with ChatGPT: the key is mechanical sympathy. Can you feel in your guts what the thing can do and cannot do? Same for blockchain/crypto: have you ever purchased Bitcoin, used it? Do you realize how it works? If you cannot grok the technology, this is not good.
Unfortunately we are back to the dismal process that ends up with an RFP of 600 questions. We are on the receiving end and get questions like, “Is the room that you use for your fax archive fire-proof?”— a question we received one month ago.
Knut Alicke: I agree, and another observation: the selection process of software is so strange. I always ask clients, “Why do you need these 500 specs?” All software companies by default will cross anything, because a software company sells a vision, not reality. Then you go into glossy meetings where they promise everything.
Let’s assume you want to buy a new car. Do you buy it by watching a 30-second YouTube video? Probably no. You want to sit in the car, touch it, drive it. Why don’t you do that with your software? Why don’t you do a proof of concept where you prove it’s working?
This comes back to capability: people need to roughly understand what’s in it. All of us teach, so we need to teach even more practical supply chain and bring this to the board. That’s why we did our book “Source to Sold,” to bring all that to the board level and make sure they understand the narrative— what works, what doesn’t, and where the impact is.
Conor Doherty: For the record, I had written down “JV—mechanical sympathy.” I knew you would say that. We’ve known each other a while now.
This brings us toward the end— two key questions. One is the board level; one is the boots-on-the-ground person. Knut, start with the board: how do you present the transformative value at the board level to persuade, and then how do you present it at the ground level?
Knut Alicke: At the board level, think about the language the board speaks: vision, growth, margin. You need to build your story to support that this new technology will help implement your vision— to be better for the customer, for the consumer, to be more profitable.
On the shop floor: everything is true that you told the board, but you need to get buy-in that this helps people in their daily life. If it’s just another tool that looks like a black box, that makes life harder, that won’t work. It needs to make life easier and help people contribute to overall performance.
If you do the Gemba in a warehouse, the people picking always have 5,000 ideas to improve. It’s not that they are stupid or don’t want; they know, but they’re often not asked. Same with planners. You need to tell the story: “Here we have something that helps you be a better planner, makes your life easier, and makes your work more interesting, because there’s new cool stuff to explore.”
Joannes Vermorel: Pitching to the board: the 21st century is the century of the mechanization of intellectual work. The 20th was the mechanization of physical work. If you don’t mechanize your armies of clerks— white-collar clerks— your competitors will do it. They will deliver more with fewer people and faster. If you have fewer people, you can be faster.
You have to look at what armies of white-collar people are doing. In supply chain planning, many companies have hundreds of people going over the same spreadsheets every single day— tens of thousands of lines. The biggest challenge is: when will those jobs be completely robotized? It’s not “if,” it’s “when.” We can disagree if it’s five or 50 years.
Conor Doherty: That was going to be my last question— thank you.
Joannes Vermorel: I am on the optimistic end. I know Knut thinks it will go through a long period improving the life of those people. But we can agree the status quo is unsatisfying. Some company will manage to do better than the status quo.
For people more operational: at Lokad we robotize, which is a tough discussion. Are you willing to embrace something that will make your job more challenging but also more interesting? The drudgery of spreadsheet descent is some kind of comfort, but it’s incredibly boring.
If you are super comfortable in an extremely boring task, that’s not good. If you have the willingness to challenge and elevate your work, this is the opportunity of a lifetime to be part of the revolution mechanizing the work. You’ll be part of the people doing much smarter things a level above mundane automation.
That is part of the Lokad vision: robotize the task and have people supervising the automation. There is another camp— take the people and make them vastly more productive. Two valid perspectives to break the status quo of spreadsheet drudgery.
Conor Doherty: Knut, it’s customary to give the closing thought to the guest. Could you expand on your perspective? Joannes seems to suggest that AI taking over the decision-making process might be more abrupt than your position. For those not familiar with your position, how do you see the near-term and mid-term evolution?
Knut Alicke: My hope— let me give two versions: optimistic and pessimistic. The optimistic: the models will be better; they will be able to learn via context or new inventions. With this, we will make life of planners much simpler and decision quality much better, which will lead to being intellectually much more challenged— which not all people will like. This will probably happen in the next five years.
The pessimistic— pessimists are experienced optimists: looking back 25 years, before McKinsey I worked in a startup doing planning software for consumer electronics. The quality of the software and decisions we had back then was, in retrospect, amazing. I still have a lot of clients not even near that. Having this in mind, I would say the pessimist says we will get there, but probably in 20 years.
Conor Doherty: So we’ve had the optimistic, the pessimistic; the apocalyptic would be Skynet tomorrow— next week— would that be appropriate?
Knut Alicke: Skynet will not happen— but not as an event anyway.
Conor Doherty: If you’re watching this in the future, send me a message; let me know. I don’t have any more questions, gentlemen. Joannes, thank you for joining me. Knut, you’ve been lovely. Thank you very much for joining us in studio. And to everyone else, I say: get back to work.
