In this episode of McKinsey Talks Operations, host Christian Johnson sits down with senior partner Nicolai Müller and partner Marie El Hoyek from McKinsey’s Operations Practice. Together, they discuss the game-changing potential of generative AI. From automating complex processes to unprecedented opportunities across industries, discover insights on productivity boosts, system considerations, and the vital capabilities organizations need for successful integration.

Their conversation has been edited for clarity.

Christian Johnson: Your company’s future demands agile, flexible, and resilient operations. I’m your host, Christian Johnson, and you’re listening to McKinsey Talks Operations, a podcast where the world’s C-suite leaders and McKinsey experts cut through the noise and uncover how to create a new operational reality. As we’re recording this episode in late 2023, it’s clear that generative AI, or gen AI, has become the topic in conversations about digital, analytics, and operations. This new deep learning technology is already making ripples with applications across the value chain.

For today’s episode, I’m delighted to be joined by Marie El Hoyek, an associate partner based in London, and Nicolai Müller, a senior partner based in Cologne. Together, we’ll be exploring what generative AI in operations is, how it’s different from digital twins and other AI technologies, its potential, and its risks. We’ll also look at what it takes to get started with these tools. Nicolai, great to have you here today. Welcome.

Nicolai Müller: Thank you. It’s a pleasure to be here, Christian.

Christian Johnson: Marie, so pleased you’re able to share your thoughts with us today. Thanks for joining.

Marie El Hoyek: Pleasure being here, Christian.

Christian Johnson: Great. So, Nicolai, can you tell us a bit about why you believe generative AI is worthy of discussion for operations leaders, especially now?

Nicolai Müller: In the past decades, there was this mantra of being faster, being more efficient, and pushing productivity. Tools we all know, such as Lean, offshoring, reviewing make-or-buy decisions, and also through technology—but we see nowadays that this productivity improvement gets more complex.

In this scenario, we now have a new technology coming in: generative AI. It promises to automate processes that, in the past, were hard to automate—areas that are more in management collaboration, which currently humans are operating, and also in complex data that you have to manage. So, in this context, there’s the question: How much will generative AI help in the search for productivity?

The McKinsey Global Institute has looked into this, and we discovered that, particularly in the areas of collaboration and management, around 50 percent of typical activities can now be automated by generative AI. Also, when it comes to handling complex data and synthesizing the essence of that, we believe there’s a huge jump in automation. This may lead to value creation across industries and functions—from pharmaceuticals, to automotive, to machinery and functions from engineering, procurement, and supply chain, to customer operations—that can unleash tremendous value. We talk about $3.5 [trillion] to $4 trillion, which is approximately the GDP of the UK.

Christian Johnson: Nicolai, what are some of the more specific opportunities that your clients are focusing on, and that you’re focusing on right now?

Nicolai Müller: Where I see our clients acting fast is in product development. And if you look deeper into product development, especially in software coding, we see up to a 50 percent productivity increase by having a machine produce code from the simple instruction, “Please give me the code for a program doing XY,” and by using tools like ChatGPT and others, a solution is generated. This is one application area where we see generative AI becoming a copilot for humans, aiding in tasks ranging from program management to procurement, and assisting supply chain managers in performing their roles more effectively.

Christian Johnson: Thanks, Nicolai. That has given us a great idea of the why and some of the opportunities. Now, let’s go into a little bit more detail about what generative AI is. Marie, what can you describe here for us?

Marie El Hoyek: Generative AI is a fascinating field, and just like the name suggests, it exists at the intersection of artificial intelligence and natural-language processing. Essentially, it involves a machine that can analyze something, and this something can now be unstructured, like language or pictures. Similar to a person, generative AI is all about teaching machines to understand and generate text or content.

Now, to add a bit more flavor, let’s discuss the different generations of large language models—LLMs. These models are the driving force behind what we refer to as generative AI. One of the first ones we commonly heard about is GPT-3, which stands for generative pretrained transformer 3. When it was introduced, it had 175 billion parameters. Think of parameters as the amount of information it had learned, allowing it to generate text ranging from writing letters to answering questions, primarily text-based. Soon after, GPT-4 was released, and we saw a leap from 175 billion to 170 trillion parameters. Consider how much more it had learned, making it more fluent and accurate, and now it could also be used for images and video.

This is the transformative possibility with generative AI. You can now generate new content in many different types of spaces. Now, that being said, generative AI comes with its own set of risks and challenges. If you imagine that it’s based on logic or probabilities, very similar to the human brain, the answers come from what you’ve learned and your sources. Because of this fact, any generative AI can give you a convincingly wrong answer—and this is what we call hallucination.

Christian Johnson: I love that term. But what do you do about it? How do you mitigate?

Marie El Hoyek: If you had a person answering you based on wrong information, you would tell them, “I want your answer from this specific book.” Similarly, you can prompt generative AI better by telling it, “I want you to answer me from this data set or to tell me where you’re guessing.”

