Transcript
Speaker: This is a little bit of an uncomfortable joke, but I said, you know, the most important use case of AI is for you to be able to tell your SVP so he can tell the CEO, so he can tell the board of directors that they're using AI.
Speaker: Because what's happened is that but we have some real hysteria going on. Boards and CEOs need to be able to tell their contemporaries that they're using AI so that they can tell their shareholders they're using AI.
Speaker: But in reality, it's how you use AI that matters. You're listening to From the Horse's Mouth, intrepid conversations with Phil First.
Speaker: Hello welcome to the latest edition of From the Horses Mouth podcast. I'm your host, Phil First, and joining me today is a real legendary figure in the world of software, Alan Treffler, who's the founder and CEO at Pega. Welcome, Alan.
Speaker: Thank you, Phil. You know, it's interesting because Pegasus and Pegasystems obviously are very near and dear to horses of all types. Absolutely, yeah. I've been covering your company for a very long time and we haven't had a great amount of opportunity to get to know each other. So I'm glad delighted to get you on today. Perhaps you could just start for the benefit of our audience to tell us a little bit about maybe yourself and Pega and when you founded the company when you founded a company and where you are today.
Speaker: Well, you know, Pega has been on a four-decade journey. across great changes in the technological and business landscape to actually bring something to our clients that's remarkably consistent.
Speaker: And when we started, it was really about bringing rules and statistical AI, being able to to find yeah what people called machine learning to find patterns into workflows and to be able to have businesses understand how to do a better job for their customers or or automate their processes.
Speaker: And you the if you actually dispute a you know charge on your credit card, there's probably about an 80% chance, given how prevalent we are in that business, that your bank is gonna use Pega to help them figure out what you should get for credit, how they should handle that, how they should talk to Visa and MasterCard.
Speaker: And so pretty sophisticated business processes have been our history. In the last three years, we've really massively changed how we engage and go to market using AI, but in a very different way than most companies are thinking about AI, and one that I think is really well-suited for the types of businesses that need predictability.
Speaker: So tell us a bit about You know, you've you've seen all the different it changes in the technology world over the last few decades. Why is this different and how have you approached this to be different?
Speaker: Well, you know, when when we um saw in 2022 the end of it that ChatGPT had come out and candidly was much better than anybody expected it was going to be at that instant, we we said to ourselves, hey,
Speaker: is this something that we really want to be using seriously in improving the lives of our our customers and their customers? And we said, absolutely.
Speaker: But when we saw how people were doing it, we didn't like the way most people were thinking about applying ai A lot of people were thinking about applying AI by creating you agents that would operate off of prompts, which basically programming and in English, that over time might call other agents. And you know we've we've seen actually in the last couple of weeks how dangerous that can be when the agents go a little bit rogue. You you might notice I have a pin that says no prompts on it, actually. Yeah.
Speaker: And the reason is not because obviously we do use prompts at times, but what we said is, look, we want to make it so our customers can describe a business problem. And then from that business problem, we want to take something called a vector database, ah a language model technology database that has our intellectual property in it from all the years we've been building workflows and doing automation.
Speaker: And based on the customer's description, based on the best practices we have, based on going out to the Internet, And asking, hey, is there a good way to do this?
Speaker: Bring together all those elements in this technology we call Blueprint AI to be able to propose to the customer a way to reimagine their business operations.
Speaker: And this has become enormously sophisticated. It actually has completely changed our business and is, i think, one of the most advanced examples of analytical and agentic and generative processing that that I've ever seen. It's available on beggar.com if people want to try it.
Speaker: And this enabled us to tackle A problem that we'd worked on for years, we'd built this incredibly sophisticated and powerful workflow engine that was able to do business rules, able to to form statistical opinions of things, and and able to really help companies improve their operations.
Speaker: But it was a little hard to use and it was hard to configure. And customers would complain that cost a lot of money to build and and took certifications and things of that type. and And candidly, we had been trying to improve that for years. We'd been really working on ease of use and and making things better. When we saw the general of AI, we said, boy, we can use this to completely change the game.
Speaker: So what we did is we used the AI. in two ways. One, at design time to help people design the business processes in a way that they thought made sense.
