Hello and welcome to McKinsey Live. I'm Lucia Rahilly, McKinsey's Editorial Director and your host for today's event, Never Just Tech, Unlocking the Full Value of Gen AI. Before we start, quick heads up that our goal with this series of live events is to give you the chance to engage with our experts on our latest research. Obviously, asking questions is a big, big part of that, so if you are registered for this event, please feel free to use the Q&A bar at the bottom of your live stream to ask whatever's on your mind. Today, Gen AI, we're all talking about it.
Many of us are piloting it, but what exactly is it that leaders need to do to actually realize the value of Gen AI at any kind of meaningful scale? Joining us today are two of McKinsey's leaders in this area, Jessica Lam and Gayatri Shanai. Jess is a partner and a leader in Quantum Black, McKinsey's AI arm. She works closely with North American healthcare and public sector clients, and she's based here in our New York office.
And Gayatri, there she is, is a senior partner, also here in New York. Gayatri is a leader with McKinsey Digital, where her client work focuses on tech and digital end-to-end, from strategy straight through very vitally, obviously, to implementation. Jess and Gayatri, welcome to McKinsey Live, and thanks for being together with us today.
Hello and thank you Lucia. Thank you for having us. It's great great to have you.
Let's start with some context. Leaders across the globe are grappling with Gen AI and how to make the most of its potential to drive new efficiencies and to generate value. Jess, before we get into the how of Gen AI, let's bring a little bit of the what to life. Talk to us about some of the potential applications of Gen AI that you're seeing and that you're particularly excited about? Absolutely, Alicia.
So our research actually estimates that Gen AI could contribute up to $4.4 trillion annually to the global economy. That's obviously an extremely large number, but there's three things that really stand out about this number to me. So first, that impact is going to be spread across industries. So some like high tech, retail, banking, they'll see north of $200 billion in economic impact. And even the ones that are lower on the impact range, like say life sciences or agriculture, they'll see north of $50 billion of impact.
Second, I think the impact is going to be across functions. So from marketing and sales to customer service to supply chain and operations, there's no function where they're not going to feel some general impact. And then the third, this one's really interesting to me, right?
The majority of that impact is going to come from just four archetypes of Gen AI use cases. It's what we call the four C's, but concision, customer engagement, coding, and creative content. Well, for instance, in healthcare, where I spend a lot of my own time, there's countless applications of generative AI. You can use generative AI for everything from personalized treatment. plans to streamlining administrative processes to reducing clinician documentation workload and beyond.
And that'll be the case in most industries. In finance, for example, it could be used for everything from fraud detection to a hyper-personalized marketing. And then there's going to be applications that cut across most, if not all, industries.
So things like increasing productivity of customer service or augmenting and improving developer speed or comparing an optimizer. optimizing supplier contracts. And that list just goes on and on.
Thanks, Jess. And Gayatri, let's move to you. One of the most compelling features of Gen AI is this widespread applicability across industries and functions, as Jess just described. But what are some of the challenges associated with using Gen AI?
And how should leaders be thinking about navigating those challenges successfully? It's a great question, Lucia. Given how much opportunity and value exists from the use of GenAI that Jess just described, all are asking why then are organizations and institutions struggling with the implementation? If you see, we are about a year and a half since ChatGPT was released. So isn't it time to see much more success stories?
And I would say, yeah, especially if you think about 2023, like when I reflect on last year, Hallucinate was the named word of the year by Cambridge Dictionary, a word that entered each of our vocabulary from days that were unprecedented. Then there were 13 Gen AI startups that each got valued at over a billion dollars. So when you look at that and you see today, there are about 200 million users that are active on ChatGPT as of this month, April 2024. And for contrast, there are 2.7 billion users on WhatsApp.
So what's really going on? Why are we not seeing more examples of Gen AI at scale? And here there are two challenges that we see.
playing at each other. The first is the risks of Gen AI are real and organizations are rightfully so thinking about those and how to thoughtfully consider those. Also Gen AI is working in a very rapidly evolving space.
With that it's becoming important that all leaders have a plan for how they can operate within that rapidly evolving environment while managing risks. We started to see AI regulations from the AI Act passed in Europe in December last year. The White House released an executive order on safe AI, safe use of AI in October last year. As such, there's both uncertainty and the need for agility in this setup to thoughtfully manage risk. On this front, I would say we as McKinsey have established and been using a responsible AI framework, which has a set of policies.
practices and tools for using in our own Gen AI development. And this framework, which includes human-centric AI development, robust data protection, privacy, security measures, as well as transparent and explainable AI, amongst others, has helped us make progress on our end. The second challenge here for Gen AI becoming really institutional has been the scaling challenges. Organizations are finding it's relatively easy to run pilots on GenAI and run a use case or even a few for that matter. However, when it comes to real value, driving real institutional value, these are similar to what companies saw on digital.
