Watch Salesforce’s Marc Benioff discuss how AI agents are transforming major enterprises

Keith Shaw: 2025 is shaping up to be the year of the AI agent. As companies continue to tout the benefits of AI-based digital labor and the automation of mundane tasks, is this the greatest tech disruption of the past 25 years?

We’re going to chat with Salesforce CEO Marc Benioff to get his thoughts—coming up on this episode of Today in Tech. Hi, everybody. Welcome to Today in Tech. I’m Keith Shaw. Joining me on the show today is Marc Benioff, Chair and CEO of Salesforce. Welcome to the show, Marc.

Marc Benioff: Hey, it’s great to be with you. Keith: So, let’s just get right into it.

You’ve written a couple of op-ed pieces in The Wall Street Journal and The New York Times, talking about the development of autonomous intelligent agents as one of the biggest transformations you’ve seen in the past 25 years. That’s saying a lot.

I mean, you and I have both been in the tech space for a number of years. You’ve seen a lot. I’ve seen a lot.

So what is it specifically about agents—not just generative AI, which was cool—but what is it about the agentic part that makes you say things like that? Marc: Well, you’re 100% right.

I’ve never been more excited about my job, what I do every day for customers, and the tremendous value we’re seeing across so many industries. The ability to have a positive impact with technology on so many people is incredible.

And yes, you’re right about the generative AI moment—ChatGPT, for example. It’s kind of like a super search engine. You can get some incredible results, but how do you translate that into business value? That’s where agents come in, and why I’m so excited about them.

We’re seeing customers achieve real impact. I just got back—literally—from Singapore, where I was working with one of our longtime customers, Singapore Airlines. You’ve probably heard of them. It’s “a great way to fly”—and they are a great airline. Their CEO, Goh, is amazing.

We’ve worked with them for over a decade. We’ve automated many of their functions—sales, service, marketing—and because we’ve managed so much of their data and they’re using many of our applications, we said, “Let us show you the magic.” And the magic was just turning on this agentic layer.

Now they not only have humans interacting with customers, but also digital agents. Watching it happen so quickly—and the value it created—was awesome. And Singapore Airlines is just one of many great examples. I think we are at the beginning of one of the greatest transformations in recent tech history.

Keith: We had someone from Salesforce on our other show, DEMO, where companies demonstrate their products. We saw a couple of demos of Agentforce. And here's a sneak peek—it hasn't even been published yet. He made me a fake customer and composed an email to me.

Then he said, “Watch this,” and translated it into French instantly. That was an "aha" moment for me. What would’ve taken 30 minutes or more happened in seconds. The speed is so impressive. Another example: one of our customers is Lennar, based in Miami—one of the nation’s largest homebuilders.

They’ve always had a vision of offering 24/7 wraparound customer service, cross-selling and upselling things like mortgages, appliances—everything you need for a house—with much greater customer intimacy. We’ve worked with them for about eight years.

They love our Sales Cloud, Service Cloud, Marketing Cloud, and they’ve built custom applications to manage their homes. They have a lot of their data on our platform. And just like you saw in the demo, they turned on the agentic layer.

Now there’s an agent from Lennar talking directly to customers. They loved it. They had attended our Dreamforce conference, went back to their Miami HQ, and ran a hackathon to explore the opportunity. They developed five use cases they’re coding out right now.

They expect huge cost savings and great value for their customers. This gets me excited. When a customer gets excited, I know we’re really onto something. We’ve seen the movies—our futurist Peter Schwartz co-wrote Minority Report. But there’s still a gap between the future and present reality.

Our job is to close that gap.

Keith: Since we’re on the topic of movies, when people think of agents, they often think of Jarvis in Iron Man. That’s the dream—an assistant who can handle tasks for you. Jarvis was an early example of home automation, which was kind of a nightmare in real life.

I wasn’t sure if you were going to say HAL or Agent Smith from The Matrix... Marc: It doesn’t matter. It’s about interoperating with the computer in an intuitive way, where it has a spooky understanding of who we are and what we want. Generative AI does that a bit.

Whether it’s Grok, ChatGPT, Gemini, or others, they’re pretty good, but sometimes they still get it wrong. Even this morning, I had an example where it didn’t quite do what I asked. These systems aren’t 100% accurate.

But when you get enough data and metadata, you can create a really great experience. Even when I use ChatGPT, I’m amazed by how much it remembers about me.

Image generation, too—it’s so good now that I get mad when I ask it to tweak one thing, and it redraws the whole image. Keith: Business examples help sell this tech. We’ve talked about airlines and homebuilders. Let’s talk about Disney.

Marc: Disney is an amazing company, but one challenge is that employees have to understand all their products—the parks, rides, cruise ships, hotels, Disney+, vacation rentals. It’s a lot. We’ve been running many of their customer touchpoints for years—whether it’s online or in-park with Disney guides using Salesforce and Slack.

