Pegasystems Founder: The Strategy to Eliminate AI Token Costs

And if you go listen to some companies, they talk about how you're going to have 6,000 or 10,000 or 20,000 agents. This is madness. They're talking about spending billions of dollars on data centers. They're going to want to get paid for this. By the way, the fact that they call it tokens is just an example of the BS that's going on. A token is .7 of a word. They could have called it words, but it's part of how they make this all kind of illusory and mysterious. Originally, you could see what you put in and you could see what you put out, and so you at least knew what you'd be paying for. Today, you don't see any of this till the bill comes. People need to understand it's not even like token costs are in their control. The thing you can control is when you use it, and there are styles of usage which let you apply AI as a business and to do it in an extremely cost-effective fashion. The key is figuring out how to—

I don't know if you can see my little button here. It says "No token cost." We run with such enormous cost effectiveness that we were able to go to our customers and say, "Nope, you don't have to pay for tokens." We took the approach that we primarily wanted to use AI to be able to really rigorously understand problems, be able to then design solutions of how you wanted to handle the different functions of your business. You want workflows that describe your business. You want agents that are capable of running any workflow. It both gives you better reliability, but critically it gives you predictability and predictable cost. They're actually thousands of times more efficient to run a workflow on a CPU than to run a reasoning session on a GPU. In the enterprise software space, you need to be able to change the way you handle situations as circumstances change. And all of that requires a level of visibility and transparency that is not typical of vibe coding type systems. And those are the problems that sink businesses. What builds trust and what builds true relationships is—

Well, thank you. I'm looking forward to it.

Well, the origin story goes back more than, well, four decades. I'd say it actually goes back to work I was doing right out of school as a consultant working for large New York banks doing software architecture work. And when I took a look at the way we did things, you know, it seemed like we were just working really hard to get the business people and the IT the solutions they needed. And I went and I visited with some of my friends who were in the manufacturing industry, which we all thought of as much less sophisticated and glamorous than banking. And what they were doing was so much more advanced than what we were doing. They were using this stuff called CAD/CAM, computer-aided design and computer-aided manufacturing, to really rethink the way that they actually design physical products. And I took a look at that and I said to my boss, "Why can't we do something similar to this for the parts of banking?" to which he instructed me that we were a consulting firm and should stay a consulting firm and I should just go do the project. But a couple years later, when the company was bought and I had to decide what to do next, I decided to branch out and see if I could build a system that would bring a lot of the principles of computer-aided design to the way you designed workflow systems in banks, insurance companies, and other firms. And that was what started the Pega journey. And if you want to fast forward through 40 years, I would say the thing that as a company that we're proudest of is that our first two clients who went live in 1984, if you can believe that, are not just customers to this day—they're Citibank and Bank of America—but they're massive clients of ours, and we're doing dramatic things for them and exciting new things across what's been five complete generational shifts in technology.

I was 26 when I started it. Yeah. So, the good news is I had worked for a couple of large Wall Street banks and had developed a bit of a reputation as somebody who could actually deliver. Nonetheless, I didn't realize till way later—I mean, it really took me years to fully appreciate—how much of a risk those first people who said yes was to this little startup that was going to actually do something pretty important in their business.

Yeah, I think what builds trust and what builds true relationships is to make sure you invest time in them when there's not an immediate return. You know, the difference between somebody who thinks that they're building a relationship and somebody who thinks they're in a transactional environment is that the people in the transactional environments only show up when there's a transaction. And sadly, the software business is full of people and organizations who are painfully transactional. And you can see that both in terms of the engagement, but you can also see that in the way that those relationships develop. So having organizations that we value, that we've been with, and that have stuck with us for long periods of time, I think really gives us good perspective, which you temper with the fact that everything is changing so fast.

"Don't go public too early" would be my initial lesson. We went public very, very early, and to be honest, I think we would have done a better job all around if we'd waited another five years. But the opportunity was there. It was 1996 and the internet bubble was brewing. When we went public, we were so advanced with our use of technology that when we went public, there was a color picture on what's called the S-1, which is the document you file with the SEC, that showed Pega running in a Netscape browser. The reason it had to be Netscape is, of course, Internet Explorer didn't exist back then. And the whole idea of running in a browser was kind of like, "Oh my god, how is that even possible?" Everybody was running on these big fat IBM mainframe screens. So we were able to put on a pretty good show, but we hadn't yet gotten the depth of business and breadth of business and depth of management that I would recommend today that somebody get before they go public.

