PegaWorld 2025 Alan Trefler keynote: Enterprise Transformation with AI: Powering the Autonomous Enterprise

Hello, Pega. It is so wonderful to be here again. And it is such an exciting time. We're at so much of an intersection between technology that would have been considered unimaginable years ago and and the needs of businesses which are more profound than ever. And we are honored here at Pega to have so many tremendous people, tremendous brands, tremendous companies joining us here to really explore the possibilities, to really understand what we can do together and how this partnership, which from my point of view has been just an amazing experience for, in effect, my whole adult life. How this partnership can continue to promote the best in both of us. You know, when I was thinking about our group here of of innovators and creators, one of the questions I had was how do you name them? You know, if you go to ChatGPT and you say: Hey, what do you call a group of flamingos? And it will come back to you and say, hey, they're called a flamboyance of flamingos. So that was taken. Couldn't do that. Well, you say, okay, let's go for something probably closer to our our customer demographic. Let's go to the really wise and knowledgeable and occasionally stern owls. What do you call a group of owls? And it's a parliament of owls, which doesn't sound very cheerful anyway, so we're not going to do that.

So I went back to ChatGPT. I actually do this a lot. I'm in all the AIs, the Anthropic and the Gemini and ChatGPT. I'm on all the time. And I went back and I fed in information about this customer base, the people we will be drawing together, the people who stimulate us and we try to work with to really push the boundaries. And I said, what do we call this group? And you know what it came back with? And I thought this was prescient. AI calls you a constellation of creators. So join me here in creating this enormous shift in technology that is going to be mind blowing. Now, one reason it's going to be mind blowing is because of all this discussion of AI and because of the magical marketing introduction of agents that have come in and it lends itself to the question, what do you want to call a group of agents that might come in to do stuff for you? I mean, they they look friendly at first, but to tell you the truth, they could become a scary, disorganized mess. And the sad word that comes to mind here without the right controls is cacophony. A cacophony of agents doing stuff. Because God knows there are some people who tell you that you're going to need thousands of them, thousands and thousands of agents driven by prompts doing things. Now, of course, thousands of agents driven by individually written prompts are all going to do the exact same thing together. They're all going to do it right. Yeah, not so much. So our view here is that this is wonderful. There's an enormously powerful set of ideas here, ideas that have to be harnessed for this next generation of business. But what we see, these folks who want to do it with prompt engineering, we just see them doing it the wrong way and leading to unpredictability.

And I was thinking about this, and I was brought back this just last week to my earlier existence. Now I will tell you that my earlier existence has a couple of unusual features to it, one of which is I used to play chess a lot, and in fact, just last night I made the cover of Chess Life and Review. Thank you. Thank you. Rolling Stone next! Cover of Chess Life. And, you know, as if you were to go find my 15 year old self who sort of looks like the guy on that side there, I would have never viewed that this would be something that could be possible. But the reality is... Oh, by the way, my wife, when she saw this picture, said: Thank God I didn't know you before you had a beard. 31 year marriage might have ended just like that. In fact, there's only one other person in this area who has seen me without a beard. I have carried this as my shield and sword for many, many years, but not notwithstanding. You know, I know a bit about chess. Not so much currently, but historically. And I still play, as you'll come to Tuesday to the simul. And one of the things that comes out of that is people know that I know about chess and the Financial Times, which I subscribe to, one of the places I go for more of a worldly view of what's going on in the news. They sent me something just this past Tuesday, and it was a really interesting chess problem. It's a chess national problem solving event for the UK. So anybody from the UK, you can't use what you learn here to respond to this or don't please.

Anyway, it sent me this problem. And the problem was, can you do a checkmate in two from this position? Now I'll tell you, in this position white is really winning. But a checkmate in two means: Hey, there's some decisive words. We can get after the king. We can go finish this off in just two moves. And I'm going to tell you, this is a really complicated position. This is a hard one. This is worth having a chess problem. And so I looked at it for a minute or two and I couldn't see a way to force a mate in two regardless of what the other person did. So what did I do? Well, I cut the picture because, you know, all these things are multimodal now. I dropped it in to the ChatGPT-4o model and I thought my work was done. So let me show you the actual dialog between me and ChatGPT-4. Goes something like this. I go, all right, here's the position shown. White can deliver mate in two. Here's the solution. It says move the knight to A5. Check. This is a check from the knight. Black has only one legal move. So I look at the board and I say, okay, that's the knight moving to A5. But it's not a check. It's not a check. So I type that back. Na5 is not a check. Now these are learning systems. I'm sure it's just gathering information. It says, oh, you're absolutely right. Well, that's good. But it's going to reevaluate the board carefully. I don't know what it did the last time but carefully this time. And it's going to say: Correct solution. Rook takes c4. White plays the rook, takes c4. Sounds very powerful. So we go, rook c4. And I go, what? And right back to it I say, but rook c4 is not legal because that is a white piece. I didn't actually say that you have just stomped on it. Turns out it can figure that out. And it goes: Well, what's it going to do now? It says, oh, it's obviously very grateful to the reinforcement learning that I'm providing. I'm going to correctly analyze it. Correct mate in two. Knight a5, check. Hold on. I thought we saw that already. And I said... this time it reassessed it actually, which is frightening here. And I went back but it's not a check.