But putting together the perfect vacation package for a customer with allergies, ride preferences, dinner reservations—it’s complicated. But AI is so good, it can do that. Let’s say you’re with a Disney guide, heading to Galaxy’s Edge or Millennium Falcon (one of my favorite rides).

If the ride breaks, instead of sending a Slack blast, the agent can instantly look at your preferences and suggest something else that’s running with a short wait. That’s a game-changer. Real-time personalization at scale. That’s powerful.

Marc: …So the AI understands all the products, what's working in real time, and what’s right for me—my preferences. That’s pretty cool. Keith: And it feels more proactive than just reactive. Back in the day, you’d check the app for ride times.

If a ride was down, it was up to you to figure out the next move. This kind of agent adds revenue ability too. It’s not just technology for technology’s sake. Marc: Exactly. It saves money, augments employees, expands revenue capability—it hits all the key points.

Whether it’s Singapore Airlines, Lennar, or Disney, those are three strong examples.

We have many more. I was just speaking with my friend Alex, CEO of Pandora. You might have seen their stores in malls or online. They make incredible jewelry. The company’s based in Copenhagen and runs thousands of stores globally. Their commerce is on Salesforce. Their stores are on Salesforce.

Sales Cloud. Service Cloud. And now, they’re using agents to create exceptional customer experiences. My dream? Imagine telling a salesperson in the store, “This customer already has these 10 charms—here’s the 11th, a perfect recommendation.” Delivering that kind of real-time value in-store? That’s going to be amazing.

Keith: It’s obvious you're enthusiastic about agents, but let’s talk about the other side—concerns and obstacles. When we speak with CIOs, we constantly hear about privacy and data security. They don’t want KFC’s secret recipe or Coca-Cola’s formula accidentally leaked. And hallucinations are still a concern.

So as we move from traditional generative AI into agentic AI, how do we build trust between customers and agents—ensuring they get things right? Say, for example, you tell the agent to buy me Giants tickets. But it buys Giants vs. Yankees.

The agent should know I’m a Red Sox fan, right? Marc: That’s a great question. Number one: we’re aiming for the highest level of accuracy possible. Right now, our Disney agent is benchmarking at 93%. Our own Salesforce agent is around 85%. Not 100%, and here’s why.

These large language models are word-based—they predict the next word using probabilistic networks. But they don’t have the multi-sensory input humans have. That’s why they’re not perfect.

Until we move to what we call “New World Models”—multi-sensory AI—we’ll be in this space of “good, but not perfect.” Now, regarding security and data sharing: Salesforce has a built-in sharing model. It knows what you can and cannot see. That same model applies to agents.

Agents operate within those guardrails.

You probably saw in the demo how the agent operates inside the Salesforce platform—not as an external tool. That’s the agentic layer.

It can see metadata—knows your phone number is a phone number—and because we manage 230 petabytes of customer data, we can offer a very high level of accuracy and security. Whether you’re an airline, bank, or homebuilder, you need that.

And yes, we want to prevent $1 flights to Singapore from being offered! Keith: You mentioned your agent is at 85% accuracy. Did that number surprise you? Marc: We were actually thrilled. Other vendors we benchmarked were in the 60% range.

So we’re excited to be in the 80s and 90s with many of our deployments. We’ve conducted over 500,000 conversations through our Salesforce agent on help.salesforce.com.

Customers log in, and the agent knows a lot about them—has access to relevant data and metadata. That creates magical customer moments. It doesn’t replace the human agent entirely. But at Salesforce, we’ve blended human and digital agents.

If a customer wants a human, that agent gets all the info on a single screen instantly.

Keith: Are companies nervous? Do they think they need 95% accuracy before committing? Marc: We’re very confident in the value we’re delivering right now. I have about 9,000 support agents at Salesforce and 75,000 employees overall.

I expect about half of those support agents can be redeployed into revenue-generating roles—SDRs, BDRs, and others—because we’ve already shifted half a million conversations into the agentic layer.

Keith: That leads to another concern—jobs. Your op-ed referenced a Morgan Stanley report saying customers are seeing 20–50% cost reductions, including headcount savings. That scares a lot of people. You’ve talked about support roles going down, but you’re also hiring AEs and focusing on reskilling.

Still, not every company is like Salesforce. Others might just see AI as a way to cut staff—and that’s frightening. Marc: You’re right. We have to address this head-on. Yes, some roles are being replaced by automation. So we need to adjust and be honest about what’s happening.

I just spent time with the CEO of a robotics company—these robots are already on assembly lines, in retail stores, and soon in homes. It’s coming faster than most people realize.

