Inside the Advanced AI Lab:
Agentic AI, Robots, and Levitating Trees

Michael Krigsman: Welcome to CXOTalk Episode 854. I'm Michael Krigsman, and we are taking a peek inside Cognizant's Advanced AI lab with Babak Hodjat, the company's CTO for AI.

Babak Hodjat: I am the CTO AI for Cognizant. I run our advanced AI R&D lab in San Francisco, where I am currently. This is where we do core AI research and feed into our flagship Neuro AI platform, which helps our clients with AI-based decision-making.

Michael Krigsman: You were one of the inventors of the underlying natural language, NLP processing techniques used in Siri. Tell us about that, and is there a link between that research and what you're doing today?

Babak Hodjat: I was working on my PhD on a distributed, AI-based technology called multi-agent-based systems. I was amazed by how powerful these systems can be. While AI was still in its infancy, it did well on toy problems but needed to scale. When you put multiple AIs together into a system, they did quite well.

A friend of mine, through a sort of misunderstanding, challenged me to make use of that same technology to understand people commanding and controlling things like the TV set to change a channel or what have you.

At first, I thought that was, you know, too complex a problem to tackle for AI. Natural language was the first non-numerical application of computers, but it had never been quite solved. It's very, very difficult due to the ambiguities of language and how complex it is.

But then he challenged me, and I worked on it. It turned out that for the domain of command and control, using a multi-agent based architecture actually makes a difference. It was a big departure from the traditional way of doing natural language processing: grammar, parsing, semantics, and context, all the while being very language-centric. So if you did English, then moving to Japanese would be a completely different ball game.

This opened up the door to many possibilities in that domain of command and control and, ultimately, led to being able to create the natural language interaction core that ended up being Siri.

The relationship to today, in some ways, is very analogous. Today we're back into a world where we see how scaling AI and moving it into businesses and various different applications in a multi-agent manner is preferred over a centralized single system doing everything. So, agenthood and multi-agent systems are enjoying a big resurgence right now.

The motivation is different. Back then, AI was really limited. To empower it, you reduced the domain of application, called it an agent, and gave it specific tasks in conjunction with other agents with reduced domains. Then, they would interact and solve a problem over that community of agents.

Today, a large language model, which is the core of agents as we define them today, is very, very powerful, and it can be many different things at the same time. So, you have to tell them what to be. It's kind of the inverse: they can be many, many different things, but you have to reduce that domain to tell them what their set of responsibilities are and the requirements for what we want them to do in a particular context. Then, we call it an agent. So, coming at it from a very different angle but ending up more or less in the same place as what we did back 25 years ago.

Michael Krigsman: When we talk about agents, or the term agentic AI is now becoming more and more commonplace, very ridden with hype, you use the term multi-agent systems. Can you elaborate on that term?

Babak Hodjat: We all are familiar with large language models and generative AI models. These models take input and generate output. The keyword is "generate." Take input – text, images, video, or all of the above. The output could also be different modalities.

An agent, however, is distinct from a model because it does things. It has a toolbox, and it can affect its environment and observe the effects of that change or using those tools.

Let me give you an example. If you go to ChatGPT and ask it to write code to make some calls to the Amazon API, it will generate the code. It will not run it.

It's so good that the code might actually work if it's not complex. You might be able to rely on and trust that code.

The distinction, though, is that an agent would be able to run that code, maybe in a container, make the calls to Amazon, check to see whether its unit tests are passing, and then modify the code if it was broken before it came back to you and said, "Here's the code." So that's one agent distinction between an agent and a large language model.

Now, regarding multi-agent systems, this means we have different functions in some sort of organization.

In coding, we might have a project manager, a programmer, someone that writes unit tests, a backend coder versus a frontend coder. We might have a UX designer versus a UI designer. And we might actually have a Kanban board where we're posting tickets and closing tickets for the coding.

So, right there, I listed a number of different tasks and different responsibilities, and that's how, as human coders, we organize ourselves. Now, we have a choice. We could either rely on a very powerful large model or single agent and ask it to do all of those. It's very difficult to set that up.

