Approaching generative AI with a beginner’s mindset

Applying a growth mindset is the best way for businesses to approach the rapid changes born of generative AI (gen AI), says Clara Shih, CEO of Salesforce AI and today’s guest on this episode of the At the Edge podcast. Shih speaks with McKinsey senior partner Lareina Yee about the transformative power of gen AI to help accelerate workflows, the importance of change management, and the top questions business leaders are asking about AI adoption.

An edited transcript of the discussion follows. For more conversations on cutting-edge technology, follow the series on your preferred podcast platform.

Unlocking unstructured data with large language models

Lareina Yee: Clara, over the course of your career, you’ve pioneered how companies connect with customers in more intimate ways and create personalization. With all gen AI’s possibilities, what is it about large language models that unlocks new possibilities?

Clara Shih: The biggest unlock is looking at all of the unstructured data that’s out there on the internet, as well as the proprietary unstructured data that’s inside companies. If you think about the history of computing, data has been limited to what we can put into a database. And that amounts to less than 20 percent of the data companies have.

There are all kinds of data—videos, phone transcripts, customer chat transcripts, messaging, emails, calendar meetings. And the ability of large language models, as well as large vision models, to analyze, summarize, and generate this type of unstructured data is game changing for knowledge work.

Lareina Yee: Can you give us a little bit of a technology primer on how you harness that unstructured data and how that connects to your work today?

Clara Shih: From a software architecture standpoint, it’s very simple. It’s the model–view–controller [MVC] tech stack. The traditional data model was limited to structured data. But now, it includes unstructured data as well, with metadata to help organize it all.

Then you’ve got the view layer. This includes messaging platforms and self-service digital-engagement channels that people use for work today.

Traditionally, the controller was code. But large language models now allow us to go from code and clicks to conversations, using natural language to create and invoke this business logic on top of the data.

Lareina Yee: How do you use generative AI in your own work?

Clara Shih: I use generative AI every day. There’s so much happening in the world of AI, so I have a pretty neat system that I’ve put in place with my assistant where we’ll use gen AI to summarize and synthesize all of the top AI podcasts out there, including yours, as well as research papers.

I usually triage which ones seem the most interesting and relevant to what we’re trying to do here at Salesforce. Then I’ll use gen AI to transcribe the whole thing and watch select clips. So that’s a way of saving time while consuming a large amount of information.

The power of simplicity

Lareina Yee: As you’re doing that, how is that thinking changing some of the ways you dream about the capabilities you can bring to your products and services?

Clara Shih: A lot of power comes from simplicity, as well as giving more capabilities to those expert users. That’s the approach we’re taking in our core Salesforce CRM [customer relationship management]. So for customers who just want to get started, end users like small or medium-size businesses, we have these powerful, simple, turnkey use cases.

For example, when a customer email or chat comes in, our large language model suggests a response grounded in the company’s officially sanctioned knowledge articles and product documentation. It’s able to cut the average resolution time on customer support issues by 10 to 20 percent. This is what we’re seeing at companies across industries and geographies.

And when the customer is ready for more, or for large enterprises that have sophisticated data science or machine learning teams, we have our AI platform. At the most basic level, they can customize the prompt templates that power these turnkey use cases to really tailor it for that organization. Then we’ve got our Model Builder. So for the companies that want to pretrain or fine-tune a model, they have a lot of data that allows them to tweak a model without code or to use their own model.

We’re also seeing that as companies deploy from pilot to production, they’re realizing they might not need the most expensive, state-of-the-art models like Claude 3 or GPT-4o every time. They’re fine-tuning their own smaller, sometimes open-source, language models to run faster, more efficiently, and at a lower cost so that they can scale.

Lareina Yee: In a world where we’re always in such a hurry, what is the onboarding process as we implement these models into our enterprises and companies?

Clara Shih: Most customers are just trying to prove out their first set of use cases and drive ROI from that. That’s why those prebuilt, GPT [generative pre-trained transformer] features that we’re working on are so important, such as with our Financial Services Cloud, where, for example, you can generate a client dossier before you walk into a meeting to discuss a financial plan with a client.

A time-saving game changer

Lareina Yee: Where do you see the future of software? How are we repacing or resetting our expectations and standards?

Clara Shih: The change management involved with this is harder than the technology itself. It’s not just the CIO [chief information officer] who’s trying to convince the rest of the company to try generative AI. It’s every employee wanting to do something, being curious, and seeing this as big, if not bigger than the internet and cloud computing were a decade or two ago.

The companies that have spent time bringing their data together, that have a data warehouse, a data lake, and everybody in their contact center using that data have a huge advantage. We’re seeing the companies that haven’t organized their data accelerate those foundational projects, because they know that’s what’s required to deploy AI securely and at scale.

Lareina Yee: You said earlier that AI promises to fundamentally change how we work. Can you tell us a little bit about the journey from a use case to changing workflow?