I was mid-30s. It was a pretty wild and dynamic time. What basically happened is that some other companies sort of in our space—so we were in the customer management space, we were in the workflow automation space—and some other companies (people will probably not remember these names because they're all dead companies, like FileNet and a company called Siebel and a company called Scopus, now all sucked up into Salesforce and Oracle)—several of those had gone public. And I went to my board of directors and I said, "Should we consider doing something like this?" And they said, "Well, why don't you go down to Wall Street? You spend a lot of time there. Meet with Morgan Stanley and Goldman Sachs and see what they say." And I walked in thinking that we would get some advice on how to do something in a couple years, and they saw something they could bring public immediately. So by the middle of that year, we were public. And it was a bit of a heady ride. On one hand, it's an enormous affirmation. We had three of the world's premier investment banks vying to take us out. So we had Goldman, a company called Montgomery Securities, which is part of Bank of America, and Cowen. And it was just unbelievable. But frankly, when you're very small, you don't necessarily have the full perspective. Our total revenue that year was $32 million on a trailing 12-month basis. And when that small, everything is constantly up in the air. So in hindsight, I do think some of this rush to massive growth and adopting massive funding—I know blitzscaling is a thing, but I would tell you, I think for most companies, it doesn't work out very well. And for us, we were able to navigate it, but it led to very hard times.

Well, some of what's going on in very small sectors of the market—people need to understand that there's kind of a frothy sector and then there's the mainstream business market. And I think you need to decide if you want to position yourself in the froth or if you want to position yourself in kind of a more marathon-like trajectory. And candidly, we're happy to be froth-adjacent, but I really wouldn't want Pega to be in the middle of this vortex. There's so much that's going on that—look, we'll do close to $2 billion of revenue this year. We've got over 5,000 staff. We've been able to do great things for some pretty amazing customers and build those longstanding relationships with them. Running for the froth, I think, also means that the froth can leave too. And some of what you see going on in the market—you look at the whipsaw events. This morning there were massive declines in semiconductor stocks that four weeks ago were considered to be the only thing that mattered in the market here. I'd rather try to pick product. We're a product company, and we like living products that are going to have sustainable value and that are going to be able to operate in, I would say, a more understandable and rational way, and let us develop that sense of accomplishment and pride that comes from that.

Well, maybe you wouldn't, but boy, they sure look like it's exactly what they want to do. I think certain personalities are drawn to that sort of vortex of attention. And candidly, they both are drawn to it and then the things they do feed it. And the media and the podcasters all seek to amplify it so you get to hear it. I think the real question is going to be: how does this settle out in three years?

I think the reality is that a bunch of what we're seeing is, I would describe as bubble-icious behavior. And so there is going to need to be some reallocation, and there will be corrections. The core value proposition of AI is very, very, very strong. It is, on one hand, magical, but not magic. So a lot of the beliefs that people are putting forward about how this is going to just destroy employment, etc.—there's a lot of self-serving behavior that's going on. I loved when Jensen Huang told everybody that if you weren't spending a quarter-million dollars a year on tokens that you weren't taking advantage of your good engineers. They've obviously got enormous self-interest in that happening. And I think as people grow a little more familiar, they'll be able to make a little more reasoned assessments of how to use the goodness that's here and how to organize it in a way that makes sense. That's been the basis of what we've tried to do as a business.

I focus on building the company. When the stock jumps massively, I think anybody who overly celebrates is just setting themselves up for the subsequent disappointment when it goes down. The real question is: can you provide long-term value for customers? Can you build relationships that are going to persist? And can you then deliver product that evolves with the massive changes that happen in the technology market? Candidly, the fact that we've survived and actually done pretty darn well all these years says something really about our ability to evolve. And I think it's kind of ironic that 35 years ago, when we chose a tagline for the company, we were thinking of a tagline that would make sense to our customers. What is the trademarked way that we want them to think of Pega? And what we came up with was the phrase "Build for Change." We wanted to tell organizations that writing in code—yeah, you can write stuff in code, but if you write things in code, you're not really set up to evolve it very well. But if you do things using the sorts of rule-driven workflows that Pega does, things that let business people and technologists come together and implement change at a crisp and evolving rate, that you, Mr. Customer, will be able to build for change. I don't think I realized at that time how much we would have to build for change given the massive shifts that we experienced in the technology and that we're still all seeing.

Yeah. It was interesting because we had been involved—I'll tell you just a funny aside. When I was in university, obviously a long time ago, neural networks were considered a failed technology. We experimented with them. We played with them. We tried to get them to do character recognition. But it was only in the 2014 time period that this idea of being able to build these generative-capable systems even began to really catch hold. But in 2010, we decided that we had a super powerful workflow system that could—if people wanted to apply for a loan or if you wanted to onboard a customer, it was just a great system to put the controls and the management and the intelligence about how to do that. We decided we wanted to add AI to it. I had been personally very involved with AI because early in my college career, I taught computers to play chess. That was something that I had gotten involved in. But we went and we actually acquired a company that was a leader in statistical AI. That's what we used to call machine learning, where you would actually look at patterns in numbers and decide how you could use that to do future steps, future pieces of work. And from 2011 on, we built very powerful statistical AI into our workflows, which led to great stuff. When we saw GPT come out, as you said, just four years ago, it was kind of mind-blowing because it was so much more advanced than what we had kind of expected. We had sort of expected that thing might be out in 10 years, but it was there, and we saw a real opportunity to apply it for our customers. But we were also a little nervous about where this was going to go and what it might mean from a cost perspective, and also wanted to make sure we applied it right, because lots of times you can have a really shiny new technology and easy to apply in ways that might not be the optimal. So we took a very contrary approach to how to use it compared to, for example, Microsoft or Salesforce or ServiceNow. We took the approach that we primarily wanted to use AI to be able to really rigorously understand problems, be able to then design them, design solutions, sort of build the recipe of how you wanted to handle the different functions of your business, and do that at a design-time effort, which is a real collaboration with AI, but at runtime, as much as possible, try to revert to workflows. Because we have this enormously powerful workflow engine and we said workflows are reliable. You trust them. You can understand what they're going to do before they do it, not just after. And they massively scale and they don't kill entire forests of trees by eating up too much electricity. They're actually thousands of times more efficient to run a workflow on a CPU than to run a reasoning session on a GPU. So we said, "Let's burn those reasoning tokens when we really need it." But once we figure something out, let's record it. Let's put it in our recipe box. And then when a customer wants to do something, you figure out if you can dip into the recipe box, pull it out, and actually do something you've already figured out how to optimize. And I think that is the absolute right way to tackle this problem. And it's really quite different.

