Driving innovation with generative AI

As last year’s hottest topic—generative AI (gen AI)—becomes this year’s deployment discussion topic, companies are keen to turn discussion of generative AI’s potential into action to capture its benefits. In this episode of the Inside the Strategy Room, two McKinsey experts discuss how top innovators use the technology to drive growth. Laura LaBerge is an expert in our Strategic Growth & Innovation service line, of which Matt Banholzer is global coleader. They are coauthors of a recent article that explains why companies with cultures that embrace innovation have an edge with generative AI. This is an edited transcript of their conversation. For more discussions on the strategy issues that matter, follow the series on your preferred podcast platform.

Sean Brown: Before we get into how generative AI can help businesses innovate, what do you cover under the innovation umbrella?

Matt Banholzer: Our definition covers not only new products but new processes and operating models that can create a competitive advantage by making you more fluid, adaptive, or cost-effective. Innovation is also about new customer experiences and ways of engaging with them, and new business models and value propositions. In the past ten years, for example, many companies shifted from selling products to selling services, or adopted subscription-based approaches. Business model innovations can also include different routes to market or using your assets in new ways.

Our research suggests that we may be transitioning to a new era shaped by new technology platforms and major demographic shifts. To thrive in this world, you have to innovate, because what got you here may not get you there. Many of your business norms, operating models, or products may not be effective in the future and not innovating may be riskier than making big bets on growth opportunities. Times of uncertainty require not only battening down the hatches but using productivity to generate cash flows with which you can establish beachheads for new growth.

Sean Brown: Your article says that top innovators excel at finding and capitalizing on these new sources of growth. How do they do that, and where does gen AI come in?

Matt Banholzer: We did a survey to find out what drives these companies’ outperformance and found that what they have in common is an innovation culture. We were shocked at the distance between the top and bottom performers, which was as much as a 1,000-percent-plus difference [exhibit]. Those with strong innovation cultures are much more likely to report that their products and services lead their industries and that their organizations are best in class in the speed of new product development. This is where generative AI comes in: it’s about developing and testing and deploying. Some of the leading companies were deploying gen AI one or two years before ChatGPT took off.

Sean Brown: What does it mean to have an innovation culture?

Matt Banholzer: We have written before about instilling an innovation commitment, the human factors in innovation, and the eight essentials of innovation, where we define the building blocks of an innovation culture. For example, are you setting bold aspirations that can only be reached through innovation? Often, companies can deliver their strategies without innovation, so it’s not surprising that they aren’t innovating. Innovation culture also means applying customer-backed insights and what the market is telling you. Furthermore, top innovators challenge assumptions and assertions, embrace uncertainty, and enable iterative development.

Sean Brown: Are there particular areas where these companies focus their attention and investments?

Laura LaBerge: One difference between them and others is that top innovators invest more in R&D and digital technology. But it’s not just more—they invest differently and get much higher returns on those investments. On average, they spend 55 percent more on digital technologies, with a focus on tech that enables them to develop strategic differentiation. Additionally, they focus on speed, granularity, and integration, reporting two to three times the capabilities in these areas than the average company—and as much as nine times more than weak innovators. These investments prewire them to take advantage of new types of technologies, so it’s not surprising that they are well ahead in deploying generative AI at scale to accelerate R&D and innovation processes. In the past, these organizations were ahead on other types of technological advances, such as the Internet of Things or design engineering. What’s interesting in this moment is the degree to which gen AI can play to their strengths.

Sean Brown: How do these companies develop the speed, granularity, and integration you mentioned?

Laura LaBerge: Regarding speed, for example, business leaders and product teams use real-time data to drive rapid improvements. They use technology extensively throughout the organization, going beyond simple automation to integrating development, security, and operations processes. Granularity is about leveraging machine learning to analyze data at scale, and integration refers to their organization-wide focus on end users and seamless embedding of control functions. Innovative companies had all these elements in place before gen AI came into play, and these capabilities turn out to be critical to both taking advantage and avoiding the risks of gen AI.

Sean Brown: What can companies that are in earlier stages of experimenting with gen AI learn from these leaders?