Currently, we have limitations on context. The complexity of what we're expecting the system to do is very high, and observing how various elements of what we are asking it to do is difficult.

For us to set up a system like that, we already are thinking in a very modular way anyway. So mapping that modular thinking, that structured thinking, to the way we set up a team to produce the more complex, larger software package that we want the team to create simply makes sense. That mapping is basically by creating multiple agents, each agent being responsible for a part of this whole.

For example, the product manager, project manager, the UI/UX designer each becomes their own agent. In the context of the agent, we tell it its job description, what it needs to do, what to expect to get from other agents, and what to return to them.

Then we set it off to operate. There are several interesting properties, not least the fact that we can think of ourselves as agents in this structure as well and define the human role within that structure. This is unlike, for example, treating this whole thing as a black box single agent that we ask to do something, and then we expect it to return something. We're not anywhere in that loop.

So that is one of the important reasons why looking at software in general, and even organizations, as multi-agent systems makes a lot of sense right now.

Michael Krigsman: So, are you attempting to then model the construction of an organization in software using these multiple agents, each of which is designated to accomplish a particular task? Is that the basic goal here?

Babak Hodjat: That's great. You've read my mind. That's exactly right.

It goes beyond software. For example, from one side, we can look at software. Software is already inherently modular. We've had this whole object-oriented approach that's been the underpinning of a lot of software principles for quite a while, and objects are just a tiny step away from agents. So, you can think of software itself as being this modular system that can now be represented by agents, and it has some very interesting properties. We can talk about that. That's really interesting.

But this goes beyond software. We can look at an organization where the responsibilities are in these nodes. They could be human, software, or a mix of both, running this organization. As long as we can actually list those responsibilities, tasks, and job descriptions, including what to expect coming in, what to expect to leave, and what to task the agent to expect from some other agent to deliver, if we can actually define these relationships, then we can actually attempt "agentyfying" an organization that's not necessarily a piece of software. And it has huge implications.

So, yeah, that's an area of interest to me. What are the implications? How do you do this in a human-centric or human-augmented way? How do you guarantee that a system like that actually works?

What are the properties you gain? Because running and thinking of everything as agents has its costs, too, right? You're running large language models, which are the brains of each agent. So, you are actually tasking; you're still running large language models for every step of this organization. Depending on the throughput, you are paying good money to operate this thing. So, on the other hand, you're getting robustness, future-proofing. There are a lot of very interesting properties that you're getting as well.

Michael Krigsman: It sounds like you are constructing a robot army that encapsulates large language models to accomplish whatever goal you want. You're nodding your head. Okay, so then we know that LLMs hallucinate; they make things up.

How do you prevent your army, or members of your army, from going rogue?

Babak Hodjat: Large language models can fabulate, or as they are commonly known, hallucinate. They make stuff up because they're generative. They're not trained to be truthful. They're trained to pass muster with their output. That is a real problem, and it's another reason for us to "agentyfy" things. When you reduce the scope and domain and make it very clear what you expect, you reduce the hallucination. That's one aspect.

The other aspect is that we need to make sure that we have instrumentation in the agent that tells us when we can trust it and when we should verify.

There is some very interesting recent research, some of it coming out of our lab here, that measures uncertainty on large language models. It's amazing. Given a context—again, that context is important—this is something that can be handled within an agent scope, not in a general scope. Given a context, for a certain input and an output that we get from the agent, how much can we trust that output?

If we now have a measure of that, and based on that measure, set a threshold, we can say, "Hey, right here, what this agent did, I can't trust it as much, so I need a human to verify it." This is another reason to look at these systems from an agentic perspective versus a one-size-fits-all LLM that does everything.

Michael Krigsman: Subscribe to our newsletter, and subscribe to our YouTube channel. Check out CXOTalk.com. We have phenomenal shows coming up with amazing people between now and the end of the year. So check it out! Join us. Greg Walters makes a very interesting point. He said, "Won't an army of agents cut down on the probability of mass hallucinations because of self-checking and self-healing? And is that possible?"