Another risk is model bias. Imagine that the model or the person has learned from the internet as its source, which is not the most respectful or kindest place. So, whenever you use a model, you need to be able to counter these biases and instruct it not to use inappropriate or flawed sources, or things you don’t trust. Another risk that is top of mind is IP [intellectual property] risk. Now, if you imagine generative AI generating code for you, who owns the code? Is it the gen AI that generated it or the requester who wanted it? These details are something we will need to iron out soon.

Christian Johnson: What I’m appreciating here is the discussion of the very limits of the data sources. That’s really critical, right?

Marie El Hoyek: It’s critical. Additionally, the fact that you need to guide your own data means you have to take care of your data and ensure its safety. Otherwise, that is also an added risk. That being said, all of these risks can be mitigated. However, we need to be aware of them, plan for them, or approach them in a way that limits them so we can control them. By the way, we’re witnessing regulations and offerings that are starting to adapt to these risks, and I expect we’re going to see quite a few changes in the near future.

Christian Johnson: Just the evolution here—the rapid expansion from 100 billion with a “B” to 170 trillion with a “T” is really dramatic. I think one thing we would now like to turn to is how this is being used and where we are seeing use cases come to life in businesses today. What are some really good examples of that?

Nicolai Müller: I think it’s a question that clients have to ask themselves: What impact do I want to achieve? In the end, we have to solve one big question and challenge: how to increase productivity, which involves efficiency and effectiveness.

When we look into use cases, we try to explore different angles. One is the question of automation. Things that currently take hours can be done in seconds. But it’s also about augmentation, where a human may only be able to work with a certain set of data. Imagine being able to access all the data in the world that exist. This was one of the big revolutions; the internet gave us access to all data. Now, with machines, we can use and synthesize that data. So we talk about augmentation. And then we see innovation.

Innovation is the capacity to come up with completely new solutions. Not just making an existing product cheaper or achieving faster product development, but now generating completely new ideas for features and services. So what have we seen? Automation. I talked about how I’m fascinated by what we can now do in software coding and the whole field of engineering. You also heard, for example, the CEO of Nvidia saying, “Hey, the era of software is over. I think we told all our kids to learn software; now you figure out software can be done by a machine.” It’s a huge evolution that we see, but not only in software.

Parts and hardware development. Synthesizing a huge amount of requirements that your customer gives to you, asking generative AI to understand what the requirements are and how the requirements differ from the last product. How do the requirements vary between products? Are they similar or different? It will help to come to a better synthesis, better understanding of the requirement, and develop faster and better products.

In augmentation in pharma and research, I think we’ll see a humongous increase in effectiveness, output, and research. We have cases in pharma where you can imagine understanding each little molecule, what kind of effect it has, and how it reacts with other molecules. It’s something that is instrumental. So we see vaccines or other pharma products being developed faster than traditionally was expected by using generative AI. This augmentation leads to a better kind of solution.

As for innovation, you may have also seen one famous German OEM in the US that has integrated ChatGPT into their products. So you can interact and speak with your car. This is innovation. But, Marie, you have also worked with me in this space. What have you seen?

Marie El Hoyek: My background is in industrials, very much deep in operations. Personally, I love all the copilot applications, especially in procurement. The idea that you can ask a friend who knows all your contracts and can answer any question by heart and in plain English is just mind-blowing to me. So, instead of analyzing old contracts, price history, and external trends, I can simply ask the questions. I’m sure there are many more cool applications in terms of content generation, etcetera, but this one, in particular, blew my mind.

Nicolai Müller: And Marie, what I observed are these humongous opportunities out there and the numerous use cases. I mean, we have been in workshops where we were sitting with our clients, and easily after an hour or two, we didn’t end up with just five or six potential use cases across a whole different function, but rather 150 or more. I see here a huge opportunity, but the challenge that we’re facing is, where do you start? What I call “happy generative AI,” where a copilot can help you in your daily job, may become a commodity that everybody can do. Where is the truly transformative generative AI? Is it leading to a differentiating factor for your business? Is it really adding value and creating value for your customers?

I think this is the challenge we face. It’s like what we say in Germany, you don’t see the woods because of the amount of trees in front of you. So where do you start and where do you end?

Christian Johnson: Can’t see the forest for the trees. That’s exactly it. When I hear all of this excitement, I also think of the classic chart that we’ve seen for technologies in general, where you have this initial sharp upward curve as everybody gets very excited about it. Then it sounds like where you’re moving is, we need to anticipate when organizations either find, as you’ve put it, that it’s commoditized or that it’s hard. And that gets us down then to value. How do companies think about long-term value and not just a set of very exciting use cases that may not build forward very much

Nicolai Müller: This is a challenging question. If you look into the Google search index, which gives you a bit of a feeling of where we are on the curve, you’ll find out that it’s now Googled more than any traditional operational questions you have. You have seen all the digital manufacturing terms out there. We have cloud computing and the Internet of Things that we’ve now seen over the years, and it’s a constant discussion.