Speaker: And then at runtime, we're going to primarily run the workflow, the business process that was designed. But at any step in that workflow, we can actually call a customer agent, make ah you know an MCP call, as they describe it, to to call to another agent. Or we can actually call a PEGA agent.
Speaker: to do things like document processing. And so we we use the agentic language model capabilities in a very harnessed way.
Speaker: yeah We harness it, we and we use your cloud, we use OpenAI, we use Gemini at design time. But then at runtime, we use it in a very limited way, which enables us to both have a high level of confidence that there is no chance this will hallucinate, our view.
Speaker: And it enables me to wear this other button, i don't know if you can see it, that says no token cost. So in a world where suddenly companies are going, whoa, I'm being hit up yeah by these enormous cloud bills, we basically say, hey, look,
Speaker: We use the AI so candidly, i think intelligently, that we do the design, demonstrate how that works in a workflow, let let humans in the loop happen at design time.
Speaker: And then at runtime, 98% the time, you don't need your missile loop because you're executing something that's been well far through and well conceived. And ah you know we think this is a really exciting way to bring AI broadly.
Speaker: into workflows which are candidly the heart of lots of but different different businesses. so So you've talked a lot about this token trap I've been reading, so um and you've touched upon it already, but what exactly is it and how could the consumption economics of AI become a serious problem as clients try to move from experimentation to scale, Alan?
Speaker: Well, look, you know, when we we had a wonderful conversation internally in early 2023 in which we said, you know, these these drug dealers are given tokens away for free. Basically, it was like $20 a month, and you could do pretty much anything you wanted.
Speaker: But we we said among ourselves, we said, look, These guys are talking about spending billions of dollars on data centers. And ah sooner or later, they're going to want to be paid for that.
Speaker: Now, since then, it's become hundreds of billions or trillions of dollars on data centers. And that's part of what's just going to continue to make these economics fraud. But I think that, you know, if you take a look at what the the AI providers are saying, I think they're trying to push the customers into one of two bad decisions.
Speaker: One decision is, hey, just use the AI to write code. And let's face it, AI is terrific at writing code. We use it extensively.
Speaker: But if you're trying to build a system of any scale, people, I think, in some cases have forgotten how hard it is not just to build those systems, but much harder than building them is the maintaining and evolving them.
Speaker: The ability to change them is is incredibly difficult. So you see people who have actually built applications, but but they can't deploy them to production because they're not sure how they're gonna change them.
Speaker: And we think that's kind of a trap that the model companies have set because when you wanna go change one of these systems, they end up reading the entire code base into the model to understand it.
Speaker: And that can eat an enormous number of tokens when that happens. But even worse than that, you know, in a workflow system like ours, you can look at the workflows and you can understand what the system does.
Speaker: and it's gonna operate completely reliably, but it's designed to help business and IT speak in a common language. When when you look at a code base,
Speaker: even if it's documented somehow, that's that's not a very pleasant language. So that's kind of trap one, the you know AI companies are setting when they say write huge amounts of code.
Speaker: But trap two is even more serious and has even greater economic issues. Trap two is they'll say, well, don't write the code. I mean, Anthropic would far prefer that you created bunches and bunches of agents right which are basically prompts fed into the model,
Speaker: to respond to do a particular task. And they want you to be able to create these agents and have these agents call other agents, which might call other agents. And then you you hear from companies like ServiceNow, which talk about creating a control tower to control all your agents.
Speaker: and And then you know people say you need a kill switch to be able to turn them over, off. We think this is madness. We it's just insane. What we think and what we've done is that you should have one agent that understands workflows and understands how to read the workflows, understands how to process them, understands how to interact with people to execute the workflows.
Speaker: And then that one agent can actually take all the workflows that can run a business and suddenly they all become agentic. And you don't have a control problem in terms of them you know potentially going rogue, but you've built governance in at design time because you've had the concept of a design time versus a runtime. And I think that this third way of doing things, this design and then run, what you've designed, both gives terrific economics from a token point of view, but more importantly, it gives predictability of outcomes.