Babak Hodjat: Absolutely. The fact is, you can even task some of these agents with the responsibility to check other agents for things like hallucinations. If I'm an agent in a pipeline of agents working together, and the response I get from another agent is iffy, or is not really what I was expecting, I can also go back to it and say, "Hey, try again," or refine my inquiry back to that agent. So, even automatically, between the agents, we can manage a lot of this hallucination.

Now, having said this, we are faced with a new reality with large language models being the basis for a lot of what we do these days. We cannot guarantee that these systems would never hallucinate. So if anybody goes out there and, for marketing reasons or whatever, tells you they can guarantee there's no hallucination, take it with a grain of salt. That itself might be a hallucination. Anyway,

that's not possible, but we can gauge and measure it. We can reduce it. The other thing we need to remember is that we're faced with that army being a set of black-box agents. So while we do have tools to try to pry into their thinking, logic, and reasoning, at their core, these systems are deep-learning neural networks that are very, very large and are black boxes. It's very difficult for us to know exactly how they're thinking.

Now, the good news is, we've been dealing with these sorts of systems since industrialization. It's just that they've been human-driven. Humans are black boxes, and we really can't read another person's mind. We do give them a set of responsibilities in a larger organization, and we know how to do that. So the good news here is that we're not faced with this unknown new era of multi-agenthood. We can borrow from all that we've learned through industrialization and setting up these organizations. We can bring those lessons in to make sure that agentyfication operates predictably, productively, and with high quality.

There's one other thing I wanted to mention before we forget, which is one thing that's amazing in this new era of multi-agency. Back when I was working on the pre-Siri days, we had to define the language between agents. It was a non-ambiguous set of instructions that these agents needed to issue to each other in order to communicate. We don't have to do that now. In fact, the native language of a large language model is natural language: English.

What does that give us? There's a reason why language evolved in humans the way it did. It's very expressive and allows us to express intent. So now you have an organization where the connection between these nodes—let's say if we had software—is expressing intent versus making specific calls.

This makes it robust and future-proof. You can replace various elements, for example, in your software. One of the examples I use is that I used to work at Sybase many years ago, and we had databases. They were not state-of-the-art databases. We had lost the database wars back then. But these databases were the standard on Wall Street for 15-20 years. They were legacy. The reason they were standard was because it was very difficult to yank them out and upgrade them or change them because they were ingrained into the software. Probably, the people who integrated them had long left.

That is a thing of the past now. That robustness, that connection between software nodes, which are now agent nodes, gives us an intent-based and natural language-based system. This allows us to replace and upgrade without caring about integration as much because the connection between the intent and the call into the specific API is handled by the large language model now. So that gives us a lot of flexibility just on the software side. For someone who has a software engineering background, I think that's a really interesting property of multi-agency.

Michael Krigsman: You said that we have been dealing with humans who are making decisions through organizations from the beginning of organizations. However, if we look at it from an ethical perspective, one fundamental difference is this: with humans, we're the sum of everything that we know, and we could call that data, and we have our biases. However, when it comes to data-driven decision-making, I don't think we can be quite so easygoing about potential biases that live in the data. Maybe talk about that aspect of it.

Babak Hodjat: There may be, and certainly are, biases even in our general-purpose large language models. We need to control for them, for sure. That's one big major risk. There's also bias inherent in the data that we've collected and are operating on.

Before I address the question directly, I want to mention this: If we create an organization that's only large language models operating based on their beliefs and their training, then I think we've done a disservice. That's not where we want to go. If an organization already has data that they've collected, historical data around the decisions that they've made, their operations, and so forth, that data is invaluable. We have to task our agents to defer to that data and perhaps even use what is now called "traditional" (as of two years ago) ML models and analytical models.

So, we have the agent, which is a large language model-based agent, operating a machine learning-based or analytics-based, data-driven model based on that data. And it's making its decisions. That is much more reliable than a single, large language model-based agent doing guesswork based on its reasoning. It opens us up to a lot of bias that is inherent from all the data in the world that is ingested to be trained. So that is super important.