The value really comes from enterprise-wide adoption and scale. And here, most organizations are struggling is how I would describe it, and are in pilot purgatory land. The organizations that are truly succeeding are finding they are able to do so by fundamentally rewiring their organizations. And this takes a focus on six core enablers. The first is strategic roadmap, letting value be your guide and having a clear sense of where you want to lead.
and where you want to head. Focus on value by the way has been a long-standing principle but CEOs and leaders must now particularly rely on it to counterbalance the pressure of do something with Genial. The second on this dimension is talent. While the roadmap is more about the what, talent is about the how and here upskill, upskill, upskill is the mantra that we see working.
Number three is operating model. Realizing P&L impact from Gen AI requires it to be business-led. Operating in close partnership with both the business units, but also the functions to find ways to redesign processes in collaborative cross-functional teams and at pace has become critical.
The fourth is technology. Technology architecture and getting the underlying technology to be ready is important. Fifth is data. GenAI makes a focus on data quality and in particular unstructured data quality really more central.
And then last but not the least is adoption and scaling. Taking a structured change management approach, including a very clear perspective on GenAI risks and guardrails that we discussed is critical. And again, here we are seeing organizations that are seeing success. in institutionalizing Gen AI invest almost $5 in change management for every $1 investment in technology. And truly here, it's never just tech.
Thank you, Gayatri. Just Gayatri invoked the dreaded pilot purgatory. We see lots of folks adopting various AI technologies for different use cases in this area or that area.
But let's turn to how leaders should think about designing the kind of solid Gen AI framework that will really help their organization and gender value and ultimately to drive out performance. Like Gayatri mentioned, technology alone doesn't create the value. It is truly never just that. But when combined with the right organizational shifts, I think we see that Gen AI can actually generate immense value that can put even more distance between the leaders and the laggards, honestly. And the six areas that Gayatri outlined, which are from our book Rewired, they're needed to deliver the kind of broad change that harnesses digital and AI.
GenAI is no different. There's just some nuances that we see. So on the strategic roadmap front, for example, you need to figure out where GenAI co-pilots can give you real competitive advantage.
So leadership of the organization needs to be aligned around the potential of GenAI and its impact on the existing transformation roadmap. This will ensure that GenAI is integrated into the overall strategy and vision for the company. And this is where our Taker, Shaper, Maker framework comes in.
To recap, being a taker means using available tools, often via APIs and subscription services. A shaper really means integrating your available, the existing off-the-shelf available models, but with your own proprietary data. And then maker means building your own large language models. And for now, the maker approach is just way too expensive for most companies.
So really the key here is figuring out when to be a taker, like when you're really just trying to get productivity improvements. versus when you want to invest to be a shaper and you really see potential competitive advantage in that use case. And then on talent, I would say you need to upskill the talent you have. We need to be clear about the Gen AI specific skills that you're going to need. So organizations are going to need to manage talent to stay ahead of the Gen AI skill scales.
It's important to bear in mind that successful Gen AI skills are about more than just coding. So in our own work, We've observed a few kind of helpful things for those that are building Gen AI models, including things like design skills to uncover where to focus solutions, or contextual understanding to ensure that the most relevant and high quality answers are really getting generated, and collaboration skills to work well with the knowledge experts and to test and know how to validate your answers, strong forensic skills to figure out the causes of breakdowns, because breakdowns will inevitably happen, and anticipation skills to conceive of and plan out for those possible outcomes and put in the right kind of tracking into the code so that you can really keep things running. And then moving to scaling, I'd say two things, right? Creating a centralized team to establish standards and then really building trust and reusability.
So I'm creating a centralized team to establish standards that enable responsible scaling. This... central team needs to focus on establishing protocols and standards to support scalability while minimizing risks and costs. So this includes things like procuring models, developing data readiness standards, creating prompt libraries, allocating resources, and actually during the development of Lilly, McKinsey's own GenAI knowledge platform, we prioritize scalability through an open plugin architecture and standardized APIs. So this facilitated secure experimentation and access to a GPT-LLM, transitioning Lilly from being a product used by select teams to a platform that was accessible across the organization.
And then on the building trust and reusability front, many people are going to be concerned about using Gen AI. They just don't, right? And if they're going to use it in their day-to-day jobs, they want to truly understand how it works.
So plan to invest more time and money than you think to really build that trust. I love that, Jess. That's a great list.