Speaker: which in the businesses we tend to work with is really important. Yeah. One of the things which I'm really concerned about is this use of what is sovereign versus what is shared. And I think it Alex Karp came out and said, you know, don't put all your intelligence on a rented model. You know, what would it what was it? A claw put Figma out of business pretty much because they figured out how to do their business. So how do you work with clients in terms of what they're putting onto a,
Speaker: an arthropic model or or or an open-weight Chinese model, that sort of thing. how do you How do you manage that with clients? Well, the way we manage it, we say, look, you you want to define your business in terms of the workflows and rules that you operate under to deliver outcomes for customers.
Speaker: You know the way I think about it is that the people who worry about what Alex Karp is worried about They're training the models to make decisions the models shouldn't be making.
Speaker: The language models are intrinsically probabilistic. So you're you're always going to have certain weight risks of variance and drift. You know these days, in the last month or so, we've come to realize they also can go rogue.
Speaker: yeah And they can start doing wild and crazy things. Thinking that you can control those with English language prompts... I think is a little naive, which is why so few of these serious systems that are based on on training or fine-tuning the LLMs, so few of them have actually gotten live in any sort of scale.
Speaker: But we have lots of systems that are live based on workflows. Workflows are reliable. Having the workflows be agentic It's just a natural way to get better automation in there and also to be able to interact differently with people or with other systems.
Speaker: So the way we would handle that or say that is, look, you shouldn't be relying on the models to do things that you can put into workflows. the Workflows are a much better way to think about your business.
Speaker: yeah So would you say like that Pega is moving more into, it doesn't matter so much if clients go directly into your own interface. They could be interacting with their own harness or whatever workflow they're using. As long as Pega is there housing the data, housing the workflow and orchestrating it for them.
Speaker: Well, we don't really want to even house the data. We want to house the workflows and the rules. And we already support a full MCP interface. So a customer can use Claude, OpenAI, i Devon, whatever their preferred sort of the tooling is to be able to interface to the Pega engine to configure it, to say, hey, you know I need a new workflow that does so-and-so, and so and the the Pega engine will respond to that.
Speaker: And then at runtime, Anything you can do with a Pega interface, for the last two years, we've allowed you to do conversationally.
Speaker: So we've allowed you to do it through a WebEx or Slack type or or conversational interface here. And our agent knows how to interact with that conversation, look at the inventory of workflows, find the right workflow,
Speaker: And then either execute it or use that workflow to tell the conversation, I need some more data or I need some more information, et cetera. So this this idea of conversational interfaces, which I thought it was hilarious that you Salesforce announces the great innovation of you know headless Salesforce.
Speaker: yeah We've been headless for eight years. And it's, I think, to be honest, It's one of those cases where the right answer is and.
Speaker: People are going to want to operate conversationally for certain things. And for some things, it makes sense to see a screen. And being able to do both... or being able to allow me to have that conversation, not directly with Pega, but have that conversation with my preferred AI model and have that AI model reach out to Pega if it's something that Pega should be doing.
Speaker: To us, it's just a natural part of our architecture. So we we love the way this is going. Yeah, it's like we're in phase two of AI, and phase one was experimenting and doing all the things that you're like prompts and things which people are realizing are not effective. And phase two is how do you scale this in a controlled way that you're not blowing costs out of proportion?
Speaker: But then how do you get to this outcomes-based nirvana that we all love to talk about? I mean, have you really started to see this in in production yet? Well, look, for the last so about 10 years ago, we realized that selling software by seats made no sense.
Speaker: And part of it is that we were really good at doing automation. So, you know, we would have customers who would come in, they'd have a, ah you know, business function, you know, a loan approval function or a ah or customer response function that maybe had a thousand people. And as a result of putting PEGA in, suddenly they only had 800 when they came up for renewal.
Speaker: And what we said is, you know, it's not really fair for you to get 20% discount because because We gave you such incredible power. If anything, yeah you should be saying how much you love us and and and paying more. So we moved away from charging for seats, which is sadly still used in some areas, to charging for the actual units of work that got done.
Speaker: Every unit of work in a Pegasystem, we tend to call a case, like a loan that's approved or a customer that's onboarded. So those are effectively work-based or outcome-based measures.
Speaker: And that is the standard way we've talked to our clients. A lot of other companies are now trying to discover outcome-based measures, but because our outcome-based measure is actually tied to how the system really works,
Speaker: It's very natural in our environment to be able to count cases and charge by the case. I think the people who are trying to yo hook on to outcomes, yeah I read an article that talked about how some of the other vendors in this space are basically doing every contract as a completely custom contract.