Matt Banholzer: There are five elements in how these companies approach gen AI. First, they know how to ask good questions. This goes beyond simple prompt engineering and thinking about the syntax—they understand what problems the business needs to solve and how to use gen AI to address those questions. Second, they focus on weeding out bad answers. That doesn’t mean simply rejecting answers that don’t make sense but always challenging assertions and viewing them as assumptions. When companies build new businesses or launch new products outside the core, they make assumptions around customers’ preferences and their willingness to pay, or whether they can manufacture the product and the sales force can sell it. In business as usual, you can assert how that will go because you have pattern recognition. In innovation, you have to question those assumptions, and this mindset translates cleanly to gen AI. When gen AI spits out an answer, top innovators ask, “Is this a useful answer?”

The third difference is that they continually build proprietary data. Gen AI is a great way to rapidly summarize and synthesize data, but its ability to drive insights from unstructured data is limited, especially around specific corporate decisions. At McKinsey, we have gen AI tools that are wired into some of our proprietary databases on company performances, market size, and so on, so the answers are synthesized in the right way, and we can sift through data that others don’t have.

The fourth capability top innovators have prewired is learning and changing course quickly. Agile practices effectively mean the ability to move forward under uncertainty, to test and learn and act without having full answers. That’s pertinent to gen AI because it allows you to say, “This gen AI workflow may not pan out, but we will test it and if it works, scale it as fast as we can.” This iterative test-and-learn loop is how organizations escape pilot purgatory.

And fifth, companies with innovation cultures have workflows already wired for no human touch. People ask the questions and spot bad answers, but many other steps are automated. To take a CRM [customer relationship management] system as an example, these companies can go from identifying customers to having gen AI develop potential prompts to reach out to these customers, to following up. You make it as easy and seamless as possible for salespeople to engage.

Sean Brown: How do you put this prewiring in place if you’re in the early stages of gen AI adoption? Can you do it in stages? Or is it all or nothing?

Laura LaBerge: It isn’t necessary to do it all at once, and certainly not to do it all at scale. The baseline is to do no harm, especially around data security. As you experiment with these technologies, you need to put in place regulatory and data security boundaries. Then, figure out where in your organization gen AI could drive the biggest strategic advantage by enabling you to accelerate or be more granular and start testing.

Matt Banholzer: Most leading companies have taken a use-case-driven approach where they pick one element they know they want to transform. The early examples were skewed to things like customer service prompts, but they can come from anywhere. I want to emphasize that companies in every sector are testing the technology. In a chemical or pharma R&D context, companies trying to discover new molecules start with a large library of candidate molecules that may be generated by gen AI or experts. Many steps follow but you can accelerate a slow early step.

Sean Brown: Numerous regulations have been introduced or proposed related to gen AI. What impact, if any, might those have on the five prewiring areas you talked about?

Matt Banholzer: There’s much debate about the executive orders and regulations that have come down. Many of them are mainly focused on how you declare your use of the tool, but back to my earlier examples, there are regulations around what chemicals can be used, how you synthesize them, safety regulations, et cetera. You can choose to use an advanced chemical, but it requires guardrails.

Laura LaBerge: It’s likely to progress along similar lines as we saw with regulations around personal data, which varied by region and evolved over time. Organizations had to stay on top of it and adapt.

Sean Brown: Let’s dive deeper into the five areas where top innovators lead. How do you ask gen AI good questions?

Matt Banholzer: Many of the skills needed to get the most out of gen AI are skills companies have honed doing product launches or applying machine learning, but we were surprised by the degree of differentiation between the top and bottom performers. The top performers understand the limitations of the tool. Just like you don’t use a hammer to turn a screw, you don’t ask gen AI questions that are best answered in other ways. It’s about avoiding garbage in, garbage out. The question needs to be answerable, and you have to understand the reliability of data, but there are probably specific questions at given points in a workflow that you can automate.

This is where prompt engineering comes in. Just asking a sales team or a researcher to use the tool and see what they get doesn’t work. However, if you know that there are five questions associated with opening up a sales lead or five elements of functional molecule groups that you always explore to get a new property, you can hardwire those questions. In early experimentation, you may provide loose guidelines and let people learn, but as they get more sophisticated, you should engineer the questions and contextualize them.

For example, some of McKinsey’s gen AI knowledge tools let us search our internal databases. Back in March, the prompt was, “Here’s our internal tool with a custom data set powered by a certain engine.” Now, the tools take a prompt and know that five or six other questions tend to be highly correlated with that prompt and they automatically push those questions to the engine to give contextual answers, as well as linking them to other workflows. But we have guardrails around what you can and can’t trust, with a focus on citations and source data.

Sean Brown: How do strong innovators deal with bad answers or hallucinated data?