And I would perhaps add two things to this, one on the technology front and one on the data front. On technology, there is a need to set up the technology architecture to scale. Creating a Gen AI model is fairly simple, but achieving scalability is complex. And here there are three critical things that make a difference.
The first is focusing on reusing technology. We have seen reusing code accelerate the development of Gen AI applications by 30 to 50%. That's a big number. And this starts with prioritizing the identification, as well as the development of common capabilities across key use cases. For example, the financial services company identified three reusable components that were applicable to about 100 use cases.
and by prioritizing their development they expedited the progress significantly. The second is focusing the architecture on enabling efficient connections between Gen-AI models and the internal systems. As you can imagine, efficient Gen-AI models require seamless access to business data and applications.
And integration advancements streamline this process, but clear guidelines are crucial to prevent technical debt. And here we are seeing CIOs and CTOs needing to define standards, like whether it's a model hub for on-demand provision of approved models, or it's standardized APIs for connectivities, or it's cloud foundations to use modular isolation zones. So focusing the architecture to enable these connections between the internal systems and Gen-AI models become critical. The last would be being clear on changes in the tech stack you need and support. Scaling Gen-AI requires changes in multiple layers across the tech stack.
So whether it's applications or MLOps or models and APIs, the infrastructure across every part of the tech stack, every layer of the tech stack. there's a combination of options you can consider in how you solution, including build by and partner decision. This obviously makes it complicated to think through these changes.
And organizations need to think through these strategies to be able to maintain optionality, to switch and experiment, and as well, avoid any kind of vendor locking. Lastly, on the data front, like we've talked a bit about data quality being important and a focus on unstructured data being important. This to me requires two real focus and priorities. One is building specific capabilities into the data architecture to support the broadest set of use cases.
Building relevant capabilities such as vector databases, data free and post-processing pipelines. into the existing data architecture in support of unstructured data is important. And in our experience, organizations that are having success are focusing on five key components of the data architecture. And these include number one, unstructured data stores, data pre-processing, LLM integrations, vector databases, and prompt engineering.
And then the last piece I would say is focusing on key points of the data life cycle to ensure high quality. One reason pinpointing data quality issues is much more difficult in Gen AI than in classical ML models is because there's so much more data and so much more of it is unstructured, making it really difficult to use existing tools. Developing multiple interventions into the data life cycle, both human as well as automated.
For example, setting minimum thresholds of quality of unstructured content to be included into Gen-AI applications has become important in this context. Thank you, Gayatri. Jasmine, Gayatri, that's been hugely helpful. We have gotten a lot of questions from the audience, so I'd like to pause here and try to address at least a few of these. First, you know, Guy, you mentioned that Gen AI is a rapidly evolving space.
And we've had a number of questions on that fast moving nature of Gen AI and how to plan for and navigate that dynamism. Jess, how would you recommend future proofing Gen AI investments to stay competitive in a world where the tech is just changing so quickly and sometimes unpredictably? Really good question, Lucia, and one I hear a lot from clients.
So to future-proof GenAI investments and remain competitive as all of this, as you pointed out, very rapidly evolving, there's a few things I might focus on. And at the risk of repeating what I said already, I think we just really need to develop a deep understanding of where GenAI can have the biggest impact in the business. So instead of chasing individual GenAI pilots.
You need to really find those high impact domains where you can truly disrupt the business model and can transform them more holistically. And that means that you're kind of looking at things beyond just the Gen AI model. And so that gets me to my second point, which is Gen AI will not drive competitive advantage on its own.
So you're going to need to combine it with traditional AI, but also just with other levers. So, for example, one enterprise decided to reimagine their sales domain. They had an AI assistant identifying a potential lead for the sales rep based on their upcoming construction projects in the next month. And the assistant would then aid the rep in crafting a hyper-personalized sales pitch, which incorporated data from demo requests.
social media interactions, known budgeting cycles, you know, you name it. The assistant would then offer real-time targeted recommendations of the bundles. So basically in order to do the use case here, right, Gen AI, other AI, other levers such as process redesign, workflow optimization, digitization, they all had to come together. So that is a great way to feature proof, right?
You're not going all in on one specific technology or one kind of specific thing. And then third, for the foreseeable future, at least, right, Gen A and I solutions will work best alongside humans and not in place of them. Right. So this means driving real investment and commitment towards that last mile adoption. And that's part of creating that ecosystem of reusable assets that I mentioned, right, which will really accelerate development and adoption.
And that helps helps future those investments. Thanks, Jess. Gayatri, anything to add or shall we move along to the next question? I think Jess covered it beautifully.