We need to take responsibility for reskilling. Some companies say they’re committed, but don’t invest in training. They lay people off and hire AI specialists instead. That’s not okay. Keith: So how do you see your role—as a CEO—in that space? Marc: I’ve always believed in retraining and reskilling.

We put 1% of our equity, profit, and time into our foundation when we started Salesforce. We’ve donated over a billion dollars—our biggest grantee is the San Francisco and Oakland public schools. It’s about education. We’ve also invested heavily in adult reskilling over the past 20 years.

Healthcare is another huge area where agents will help. I just heard from one of my technicians—she broke her foot and is seeing a doctor. But doctors and nurses are overwhelmed post-pandemic. Appointments are hard to get. An agentic layer can help triage, offer advice, and guide next steps.

It’s going to be game-changing in healthcare. I personally ruptured my Achilles last September. My local doctor wanted to do surgery, but I opted for regenerative techniques. I read Tony Robbins’ book Life Force, and decided to go that route. Six months later, I’m walking normally again.

Marc: Just think about this—what if the agent could help guide those decisions? Imagine it assisting in pre- and post-operative care, or in oncology. Cancer patients dealing with complex treatments like chemotherapy could have 24/7 support from an agent. That would make a huge difference.

Keith: I live in a rural area, where there’s only one orthopedic surgeon. So you’re limited to whatever that person knows. But what if that surgeon is augmented by AI? Suddenly, they’re the best orthopedic surgeon anywhere because the AI helps read scans, labs, and medical histories. Marc: Exactly.

And they can say, “Yes, we’re going to operate,” or, “No, let’s go the regenerative route.” That’s where AI really starts making a difference. Let me give you an example. I’m an investor in this amazing company called Artera. In prostate cancer, you can go in a lot of directions.

You have blood tests, scans, and so much data to ingest that even top urologists struggle. Now, top-tier doctors at elite institutions have intuitive instincts from years of experience. But local urologists don’t always have access to that level of insight. Artera got FDA approval.

It helps make your local urologist as good as the best. And it’s found that many patients shouldn’t receive treatment right away—they should go on active surveillance. That’s a decision that AI can now help make. They’re working on breast cancer next. This is cool stuff.

This is where we need to go. Whether we’re talking about Singapore Airlines, Lennar, Disney, UCSF, or the healthcare system—AI can make things better across the board.

Keith: Was your doctor pissed off when you told him you were going to try something different based on your own research—or maybe with AI? Marc: [Laughs] He’s 58, I’m 60. Great guy. He hadn’t really heard about the regenerative stuff before. We went over the scans together.

I had my ChatGPT open. I might also have a few bits of knowledge in my head—I’ve been studying regenerative biology for a while.

So I said, “Let’s consider a regenerative path.” He said, “I don’t know anything about that, so if you go that way, you’ll need to find someone else who does.” And that was okay.

But the idea that I didn’t need to get sliced open and stitched up—that I could regenerate on my own—that’s amazing. We’re not afraid of putting a Band-Aid on a cut and letting our body heal. That’s what we’re doing here. And yes, AI is going to help with that.

Keith: A couple more questions. Has generative AI been overhyped? There's talk of an “AI bubble.” Companies are spending too much, losing money. Will agentic AI help solve that by offering a better economic model?

Or will it make things worse, where VCs just keep pumping money in to stay afloat? Marc: Good for you for asking that. Here’s my honest answer: Number one, yes—when some of this rolled out, it was overhyped.

It should’ve been framed more accurately as the beginning of a very long journey. When Microsoft launched Copilot, it disappointed a lot of enterprises. Initially, developers liked it with GitHub, but then companies like Cursor or Surfboard started leapfrogging it. Technology is a continuum—it gets cheaper, better, easier to use.

That’s exciting. Number two, because of global chip restrictions, companies like DeepSeek had to move away from expensive transformer models. They shifted to Mixture of Experts (MoE) models, which dramatically lowered costs. Our internal benchmarks show that combining MoE with open-source technology—something we at Salesforce deeply support—changes the game.

We’ve contributed heavily to open-source AI over the last decade. Prompt engineering, for example, was invented by our research team. We open-sourced it through an MIT license, which means people get it mostly for free. So DeepSeek proved there’s another way—cheaper and more efficient. That’s what’s exciting about technology.

There’s always a new way, a better way, a next thing.

Marc: When I first started at Apple in 1984, programming on the original Macintosh in assembly language, I couldn’t have imagined where we are now. It’s light-years ahead. And in 30 or 40 years, we won’t recognize the world we’re in—unless we’re watching old movies.

We’ll say, “Wait… I’ve seen that before.” Keith: I talk often with Mike Bechtel, the Chief Futurist at Deloitte. We go back and forth on this: tech has both benefits and downsides. You have to acknowledge both—or you risk veering too far in one direction. Marc: 100% right.