Laura LaBerge: Cross-functional teams have always been important, but they are critical with generative AI. Remember that the aim of gen AI is to create new answers. In art, the tool learns from looking at images and then creates new ones. The same goes for literature and code. When you ask questions about patents, for example, or regulatory changes, you have to be careful not to ask in such a way that leads gen AI to generate an article that didn’t exist or a citation that isn’t real. If you don’t use cross-functional teams with broad perspectives who can spot things that don’t make sense, or you use forms of generative AI that don’t show the sources they draw on, you can end up with these hallucinations.

Another element innovative companies have to avoid these pitfalls is control functions seamlessly embedded into the workflows to help mitigate risk. Regulations regarding applications of data and gen AI are changing, so you want to ensure that teams experimenting with these tools are connected to those paying attention to regulatory changes and protecting your proprietary insights and data. You don’t want to accidentally make something public by using an open-access gen AI tool.

Matt Banholzer: This is an area where enterprise leaders can add a lot of value. As resource allocators or decision makers, you can say, “If we’re going to use gen AI, it won’t be five people in the IT department but a cross-functional team that includes some sales and P&L members.” You can also integrate control functions and feedback loops. Often, leaders say, “Let’s just have five interested people experiment because I’m not so familiar with this.” Instead, you should say, “I will lead from the front because if I do this right, we can have five to ten times higher odds of success.”

Sean Brown: How would you recommend companies approach investment in proprietary data to feed gen AI models?

Matt Banholzer: Very few companies apply gen AI across the company because gen AI without proprietary data doesn’t provide much insight. At the same time, you don’t want to overengineer the first use case by hardwiring many different data sets. Typically, companies take one or two use cases that might have a lower proprietary data load or can rely on one or two data sets linked together, then expand from there.

Sean Brown: To your earlier point about top innovators learning fast, is it better to learn from in-house experiments or to hire talent from the outside that already brings some expertise?

Matt Banholzer: When we looked into what drives high-performing innovation teams, we found that several traits matter. People tend to over-index on some elements, such as data science or developer skills, but softer innovation skills are just as important. They include having a bold vision and understanding where a new product or service can fit, skills at collaboration and being able to navigate an organization to rally resources, skills around continuously learning, and the ability to marry the conceptual to the analytical. You probably do need to bring people into the organization, but in a way that complements those skill categories. It’s also not about having your existing team do new things but thinking about the skills your team has and adding individuals with the skills you are missing.

Laura LaBerge: On the talent point, most of the top innovators have executive leadership teams with a much higher percentage of tech-savvy leaders than other organizations. As for agility, one of the biggest differentiators is the ability to be agile organization-wide. Top innovators are way ahead of others on this. Think of a bus: if one wheel is going at 200 miles an hour and the others at 20 miles an hour, you’re not going to get anywhere fast. Many organizations invest in technology or analytics in specific spots and get low ROI because the organization can’t act on the insights, or worse, it takes stutter-step actions that merely signal an opportunity to the market that others then capture. Don’t waste money out-investing in parts of the business that will accelerate beyond your organization’s ability to execute. You need to unlock the critical bottlenecks.

Sean Brown: How should leaders build these capabilities so the organization is ready to venture deeper into gen AI?

Matt Banholzer: Business leaders should think about ways to instill these practices. Can you run an experiment on a no-touch interaction, being mindful to establish guardrails? Can you make your budgeting process more periodic or with a metered-funding approach versus annual budget cycles? In our research on thousands of companies, the most innovative organizations have a quantifiable aspiration for what they want to get out of innovation. They allocate resources rigorously. It’s not done in silos—this much to M&A, this much to capex [capital expenditures], this much to R&D—but in an integrated way, and dynamically, almost like a venture capital firm would: you get some funding to deliver a set of proof points that then effectively gets you through to series A. Then they accelerate and derisk. And they’re fearless in learning, ensuring that noble failures are celebrated.

Sean Brown: What one piece of advice would you give leaders who want to get smart quickly on gen AI?

Matt Banholzer: Use it. This goes to the core of the agile approach. At McKinsey, we jumpstarted adoption because people built the tools and then saw the flood of use. Showing, not telling, is incredibly important.

Laura LaBerge: As business leaders, you can help set the direction for where in your business an acceleration could bring the biggest strategic distance. Where could the types of advantages and answers that gen AI can deliver help? There is a bit of a shiny objects syndrome with gen AI right now, but this tool is not appropriate for every type of question, so help your organization be thoughtful about where to deploy it.