The focus on value, the priority of like where is it really a competitive advantage and how do you use it with other levers? And then ensuring that this is in human augmentation and not in lieu of humans. It's a beautiful way of synthesizing what she said.
Great. Okay. Next, we talked a bit about the talent question, but it's obviously top of mind. We've had a lot of questions on it. It's clearly something organizations feel they need to act on quickly.
I think it might help to get a bit more specific about the skill sets or the expertise that organizations need to build Gen A capabilities and solutions. Gayatri, why don't we turn to you here if you would jump in. Yeah, sure.
Look, this is a question every organization, every institution is asking itself, right? Like, what are the real skills needed and how do they get there? I would say there are three kinds of skills that we are seeing organizations need as they build Gen AI capabilities and solutions. The first of which is technical, and Jess covered it earlier when she mentioned all the additional skills you want to ensure your data scientists and engineers possess. as we dive deeply into Gen AI.
So I'll build on top of that to say, in addition to the technical skills, to the technology skills, you will need talent in two other categories. The second category of talent needed here would be experts in ethics and responsible AI practices. This will include folks with backgrounds in privacy, in legal, in specializations, in mitigating biases in AI. and will be an important category of talent to bring into your practices of developing Gen AI applications. The third bucket, but an equally important bucket of talent needed, will continue to be your subject matter experts and your domain experts who have deep industry and functional background.
They'll be core to Jess's points around ensuring true business value continues to be unlocked. And that you're not just building shiny objects, but you're delivering real long lasting sustainable impact using Gen AI. Very helpful. Let's do one more quick one here on measurement.
So we've talked a lot about how organizations can set up for success, but how best to gauge how they're actually doing. Jess, let's turn back to you. How do you recommend or how do you see clients measuring success in this Gen AI context?
Yeah, there's no one way. Right. So this is going to require a pretty comprehensive approach. And you're going to need to look at a pretty holistic set of metrics.
Right. There's, of course, some of the kind of financial and operational metrics that you would be looking at, but also customer satisfaction, employee satisfaction, societal impact. actually how your organizational capabilities are evolving, right?
There's a number of categories that you're going to need to look at. It's not going to be black and white on kind of what exactly you can measure here. And across these categories, you're really going to need to design, to measure, monitor, and evaluate the metrics that matter most, right? Because there's a lot of stuff you could be looking at. And these are going to include, you know, things like I mentioned, right?
Financial ones for revenue generation, cost savings, productivity, user satisfaction. adoption rate metrics that will be really important to understanding how people are actually doing and feeling about using this technology. Then those metrics that cater to customer employees.
Then really a full, fulsome set of metrics to understand any potential bias and ethical implications. You want to make sure you're actually monitoring what you are putting out there and what effect that might have over both the short and long run. Thanks, Jess.
And just one more, since we have an additional minute, there's a follow-up question here on the types of metrics that are emerging in success stories. Do you have any real examples that might help bring this measurement question to life of how a client that is measuring their efforts successfully, either of you could take that? Can I just start with saying I have not yet seen anyone have the full set and suite already up and running and monitoring. I will say I have seen a number who have found that their existing way of monitoring different use cases and these things is not quite sufficient for Gen AI.
So there is a lot of revamping going on at the moment as they are seeing what is coming out. actively monitoring what they're observing, and then seeing how that's different with generated AI, right? And how that really leads to some new and kind of different types of metrics in some of the categories I was just going through, right?
But adding in some of those kind of broader metrics to make sure that they're really understanding the full impact of these models. Great. And I would build on Jess's example, because I've seen something similar with my clients.
They started with a lot of efficiency metrics. And then as we talked about today, adoption and scaling is one of the most important enablers to allow for Gen AI to get accepted through the organization by both employees and other stakeholders. And a particular focus on efficiency metrics by itself is not helpful. So these clients are now learning, these organizations are learning that it helps to have a much more holistic set of metrics.
and also a focus on effectiveness and quality and how many humans are you really augmenting using these tools versus pure efficiency. Okay, very helpful. That regrettably brings us to time.
This has been such a great discussion. Jess, Gayatri, thanks again for your insights today. Thank you, Lucia.
And as always, many thanks to all of you in our McKinsey Live audience. You'll be able to find a replay of this and of all our previous McKinsey Lives on our website at McKinsey.com slash live. Stay tuned also for our next McKinsey Live, Capturing the Power of Productivity Through Tech Investment, a related topic, with a couple of our top leaders, senior partner Olivia White and DNA digital leader Rodney Zemel. That's on May 20th at 10.30 a.m. EST.
Have a great day or evening and see you next time. Be well.