Clara Shih: Previously, when a customer was on the phone or chatted or messaged in a question—they may have had an issue with a product that they purchased—the service rep working with the customer was on the clock, trying to match up the customer’s setup, warranty, and entitlements with the right knowledge article.

When they found the answer, they would be copying, pasting, and frantically formatting it in a way that was appropriate to send to the customer. Then they would send it, hoping the customer wasn’t angry for having waited so long.

But with gen AI, all those steps I just described that would traditionally take several minutes, if not hours, can happen in an instant. So customer wait time and resolution time have gone down, which means that companies really can do more with less.

No job left untouched

Lareina Yee: If I’m thinking about my everyday consumer experience, I’m excited to be served faster and better. But if I’m an employer or a customer care rep, is this the end of my job? How do I think about the opportunity here?

Clara Shih: That is such an important question. Every job is going to change. Everyone is going to have to write a new job description, because so many of the tasks that were time consuming before can be automated or accelerated. It means we have to learn how to leverage these technologies to do our jobs better and more efficiently. We also have to discover what more can be done now that this time has been freed up.

Lareina Yee: So if I’m a call center rep or a sales rep and I say, “I want to be tech forward, but I worry about my job and want to invest in the type of skills that I’m going to need in the future,” what types of things would you suggest for those people?

Clara Shih: Every knowledge worker, whether you’re a salesperson, marketing manager, or customer service rep, should learn how to prompt so they can interact with large language models effectively.

You’ve also got to learn AI safety. For example, just as you would never enter your social security number into some random website you don’t know, you have to be really careful about what kinds of data you put into a large language model.

People also need to recognize that AI is constantly evolving. It’s not like you learn prompting once and then you’re done. Because the models are improving at an exponential rate, you have to keep practicing and improving as time goes on.

The four Ds of AI risk

Lareina Yee: What gives you optimism and what gives you concern at this moment of inflection for humanity?

Clara Shih: What gives me optimism is that, as we’ve seen with previous technology disruptions, it usually opens up a world of opportunity. People who traditionally didn’t have access to certain resources, education, or certain skills can now participate in this new economy and these new possibilities. I think that’s really exciting.

And there are so many unsolved problems in the world. Certainly, in the business world, all of us want to grow faster and run more efficiently. But if you look at healthcare or education, being able to scale those services to a much broader part of the population just wasn’t possible before. Then there’s the climate question, and new drug discovery. There are really endless possibilities for applications of these new technologies.

What gives me pause is that with every major new technology, there come new risks. Sometimes, those risks are anticipated up front. But oftentimes, there are unknown unknowns. Certainly, that’s what we found with the internet and cybersecurity.

I think in the AI realm, that’s going to be true, as well. One of the frameworks I like is thinking about risks in terms of the four Ds. The first D is data privacy, data security—both consumer data privacy, as well as businesses making sure that their data doesn’t leak outside of the company or into a model. Within any organization, different people have different access to different data, so you need to make sure that gets honored with any type of AI copilot or interaction.

The second D is disinformation—the use of gen AI to generate misinformation and disinformation, which is especially important to consider, given the US elections coming up. We saw misinformation and disinformation in the last election, and I think the risk is even greater now, given how freely available these AI tools have become.

The third D is discrimination and the fact that these large language models are trained on the corpus of consumer data on the internet. There’s a lot of biased, toxic, racist, and sexist content out there.

The fourth D is really a longer-term concern, which is displacement. Looking back, some jobs were replaced with the internet, but most jobs were completely transformed. Imagine being a salesperson today who doesn’t know how to use a search engine or email. I don’t think you’d be able to find a job. The same is going to be true with AI skills. But there are also going to be new jobs created that we can’t even fathom today. That’s what happened with the internet.

Learning to see a digital bot as a coworker

Lareina Yee: Clara, you’re on the forefront of leading with technology to promote change. If we were to zoom ahead ten years, can you paint us a picture of how sales, marketing, and service might be different?

Clara Shih: When we talk about sales, marketing, and customer service departments today, we’re dealing with our customers within our organizational structure. And customers don’t think that way.

We, as customers, have products and services that we buy and then have questions about. We’re interested in upgrading, getting something new, getting cross-sold. We don’t expect or want to be bumped from department to department to have those needs met. We just want someone, whether that is an AI or a concierge, who knows everything about us, what we need, and everything about the products and services of an organization. So the goal of customer-360 vision will be actualized by AI.

Lareina Yee: We’ve been talking about personalization for years. With generative AI, our ability to meet people not as segments or groups but as individuals is radically different. And that’s going to change how we market, how we sell, how we respond to questions.

Clara Shih: Yes, I think that’s on the customer side. Organizations are also going to fundamentally change from being organizations of people to organizations of human and digital workers working together.

Lareina Yee: Can you tell us a little bit about what that looks like? Part of me is excited, and part of me is scared.

Clara Shih: We already have this in the form of self-service bots that many of our customers have deployed both on the customer care side as well as for commerce. And you can think of a bot as a digital coworker. In the past, you would have had someone servicing a customer on the phone or on chat or through messaging or email, helping guide the customer through a transaction. These bots have gotten really powerful, especially now that they’re powered by large language models. So that’s an example of a single digital worker.