Generative AI in operations has just started to pick up, I would say, in the first quarter of this year. And it has, in terms of the amount of searches people are doing, overtaken everything you can imagine. This may give you an indication that there is a huge hype out there. But has this hype and all the dreams come true yet? Indeed, people are now starting to recognize that things are easy, like the low-hanging fruits, but actually, the real core is still challenging to implement and also to make your company adaptive to changes. So we are still on the verge of answering one important question when it comes to generative AI. Is it now just another tool kit in your operations, like lean or digital or any other artificial intelligence—that is, predictive maintenance—and enables levers you can pull? Or is it a disruption on its own? Is it changing the way you operate? I think these are two scenarios I can imagine.

I tend to believe that in the next two to three years, we’ll see these two questions answered. And it may differ completely by player or by industry what the outcome is. Let’s talk about disruption. Imagine that coding is now easy. Often, you have, for example, an automotive OEM defining requirements, and then you have a supplier more or less programming the code. If now that code can be programmed by machine, do you need a supplier anymore? It can be disruptive and threatening to say that the raison d’etre, or the reason for the supplier to exist, is actually gone. So this is an extreme of a disruption.

For example, for a very research-heavy company, suddenly, if you tap into completely new sources of data, you come to a completely new set of products. And finding the language model that suits you by adopting generative AI in ways that are differentiating may help you to move faster and with better products. I think this is the most pressing question that clients have to answer.

Christian Johnson: I think one of the things we’re struggling with and organizations seem to always struggle with when it comes to a new technology or a new methodology is how do you scale? We talked years ago about pilot purgatory—this idea that you try a bunch of ideas, but then they’re never really cohering in a way that creates lasting value. So how can organizations think about this in a way that they can minimize or even avoid that kind of stagnation with this idea?

Marie El Hoyek: This is a good question, Christian. Generative AI might be relatively new, but we have years of experience in scaling digital transformations. To your point, one of the biggest challenges is the pilot trap. Building a pilot or innovating with the technology is great, but transforming an organization is a whole different playing field.

Nicolai talked about the business-led mindset to prioritize applications that are useful with real business ROI. Beyond that, getting a real impact out of any digital change, and for generative AI in particular, will always be both a human and systems question. The way I’d summarize it is, without people, the best technology has no impact. We need to take our people on a real change journey to build the capabilities to use this technology, develop this technology, but also just to know what you can ask of this technology. And by the way, in terms of developing it, there are new skills that are needed here.

Christian Johnson: So what sort of capabilities do organizations really need now?

Marie El Hoyek: I’m thinking about prompt engineering, for example, which is the ability to ask a question really, really well. Now, number two is in terms of systems. There are fundamental questions that businesses should consider early to ensure that whatever they decide leads to capable, consistent, and safe technology usage. You don’t want to end up with ten different decisions on the technology because pilots are going left and right.

So you’re going to be wondering, do we build our own language models? Do we work with partners? Do we get off-the-shelf solutions? Where do we put our data? How do we process it? These questions are better learned early, and you need to make a conscious decision about them, to ensure that later on, as you use generative AI more and more, your solution is safe, scalable, and consistent. So, yes, for me, it’s both the people element and the systems element that will enable us to go through to the finish line.

Christian Johnson: Excellent. Thank you very much. We’re now nearing the end of our discussion. But before you go, I’d like to ask one final question, which is, what should our audience be doing now to bring generative AI to their organizations? There’s so much noise out there. We’ve got a strong idea of that with the Google searches. So how do you start to cut through and make a solid start?

Nicolai Müller: I would recommend two things. First is to start with a pilot, and I would even use the term “play” with generative AI. The cost of doing nothing is just too high because everybody has this at the top of their agenda. I think it’s the one topic that every management board has looked into, that every CEO has explored across all regions and industries. So it’s important that you start and see what generative AI can do.

In parallel, you need to really think about your strategy. When I talk about strategy, it includes a couple of elements. It’s a question of how will this impact my business? Where will it lead to improvements? Where will it not lead to improvement? Should I go fast ? Should I not go fast? Do I have solutions out there? Do I need partners? Can I rely on existing LLMs out there, or should I build my own? I think this is the whole question of truly understanding what generative AI in 3 to 5 years means for us.

Then there’s a layer in the strategy, which is about getting the data technology right. It’s understanding how you want to put governance and organization in place, which can build solutions. And there’s the question, where do the competencies in my company actually come from? Can I build them? Do I need to acquire them? So you need to be thoughtful about the whole question of competencies needed.

And then there’s the question of actually making the change. We often hear that this is the most important thing. You need to make people work with generative AI. You need to capture the early wins, but also things that are more challenging.

Christian Johnson: Excellent. And Marie, anything you’d like to add?

Marie El Hoyek: Yes, Christian. Nicolai, last time we spoke, you talked about this fresh breath of innovation in our companies, and I love to repeat this. You can see it in our discussion even. This gives us the ability to dream again, to come up with new things, and to hope for more impact. And I think, to some extent, we just need to learn, and start doing it, and start capturing it.

Christian Johnson: That’s a lovely ending, Marie. Thank you both, Nicolai and Marie, for sharing your expertise and experiences of generative AI with us today. It’s a topic that we don’t see going away anytime soon. So, your advice on diving in, but with both eyes open to risk mitigation and value creation, is a great note to end on.