Speaker: and And that doesn't really sound like something that's going to make customers happy in the long run. Yeah, it's not ultimately about starting to engage beyond the walls of the IT office. So I look at a company like Palantir, and they only sell to the CFO, the COO, the CEO, because it is an outcome-driven conversation.
Speaker: and it sounds to me like, you know, when you sell to the business lines, they like buying in an outcome-driven model, because it's like tie this to revenue, or tie this to remittances produced or something, right? So...
Speaker: Yeah, and every one of those is an example of what we would call a case. I mean, what what we've always believed in trying to get the business to be central to this, not just in terms of buying, but also in terms of implementing.
Speaker: And this whole blueprint AI technology is really aimed at the business, but in a way that IT t finds itself. helpful. You know, I think what a lot of companies that are selling to the COO or CEO are doing, you know, it's kind of interesting.
Speaker: I was in Europe two weeks ago and I was asked by some customers, you know, what what I see is the use cases of AI. And this is a little bit of an uncomfortable joke, but I said, you know, the most important use case of AI is for you to be able to tell your SVP So he can tell the CEO, so he can tell the board of directors that they're using AI.
Speaker: Because what's happened is that as as part of this hype bottle bubble around AI, which has taken something that is very real and kind of made it so yeah we have some real hysteria going on, boards and CEOs need to be able to tell their contemporaries that they're using AI so that they can tell their shareholders they're using AI.
Speaker: But in reality, it's how you use AI that matters. And, you know, we think this idea of design time and runtime as both places where AI can be used selectively, both makes it easier to use, but also makes it real.
Speaker: Yeah. So, I mean, we've been talking for years about, know, do you design the business around the technology or the technology around the business? And there's a lot of noise out there that you've been alluding to ah where, you enterprises think, oh, we have to like completely redesign our processes to take advantage of AI. But everything I'm hearing from you is you don't have to do it that way. It's it's it's just thinking how real do we leverage this?
Speaker: And it's not like an immediate, let's just blow everything up. I mean, what's your take on this longer term? um Well, I think you should be able to. any Any good technology provides an on-ramp that allows you to operate in a continuous learning mode.
Speaker: And if you think, you know, some people think that their business is all about their data, right? I think that data is important, but yeah the business is really what are the business processes that you offer your constituents?
Speaker: You know, what do you offer your customers? What do you offer your staff? but You know, how do you operate in this business? And being able to define those processes in a way that AI can help you define them,
Speaker: help you improve them, and then help you evolve them, and then help you execute them, is a way where you get to not just you know k crawl, walk, run.
Speaker: you know You get to crawl, walk, run pretty fast. And you do it in a way... That is predictable, that's understandable. you know I think you know the first time, we we work a lot with banks and insurance companies, you i the first time a bank COO or president has to tell the regulator that the reason they approved this transaction to a blocked out company was because the AI told them to.
Speaker: That's not going to be a conversation that leads to a second conversation, right? Nobody's going to accept that, particularly with how unforgiving, candidly, people are about mistakes made by machines. We need processes that work, that are predictable, that we can look at in advance of their execution.
Speaker: And, you know, it's not just efficient to know how the AI came to a decision. You want to know what the AI decision criteria are. yeah What we do as part of Blueprint is we say, hey, if you're going to make a lending decision or you're going to make a decision about whether or not to onboard this customer, tell us what the criteria are. Use all the intelligence that's out there from our experience, from whatever the customer of the bank has in terms of its procedures, from whatever the regulations say.
Speaker: Use all of that and create the framework that's going to control the decision. and then put that framework into a decision table or decision tree or something that will operate with complete transparency and predictability.
Speaker: And in that case, you've used AI to really help improve your process or improve your decision making. yeah But at runtime, when you actually execute it you're going to be able to know why it made a decision.
Speaker: yeah The current use of AI, where you can't really be sure why it made a decision, and you you don't even get to know in advance what decision it's going to make.
Speaker: I think the fact that it's kind of magical, and you know I love using it, and it's really you know enormously fun, people have forgotten that it's these last miles or last inches that are what make a company either you're compliant and ultimately efficient or not.