But think about having multiple digital workers handle different types of tasks in different domains where they might have RAG [retrieval-augmented generation] on different data sets, which increases their focus and accuracy.

And then, think about having some of these AIs work with other AIs, just like they would with a human. Humans are always at the helm, supervising, collaborating, and calling the shots. But now, they’re able to scale and do much more in any given amount of time.

Top corporate questions about AI

Lareina Yee: Clara, in addition to building the technology, you’re also talking to many of your customers. What things are you hearing from them?

Clara Shih: There are three recurring themes that I’m hearing when we talk to our customers about AI. The first one always is around trust. Our customers, regardless of whether they’re CIOs, chief revenue officers, or CMOs [chief marketing officers], recognize that there are new risks inherent with large language models. So we talk a lot about trust.

Once customers get comfortable from a trust perspective, the next questions are always, “What are the business outcomes that we’re going to drive? These generative AI costs add up, and so there better be a there there. How are we driving efficiency? How are we cutting down the average resolution time of a customer service case, increasing deflection, increasing customer satisfaction, driving higher conversion rates, and reducing the sales cycle?”

Talking about business outcomes is very important. There are short-, medium-, and long-term types of impact here. So we have a lot of conversations with customers about each of those time horizons.

And then, the third one, invariably, involves the topic of people. “How do I bring my entire organization along? How do I get my leaders on board thinking this way?” And that is a tremendous change management challenge that we spend a lot of time on.

The paramount importance of building trust

Lareina Yee: These are three fantastic questions, Clara. If I could just unpack the first one, around trust, what advice do you have for executives looking at different types of software and partnerships?

Clara Shih: So much of AI is around data. And so the top trust questions are, “Where is the data housed? How is it being used? Is it being learned by the model? Is there a risk that the data can leak out of the organization? Are the organization’s internal sharing rules being honored by the AI?” Those are all really important questions.

And on the consumer data side, we get asked, “How are we protecting our customers’ data? How do we make sure there are ethical guardrails, knowing that these models are trained on the corpus of data on the internet, with all the toxic, inaccurate information out there?” So we need guardrails to improve accuracy and relevance.

Shattering the STEM barrier

Lareina Yee: Clara, one of the things we’ve seen across all these technology transitions is that women tend to be high users and fast adopters. But we see very few women actually developing and leading the technology. You’re one of the few. Can you tell us a little bit about your experience in being one of the few female technology pioneers and how you hope to encourage more women to have a seat at the table?

Clara Shih: This is a dialogue that has started to gain more steam in the last few years, and it is long overdue. I think there are not enough women builders and leaders in tech because of the way that we socialize girls when they’re growing up and because STEM [science, technology, engineering, and mathematics] education can be very intimidating.

We see a lot of middle school girls who start to feel bullied and intimidated. That’s when they start to shy away from those harder STEM courses in high school and then STEM careers. But there are interventions starting to take place.

There are some fantastic organizations out there, like one I’m involved with, Girls Inc., that are trying to intervene and help girls think of themselves as future technologists, future scientists, and future inventors so that we create a healthy pipeline of tomorrow’s leaders.

Capturing AI potential with a ‘beginner’s mindset’

Lareina Yee: One of the things you’ve championed is a “beginner’s mindset” in how you think about growth and leadership. Can you tell us a little bit about your philosophy on that?

Clara Shih: It’s about having a growth mindset and not feeling like we ever become expert enough to never question first principles or rethink assumptions. In this world that is changing so rapidly and where the pace of change is accelerating, all of us are beginners at all times. And so to come with that fresh and open mind and look at possibility, I think that’s how we’re able to capitalize on major disruptive moments like the one we’re living in right now.

Lareina Yee: You were a Marshall Scholar at Oxford, where you majored in internet studies. At the time, that was very pioneering and new. If you were to go back, as one of these very resilient young women with a growth mindset, what would you study?

Clara Shih: I’m really proud of what I studied, which was e-learning. At the time, Hong Kong had just been through SARS, and it was really an experiment on whether kids of all ages could actually learn through the web. And it was a test of the very limited capacity of the technologies back when I was doing this.

And, ultimately, it didn’t directly affect my career. But that time off, being in the Department of Education at Oxford, going for long walks on campus, reading books that had nothing to do with technology, thinking about how society affects the technologies that get created, and how those technologies in turn shape us, was personally very rewarding. But I think it also allowed me to start to really develop that beginner’s mindset toward everything else that I’ve done in my career.

Lareina Yee: On that inspiring note for young, enterprising entrepreneurs out there, as well as for executives who are trying to apply a beginner’s mindset to this, I would just like to say thank you, Clara, for joining.

Clara Shih: Thank you for having me and for the great partnership between our companies.

McKinsey has a long-standing alliance with Salesforce to drive impact together for clients.