It still doesn't take away our responsibility for the data itself that we've collected, which might inherently be biased.

So the principles of responsible AI, which we had to care about even before large language models, still carry forward and are even more important now. We need to build them into the framework we use in creating and "agentyfying" systems.

To be realistic, in the past couple of years, that has taken a back seat. And it's scary because we've been distracted by the power of large language models and generative AI, and we're like, "Wow, this thing is so useful. I'm going to go all in," and we'll talk about the responsible AI part of it later. It's become like an afterthought. I think that's risky. That's one of the things that we are really emphasizing in the steps that we recommend as far as how to agentyfy. Built into it is a framework that is domain-specific and has human-centricity as well as responsible AI, ethics, and governance considerations built into it.

It's important, and there's no simple, general statement I can put out there and say, "Do this, and then you're responsible." It's very domain-specific.

Michael Krigsman: We have a very interesting question from LinkedIn. This is from Avi Singh Malhotra. He says, "How do you envision the balance between human capital and agentic AI workers evolving over the next decade? And what strategic shifts should organizations make to capitalize on this workforce transformation while maintaining a competitive edge?"

Babak Hodjat: We did research along with Oxford Economics on the impact of generative AI on jobs. The methodology was really interesting. We took all the jobs in the US job report and broke each job into its constituent tasks. Then, we said, "Okay, here's where generative AI is today, and here's where we think it's going to be in five, six, seven, eight years from now."

Which one of these tasks is it going to be disrupting? It's different depending on the job. Amazingly enough, it impacts 90% of jobs, regardless.

This is the reality. The way to look at it is to redefine and help move the job description so that the focus is on the part of that set of tasks that humans are responsible for. This will be the exclusive domain of humans. And yes, we need to endorse and bring in generative AI and agentic AI to cover the rest and to help those jobs. This report also included those jobs that would have a harder time keeping up, to help those folks reskill, upskill, and come in.

There are cultural reasons why humans, at least in the foreseeable future, will have an important, central role in any agentic system that we build. There are cultural reasons, for sure. As an offshoot of those cultural reasons, there are going to be preferences built in. For example, look at the arts. Someone asked me, "Arts now, you can produce many pieces of art that are satisfying to a lot of people. It's all generated by machines. What is the future of art?" We're at a transition point like we were with photography. People thought that painting would die because of photography. You can just click, you get a photo, and there's so much of it now. Photography was able to redefine itself.

There is a little bit of that. There's also value that we ascribe to the human part of the role. Like, if this is created by a human versus this being created by a machine, the value of that for human society is going to be different. And that value is going to ensure that humans have a role.

There is a lot to unpack here, and of course, I'm not a sociologist, so it's a little bit outside my comfort zone to opine. But, suffice it to say, there are many reasons why, at least in the foreseeable future, humans will have a very important role.

Going back to the Oxford study that we did, I think one of the first things we need to do within an organization is list the jobs and list the tasks that are going to be disrupted. There's going to be time. Yes, this disruption is unfolding quickly, but we will have the time to upskill, reskill, and as we agentyfy, trust me, as humans, we're going to be much busier operating these agents and agent-based systems.

Michael Krigsman: Over time, social expectations and cultural norms also change to incorporate the outcomes that technology gives us. What used to be science fiction and impossible, today we accept as a natural course of events, such as getting on an airplane.

Babak Hodjat: That's exactly right. There may be a future that, if we try to envision it right now, would make us very uncomfortable, including those of us who are working in AI. Just picturing it! But yes, societies change, value systems change, culture changes.

But we don't have to worry about that right now.

Michael Krigsman: Lisbeth Shaw, on Twitter, says, "Can you repeat your comments regarding GenAI versus agentic AI and the difference in their focus?"

Babak Hodjat: A generative model generates stuff. That's what it does. You give it inputs, and it generates stuff. An agent has the facility not just to generate, but then to actuate. So that actuation, that doing something versus generating something, is the distinction between an agent and a generative model.