Well, it's interesting. I don't know if you can see my little button here. It says, "No token cost." We announced at PegaWorld in June that our uses of AI for our customers, the customers don't pay for tokens because—let me give you the analogy. Let's say you wanted to open a restaurant. You might go and hire a great chef. You might create a test kitchen. You might go into that test kitchen and you might create the collection of recipes and then decide which you want to put on your menu. When a customer comes in and they're going to order off the menu, you might tune it up a little, but basically you're able to scale and know what you're doing. That's our philosophy with AI. We want to be able to use the tokens—and we do use a lot of tokens—to do the design, to figure out how to make the best dishes. But then when you then serve those dishes 10,000, 100,000, a million, 10 million times, you're not reimagining how you want veal scallopini to be. You're actually using something that's already been determined. It both gives you better reliability, but critically it gives you predictability: the predictable outcome you'll get by doing something that you've had a chance to exercise in advance, and the predictable cost. We run with such enormous cost effectiveness that we were able to go to our customers and say, "Nope, you don't have to pay for tokens. We got that covered." And the reason we have that covered is because of the way we architected this to use the full power of AI, but to use it upfront and then very selectively. We do need to use some at runtime, but we don't use it to re-envision everything you're trying to do. We just use it to summarize or to do a little translation. And the main heavy lifting comes out of what you did at design time.

It was interesting because when we saw what OpenAI was doing and they were kind of giving it away, and then you got Anthropic, and maybe it's $20 a month, which doesn't remotely cover anything meaningful. But it struck me that these guys are a little bit like drug dealers. They were out there passing out packets of tokens and getting people hooked on the very cool stuff. And inevitably, when you have an abundance of supply and no consideration of cost, you make decisions that take advantage of the supply. And people got sucked in. They went hook, line, and sinker. And we did a little bit of math and we said, "My god, they're talking about spending billions of dollars on data centers." Now, that was before they began talking about tens of billions and hundreds of billions and now a trillion dollars on data. They're going to want to get paid for this. They're going to have to get paid for this. And we said, "So, how do we make sure we use this as if it was paid for?" To tell you the truth, when I think about things that I think we did well, I think that style of decision-making—yeah, let's use it, let's use it assertively, but let's use it as if we were already paying for it. And lots of organizations didn't do that, and those are the ones that are getting whipsawed. And candidly, it's just going to get worse. What a lot of people don't realize is that when you originally asked ChatGPT a question, you give it a question and it'd give you an answer. You could see the number of tokens you used to ask the question and the number of tokens that came out. By the way, the fact that they call it tokens is just an example of the BS that's going on. They could have called it words. A token is .7 of a word pretty much on average. They could have done it in words, but then people would understand it. Instead, they call it tokens. They create a level of abstraction. It's part of how they make this all kind of illusory and mysterious. But anyway, words, tokens—they're all just units of counting of the size of the text. Originally, you could see what you put in and you could see what you put out, and so you at least knew what you'd be paying for. Today, all these foundational models, including the ones we use, they take reasoning steps. So you'll see them thinking or sometimes they'll tell you that they're frolicking or they're reasoning. And what's happening then is those models are generating tokens to themselves, and then sometimes they'll spin off sub-agents that go and do pieces of work, and they talk to those sub-agents with tokens, and those sub-agents talk back to the parent agent with tokens. And sometimes you don't see any of this till the bill comes. So people need to understand it's not even like token costs are in their control. The thing you can control is when you use it. But I think there's going to be a lot of pain in this area before it settles down. There's going to be a lot of people chasing fantastical next models where the Chinese open-weight models somehow make us all cost-effective. The reality is there are use cases for which this is already incredibly cost-effective, and there are styles of usage which let you apply AI as a business—in our case, apply AI to your workflows as a business, which, when I look at businesses, I see reams and reams of workflows that describe what they do—and to do it in an extremely cost-effective fashion, which we've chosen to take the mystery out of by saying, "Hey, there won't be token charges."

I don't see how they would do that. I think it would be pretty radical for the government to step in and influence private companies' pricing models. I think the thing that really influences pricing models better than anything is competition and judgment. And I do think that competition and judgment will prevail. That will be how people put some of these genies back in the bottle.