And so it came back and it said, okay, okay, let's carefully and accurately solve the problem this time. It's going to really try. It's going to say, rook to d3, check. At least a different attempt. So it takes that rook, moves it to d3. Right. And the check comes from the bishop on the second row. Sounds possible. One little problem. The king is in the corner and is now attacked by the queen. And you're not allowed to make a move that puts your king under attack. So once again, it's just another completely illegal move, and it goes back to the beginning and says, you're absolutely right again. We're going to do no incorrect assumptions. Correct move, knight to A5, check. So I love the way it reasons. I wish I had played against this in many tournaments. In all seriousness, you know, this was just a little sobering interaction, comparable to one that I have all the time as we do the work that causes us to think and design. But the problem is not that AI is bad. It's unbelievable. The problem is you have to use the right AI for the right things. So what I did is I popped that exact same position into this program called Stockfish. Stockfish is an AI, been worked on for a long time. It has its roots in architecture, closer to how I got into AI. It doesn't use language models. It uses a combination of A/B searching and rules and a whole variety of different techniques. But drop the position in...that's, by the way, their logo. Stockfish. They obviously don't generate images or they would have come up with something better than that. And it declared that here, there's a checkmate in two. And it found the move, which is a really subtle move.

It doesn't take anything, doesn't check anything, just positions the queen. Turns out black can't do anything but immediately do something that leads to mate in one. These are really the creepiest chess puzzles that you have here. But the important thing is it found it in under a second and it was right the first time. And it would never do something stupid because it's the right AI for the right purpose, which is absolutely critical and candidly, largely forgotten by the pundits who just want to dump things in and hope that everything goes right. So what do I take from this? Well, first thing I take is as we knew language models aren't everything. Language models are great for some stuff, but for other things you want to use other forms of AI. But second, as you think about how you want to run your business, you need a conductor. You need a way to direct the business, to use the right AI, or use a person, or do the right things at the right time. And as much as musicians work in professions from a score, you want to be able to have sheet music that describes the workflow, as it were, that the music should go through. Because if you have that, you can take advantage of all of this power. But make sure that before the audience shows up that you're synchronized, that you've got a symphony, that is all going to make sense. And that, to me, is the metaphor for finding the right AI and putting it in the right place. And at Pega, what you're going to see is lots of AI elements, but we're going to spend some time trying to explain what the right places are, because I don't think it's well understood. And I think people must understand it, because otherwise the parlor tricks can be overwhelming. So what does that mean pragmatically? Well, we think the right way to conduct your business, the right way to design the way you want to operate is through Blueprint.

Blueprint is a set of agents that are creative, that can reason, that can challenge you, but they're doing it at design time. They're doing it as you're planning what your musical score should be. They're doing it as you're planning, not when you're engaged with a client. And that is such an enormous distinction that is sometimes missed because there are use cases where this doesn't matter. If you've got a knowledge worker, you're building an investment banking proposal who knows everything there is. They can kind of brainstorm with the AI. But if you've got somebody trying to decide if they should issue a loan to somebody, that brainstorming had better happened already to figure out what the rules and the right way to do things, and the laws and the other things are. So we have a foundation for creative reasoning, for building the best possible things. And we've complemented that with the ability to pull in client documents. If you've got manuals, documents, standards from the client, just pull it into Blueprint. Pega's years and years, decades of best practices which know a tremendous amount about workflows and the way businesses work, and actually go out to the internet and repeatedly ask the internet questions about parts of this and not believe any of it. The whole point here is for the Blueprint agents to grind together and try to come up with something better than any of the individual inputs. Now you're going to see us doing things where we consume legacy systems and replace them. We're not just doing code replacements. Some people are doing this. I'm going from COBOL to Java. Who cares? System's still going to be crap. What you really need to do is go from COBOL, Java, whatever it is to what it should be. And that's what Blueprint does.