Speaker: And suddenly the fact that it's gotten so much more expensive, candidly, should make people just more aware that the way they use it is as important or more important than actually using it. It's interesting, isn't it? exciting and And any technology that permeates your personal life so effectively, i think, has that bigger impact. Because you know if you're an efficient user of AI for your personal life, you your shopping recipes and however you like to do things, you go to work and you think, we've got to use this technology better.
Speaker: But where are we going now? We we have AI becoming ridiculously political issue. The panic around, i mean, the stuff coming out from the frontier phones is bizarre about you know end of humanity and all this all this stuff.
Speaker: Do you think we're we're just approaching a kind of correction that's going to change everything and we're going to have a very different conversation? Or or is this just a trajectory that we're going to continue on for for the next couple of years at least?
Speaker: I think there will be corrections. But, ah you know, the thing I would say about your personal life example is when when you use AI in your personal life, and also for some of the work uses you use it for, you're using it in an environment in which it's intrinsically being curated by you.
Speaker: right It's actually interacting with you. And the fact that it may be probabilistic and may suggest things that don't exactly make sense is actually in some ways a feature because it's going to challenge you to think differently.
Speaker: you know I think for somebody to use AI in their marketing department, to be able to brainstorm how they might want a ah new approach to our customer base to be, boy, that' that's really a terrific and exciting use case.
Speaker: And from my point of view, it's safe. because it's curated by a knowledgeable expert worker. And that's true for you and your home use cases.
Speaker: I would not turn your credit cards over to AI and tell it that it should figure out what your home entertainment needs should be for the next yeah year and go buy the equipment.
Speaker: I think you would find that it might do a brilliant job, but it might not. and so So deciding what you want to be in the middle of is super important.
Speaker: and And what PEGA does, we're we're really trying to use the ai to optimize environments that are what I would describe as operational in nature, that are the types of things yet you often do, john not just over and over, but you do to basically represent how your business works.
Speaker: And in those cases, I think you would need to use the AI very, very differently from when you're using the AI to brainstorm. Yeah, we do the brainstorming as part of this blueprint thing, but we do it at design time.
Speaker: At runtime, we want to use the AI very narrowly to do a translation or to summarize a document or to pull fields off of something. Those narrow uses, by their nature, limit the range of bad outcomes that you can get.
Speaker: And yeah i think that is very, very different than you know if you went to and talked to the Claude folks, they'd say that you've got to create agents by the dozen.
Speaker: And the agent should call agents. And by the way, the agents might actually invent their own cadre of agents. And, you know, we talked about how, you know, the frontier models have been talking trash in the last couple of weeks about themselves.
Speaker: ah You know, it's kind of like they're telling, you know, customers, we don't know how to control this stuff. But you should buy it. And, you know, as ah it's an interesting marketing technique.
Speaker: And you can't argue with some of their results yet. But I don't think it's a recipe for getting really ah successful and predictable outcomes to those companies.
Speaker: Yeah, I mean, I think ultimately something like Claude will become like a Ferrari. Like it'll be used for things like life sciences and biotech and stuff that's really sophisticated.
Speaker: But for general enterprise usage, end of the day, people aren't going to be able to spend millions and millions of dollars a year on maxing out tokens and agents just to do basic work, right? Yeah, but if you use Claude in the design, capture that design and workflows,
Speaker: And then at runtime, yeah you can use Claude for just being able to like summarize a document. doesn't cost that much to do that. So it's not Claude as much as how you use it.
Speaker: If you say that you're going to treat Claude like an AGI, you know ah I guess we're supposed to call it super intelligence now, but you're going to have the artificial general super intelligence. If you're using Claude like that, then you're actually candidly abdicating your responsibility to provide the ongoing guidance and to provide the curation.
Speaker: And it as part of that, you will get you know prices you can't afford. But you know the fact that we don't charge for tokens isn't because yeah we're trying to be charitable with our customers.
Speaker: It's because if you use this stuff right, you can really be very intelligent and frugal in terms of how you do it, but still get the benefit. but What I worry about in some ways, and you know we're starting to see this as part of the political landscape, is that because it's not just cost here, it's the sort of ah environmental impacts, people having data centers built, you know, around the corner from them and now saying that maybe that's not such a great thing.