Tech is never inherently good or bad. It’s what we do with it that matters. When we started Salesforce 26 years ago, that was our core belief. That’s why we created the 1-1-1 model: 1% equity, 1% profit, 1% employee time to philanthropy.

We built a company with a new technology model, a new business model, and a new philanthropic model. We’ve tried our best to evolve with technology and to live those values. Keith: You mentioned Clayton Christensen—his books The Innovator’s Dilemma and others inspired me too.

You even wrote a foreword for one of them, right? Marc: Yes. And that’s part of what makes this industry awesome.

Keith: Okay—one final question. You mentioned Salesforce has been around since 1999. According to my highly sophisticated research (Wikipedia and ChatGPT), you were at the DEMO conference in 1999. That’s our brand now.

I think Stuart Alsop started it, and Chris Shipley hosted it when I covered DEMO in the early 2000s. Marc: I remember! I was with Pat McGovern, the founder of IDG. We were at the airport afterward, in line to board a plane.

He said, “This seems very practical—I’d like to invest in your company.” He became one of Salesforce’s first investors. And I never missed a DEMO conference. Back then, you had to go to the conferences to keep up with the industry. Keith: You’re the second billionaire I’ve met.

The first was Pat McGovern. I worked at IDG for 20 years—wrote the “Cool Tools” column and the Holiday Gift Guide. He’d hand out Christmas bonus checks in person. Knew everyone’s name. He was amazing. Marc: He was one of the great people in the world. I miss him terribly.

He had such a unique brain—he funded massive brain science research at universities. Keith: So, when you look back at 26 years of Salesforce—is it the company you imagined when you started? If you had to start it today, could you? Marc: We launched in a highly constrained environment.

No VCs would fund us. We raised our money privately. People like Pat McGovern—who we now call angels—helped get us going. I’m forever grateful. I still think of Salesforce as a startup. I’m just a founder, an entrepreneur running a startup.

Marc: Last week, I had a three-day product offsite with our tech execs. Dreamforce is coming in October, and we’re planning the next version of Agentforce and four or five other big things. Tech is changing so fast. The opportunities are enormous. It’s awesome. We left that meeting completely energized.

Execution is the challenge. That’s what startups do—we have vision, values, plans, obstacles, metrics, and priorities. It’s a constant crisis of prioritization. Sure, we’re public. We have big numbers—$40.9 billion in revenue projected this year, and $12 to $14 billion in cash flow. But that’s not what I focus on.

What I focus on is customer success.

Marc: When I was just in Singapore and Japan, visiting our biggest customers and talking through the same stories we’ve discussed here—that’s what excites me. That’s what keeps me up at night, in a good way. Then I visited the founder of a robotics company.

His robots use specialized reinforcement learning models. I was blown away by the potential. And I thought, “Why didn’t I meet this guy two years ago so I could invest in the seed round?” We’re in an awesome industry.

We stand on the shoulders of people like Pat McGovern, and the pioneers behind Agenda, Release 2.0, and the DEMO conference. Most people don’t even know about all that early history—but it mattered.

Keith: That makes me feel a little old. When someone mentions a 1980s PC, I’m like, “Yeah, I worked on one of those.” Marc: Same. I was 15, working in a jewelry store in high school.

I’d walk across the street to Radio Shack, where they had a TRS-80 Model I. It was $499. I asked the guy how it worked—he didn’t know. So I figured it out myself. I learned BASIC. No teacher, no books—just trial and error.

Soon, we were moving into 6502 processors, Commodores, Apple IIs. Then the show was on. I started writing arcade games in assembly language when I was 16 or 17. Keith: My dad had a TRS-80 Color Computer, hooked up to a color TV. I programmed it.

Later, I wrote a dice-based football game for IBM PCs. I printed the code recently and thought, “Wow, Keith, you were really smart at some point!” Then I went to college—and college pushes that stuff out of your brain.

[Laughs] Marc: Now, you can just ask your AI assistant: “Give me five lessons on how to write in BASIC,” and it all comes back. Keith: [Laughs] Yep.

Keith: Marc, I’m way over on my time. Thank you so much. I hope we can do this again. Marc: My pleasure. Let’s do it again soon. We didn’t even get to Heathrow Airport or Pfizer. So many great stories, and so many cool companies doing amazing things with Agentforce.

Keith: When the next big thing comes—after agents—we’ll bring you back to talk about that too. Because it’s always evolving. Marc: Anytime. Always available to you. Keith: Thanks, Marc. That’s going to do it for this week’s show. Please like, subscribe, comment—do all the YouTube things.

Join us every week for new episodes of Today in Tech. I’m Keith Shaw—thanks for watching. (Editor’s Note: This transcript has been edited for clarity)