Now, one of the things I'm super excited about, because this ties in to our legacy transformation pitch that we are really launching here at PegaWorld, our new capabilities is that we have deepened our relationships with a set of key partners, with two of our diamond partners, EY and Accenture, and a set of other really important partners. Where we have given to them a place, a Knowledge Buddy, where they can put their IP into a language model setting. And if one of their staff works with one of their clients, that IP will be incorporated into what Blueprint grinds together and it will put their best practices into the mix as well. Unique to them, no other partner, no other Pega person can see the IP that these people are putting in, and I think this represents a huge step forward in not just improving collaboration, but improving innovation in the whole way that this works. And it's really important to us that all of this imagineering happens in the right place. You want to be able to envision the way you want your business to work. You want to be able to create, think, be free, but you don't want to confuse that with the way you want the business to actually run. Because when it runs, you want the AI to be predictable, to be able to be reliable so that two people whose problems are basically the same but are slightly different, don't get different answers. That's how businesses get into trouble. And candidly, that's the way that people lose faith in the capability of AI. You need to be able to make it so that when you run in production, when the real, real bridge gets done, it's solid and reliable and not built on creativity that might suddenly appear.

And to be predictable at the right time, you need to be able to have the agents orchestrated in a way where the design, as we do it, is really separate, easy to do, able to have as many workflows as you want, but you're able to reflect what you're building in terms of the articulated workflows that run your business. You're able then to have a conversational interface which uses a language model at runtime, but is only using it for conversation, only using it to translate from what the person wants to do to find the right workflow. Put the right data in the right workflow. You need to be able to then automate based on that so you can get as much automation as possible, and then you need to be able to coach people. If a human is involved, how do I use the language model deterministically to coach the person? And then finally you want to make sure you've got knowledge, knowledge to close the loop and feed into the design, but also knowledge that might help the person do their work or help the coach, or knowledge like Socrates, which might help train you. This concept of predictable AI, I think is unique to Pega and is enormously, enormously important. So what does this mean? Well, for me, this is a complete vindication of our Center-out architecture, the idea that you want to be able to have the work that your business does at the center of your business. Now designed by Blueprint. The thought that you want your channels to be thin and light and driven by the center so you don't have business logic and business rules in the contact center channel of the web channel, the mobile channel.

The thought that customers and employees and automated agents, whether they're from Pega or from other companies, can interoperate, whether it's with traditional APIs or whether it's new standards like MCP be able to make it so it all becomes part of an operational foundation of excellence. And then, of course, on the data side, it's absolutely critical to have encapsulation so that inside that diamond you have coherent and consistent rules, even if the data on the outside is somewhat or very different. And that idea of the live data agent that does that mapping eliminates all of this debate about data fabrics, this idea that people are going to create data fabrics and somehow you're going to get all your data right. Look, don't get me wrong, I love it when the data is right, but waiting for the data to be right is something we've been doing for 40 years, and it's never going to be right because there's always another data source, etc.. Our approach is to use AI to accommodate the variations of the data, but still apply them to common processes and rules. And I think architecturally, that's not just more realistic, architecturally it's better in any sort of volatile environment. And this idea of Center-out is central to the concept of legacy transformation, where in this release and you can see it, Pega can generate cloud native databases. So if you've got a system you want to kill and that system has some data in it, and you don't want that data in Pega, you want that data to be in a database of yours, external, we will generate the bindings and the database definition language using the AI. So that database is easily set up. This is where our partners, I think, are going to be priceless in terms of being able to help us make this all work. So it's an interesting time, and it's an interesting time for people to realize that doing this requires a bigger look than just an individual system. And what I'm pleased to announce is the concept of the Agentic Process Fabric.

How all of your engines, all of your workflows, all of the elements of your business can be pulled together to operate as a coherent whole. That you're going to be able to treat this disaggregated set of Pega systems as if they were one. You can have a conversation with an agent that can fire a workflow in one system, and then stop and fire off a workflow in another. And this Agentic Process Fabric has, at its heart, predictable AI so that it's always predictable, even as it does the work across the enterprise. It's cloud native and it's Center-out. And those are the elements that these businesses need to get started, advance and know there's a destination that will hold everything together. We're super excited about this. We have it on significant display in the Innovation Lab. You'll see some demonstrations of this today, and I will tell you that the changes to AI have beautifully aligned with Pega's architecture. And we've been working nonstop to use this to address the questions people have about how do we make it easier to use faster to implement, more robust, broader. I think we've got those questions nailed. So have a wonderful time at the other sessions, at the breakouts. Check out the Innovation Lab and thank you, you constellation of creators, for coming! Much appreciated! Thanks!