Speaker: Well, yeah, not great having a data center built around the corner from you, but it's also not great using these incredibly consumptive, incredibly consumptive GPUs, which are are kind of, you know, miraculous chips to be candid. It's really, really brilliant stuff. But why are you using them for things you could do with the deterministic workflow?
Speaker: And the answer is, because it's designed wrong. And that's really how we've gone at this problem. And it's designed wrong for cost, it's designed wrong for predictability, and it's designed wrong for the environment.
Speaker: This has been fantastic, Alan. One thing maybe to leave us with would be, you know, you've had a tremendous career in this industry. What advice would you give to maybe, you know, younger folks coming out of college,
Speaker: pursuing an IT career in this current environment, what would you leave them with in terms of how should how how they should focus? ah So I think, you know, it's interesting because I was asked recently whether somebody's child going to university should study computers.
Speaker: Obviously, they worked out pretty well for me. But I think that it is still pretty important ah to understand how these technologies work at some level.
Speaker: even if we're not going to be using them, you know in the same old way we used to program them, because a lot of that stuff is irrevocably changed. But I think I would advise people, ah you know, in the studying parts of their career, which they should continue long after they leave school, to really just expose themselves to what's going on and do it with a level of healthy skepticism.
Speaker: Realize that, you know, we've been waiting for fusion. Fusion's been five years away for 40 years. You know, so understand that some of these things don't occur on the timelines that the pundits want them to.
Speaker: And understand that some of them are just bulls. But sometimes they're kind of magical. I mean, who would have imagined that by eating large corpuses of the internet and language, that you could have things that actually were almost indistinguishable from a human? I mean, that would have been something that when I was in university, we would have thought, ah boy, that wouldn't work that way. So got be open-minded, but you also have to be highly skeptical.
Speaker: And you always have to be learning. Those those are i think, the... the critical elements of not just having a successful career, I think of having a successful life.
Speaker: Yeah. 20 years ago, 30 years ago, you probably only have to retool maybe once or twice in your career. Now you've got to like retool your s skills constantly. it's ah It's a different mindset. It is. And the only thing I'd say that's good about the that is that it keeps you young.
Speaker: So I think if you're constantly if you're constantly learning, that's you know that's both a skill set that needs to be cultivated, developed, but it's also, I think, very reinforcing to both achievement and creativity. I mean, I just i found the last three or four years invigorating. like It's sort of like, this is exciting stuff.
Speaker: It's a bit of anxiety, lot of opportunity. Well, Phil, you probably didn't have anybody explain that your business was dead because, you know, of Saspocalypse. You know, part of part of this whole, you know, some of what's happened in the last year is that there's been terrific, I say i would describe as bad categorization, where people haven't realized that you know, there are lots of different types of software.
Speaker: And yes, some software is dead, you know, candidly. There are things that you can just do now, or you can just program now that you didn't use to be after, but there's lots that's not. And that does make it invigorating, I will say, but it does make it sort of just a really interesting environment here. And and the the level of overreaching by companies as as achieved levels we haven't seen since the dot-com bubble.
Speaker: I'm not sure why Anthropic thinks it should be worth, well, actually, i know why they think they should be worth $2 trillion. dollars it's It's because SpaceX was worth 1.8.
Speaker: But, you know, this whole, I would say, highly imaginative way of looking at things that are real But overreaching with them, I think ultimately people need to caution themselves to both look for what's real, um you know, not be ah skeptical of the core, but i'll also be skeptical of some of the claims and some of the approaches.
Speaker: Yeah. And on that note, I'd like to thank you very much for your time. i really really enjoyed meeting you today, Alan, and hopefully we can get together in person soon. But this was great. Well, we should be able to pretty easily. And thank you. I i think your your newsletter provides a healthy periodic dose of skepticism, which is very much appreciated.
Speaker: Thank you very much. i appreciate that. Thank you for listening to From the Horse's Mouth. Don't forget to subscribe and like wherever you listen to podcasts. Got something to add to the discussion? Drop us a line at fromthehorsesmouth at hfsresearch.com or connect with Phil on LinkedIn.