It's as if you have a little sandbox that you give the model, and you say, "Hey, if you're writing code, run the code, too. Check the output, and maybe modify it if it's broken before you respond to me." So in writing code, that would be an example. But it doesn't have to be a sandbox. It might be actually operating something, like "Make this call." "Somebody just submitted a request for new software." "If you get this request, get all the information and verify that they're who they are. Here's the API for verifying them, and here's the API for, I don't know, opening a ServiceNow ticket to get them the software. Do it." That's what the agent is going to do. A purely generative system is not going to do that unless you give it the facilities to operate, call APIs, and do stuff. That's the main distinction.

Michael Krigsman: We have another question from LinkedIn, from Greg Walters. He says a couple of years ago, he wrote about AI bringing about the end of the C-suite, which on the surface seems like a silly question. But to what extent will AI take jobs from senior business leaders? After all, any half-decent LLM knows more than any of us.

Babak Hodjat: When we look at an organization, as we go up the organization, the decisions are more and more consequential and less data-driven. The focus has always been to automate, verify, improve, and optimize decision-making at these lower nodes. And somehow pay a lot of money to the top nodes. I'm saying that as someone who's in the C-suite as well, so I guess I'm speaking against myself there. But the fact is that now we have the capacity to help with that level of decision-making as well.

Now, when I talk about AI decision-making, it's usually misunderstood as, "Oh, we're always doing that. Like, AI helps us. It gives us insights, and then we use the insights to make decisions." So it's almost taken as a given that humans have this exclusive domain of making decisions. That's what I want to dispel.

We're not very good at making decisions. We have to face it. We are usually emotional, and we're subject to the latest thing that just happened or the latest news we get. The recency of something affects how we make decisions. We're myopically focused on a single outcome or objective at a time, whereas most decision-making is about balancing more than one outcome. And we don't really have the capacity to look at a lot of data to make our decisions. So often, we walk in with a gut sense of what needs to be done that is very skewed and biased and often could be irrational. And we make the decision.

We're paid for making decisions, taking the risk. Because when you think about it, if we're not very good at making decisions, but the decisions we have to make are consequential, we are taking a lot of risks by taking them. So the value we have is in taking the risk versus actually making quality decisions. And I think AI systems can be very good at augmenting us in making decisions.

They can already look at many more factors than we can at any given time. They can optimize against more than one outcome at the same time. They can optimize for, "I want to improve revenue while reducing cost while improving sustainability while reducing bias." It's like, right there, there are four different KPIs that I just mentioned. I have to have all four in mind before I make a consequential decision.

So I think it's really important to start bringing AI into our decision-making. There's a good methodology for capturing the historical decisions you've made up until now, bringing in the data that is consequential to decision-making, and having an AI tell you, "In this context, here's what I think you should do. Here's what is going to happen if you do this. Here's how certain I am in my prediction of the consequence of your action. And here's an explanation of why I think you should be doing this."

Also, if you want to do something else, let me know, and I'll try to predict what the consequence of that alternative will be like.

Michael Krigsman: I find myself surprisingly coming down on the side of Greg Walters about AI replacing senior execs. Let me tell you why, based on what you just said. There is an entire field of study that we call "the wisdom of the crowds." LLMs represent the ultimate wisdom of crowds because LLMs, if they're big enough, have all the information. So if we have to make a judgment, doesn't it make sense to call on the LLM to make the decision? The LLM can sum up human experience, history, knowledge, actions, and outcomes to guide us to the wisdom of the crowds that will actually lead to the best decision. And so, from that standpoint, why do we need so many senior execs?

Babak Hodjat: No, I disagree with that for several reasons. One is that the LLM is fixed in time. It's only trained up until a certain point. "PT" and "GPT" stand for "pre-trained." These foundational models don't learn after they're trained. So don't rely on an LLM alone for decision-making.

The second thing is, yes, it knows a lot of stuff. The wisdom of the crowd is in there. Maybe you can rely on it for some of its reasoning derived from that wisdom. But it's the wisdom of the crowd. It knows nothing about your specific context, the here and now of you and the decision you need to make and the consequences of that decision.

So don't rely on that. You can consult it. In many cases, it will give you some good advice. But don't rely on that alone.

Thirdly, if you do have data, you can rely on historical data for your decision-making—data from decisions you've made in the past and the consequences of those decisions. You should use that data and specific ML and analytic models built on that data. For example, if you have a digital twin, or an ML-based digital twin, of the subject of decision-making, you can optimize against that digital twin and come up with a much better decision strategy than any LLM could.

Let me give you an example. Let's say I have a refinery, and I need to make some adjustments to the refinery process to give me more throughput but reduce emissions.

Would I go and ask an LLM what to do? Maybe at a high level, it'll give me some tips on what to look into, but it knows nothing about my specific refinery, the state I'm in right now, the price of oil today, or all that kind of stuff. So, okay, I have to give it all that information. Then, it doesn't really know much about the alternatives available in this particular refinery, the changes that I can make. There are probably many combinations of changes I can make that are unknown to the large language model.

On the flip side, I have a lot of data about the refinery, and maybe I even have a simulation of the refinery. Putting an ML model of the refinery together with a simulation of the refinery would give me a digital twin that I can optimize against. And that optimization is this: which knobs and levers to pull and turn to get the best multi-objective optimization.

I would go with that any day over an LLM alone-based system. So, yes, for some cases, an LLM might be a good sounding board, but in many complex, important decisions, we have to create these agent-based systems that have these other facilities and tools at their disposal.

Michael Krigsman: Wilson Suarez has a question that I think will lead us into evolutionary computation, another area of focus for you and the Advanced AI Lab. Wilson Suarez's question is this: What role do emotions play in a world where humans don't make good decisions, and AI is impacting various scenarios?

Babak Hodjat: Emotions, I think, are very context- and culture-specific. They are an abstraction of a set of sensations, like a sequence of sensations that we have. They're typically not the same between two cultures.

So, making emotions central to our decision-making is more of a consequence of how we think and how that affects us, versus a sequence of steps or an analytical approach to decision-making. I don't think of emotions as coming from a different source and informing decision-making versus analytics or data-driven decision-making.

Let's set that aside for a moment. Regarding emotions, I would say that at least the sequence of sensations may have been evolved. I'm trying to build a bridge to evolutionary computation here. There is an interesting aspect that might be missing, which is creativity.

And creativity is interesting because the sum total of humanity's knowledge—captured mainly in text—that informs a large language model is just that. It might not capture very new stuff that we might encounter or need in very new situations.

While a large language model might be able to create new things that we've never seen before, and are interesting, in an interpolative way, on a go-forward basis, we might need to borrow from how evolution adapts to new situations in nature. We call that "creativity."

There's a lot of interesting work being done on how to rapidly search for new solutions, creative solutions for new situations, by borrowing from population-based approaches. What does that mean? It means that rather than having one model—as in the case of the large language model or a deep network—you have many variations of models that operate differently. Many of these are useless; they're too creative for their own good. But some of them are useful and different enough that you want to keep them around and maybe mix some of their traits with some of the other models. Through multiple generations of doing that, you end up with new and interesting, adaptive systems. That's evolution.

And evolutionary computation: up until now, evolution has been primarily done through methods that many of your listeners might be familiar with: genetic algorithms. So we have one model, or we have two parent models. Both are kind of fit for their environment. We borrow some of this, some of that, and maybe tweak something randomly. We have a crossover, we have a mutation, and we create a new model. Then, we verify if that model is interesting. Sometimes, it is more interesting and better than the two parents, and we keep it around.

Recently, though, we've observed that you can pull large language models into this algorithm. It's as if you have a human, with human reasoning, looking at the two parents and saying, "Wow, this trait here is interesting, and that trait there is interesting." And through their reasoning, they are going to use those rather than randomly selecting. You're being very mindful about how you generate that child model. You still need a healthy dose of randomness and so forth, but it has enriched evolutionary computation in some ways.

There is this give and take. Evolutionary computation is bringing in more creativity into state-of-the-art AI, which is large language models. But at the same time, large language modeling is also bringing reasoning into how we do evolutionary computation.

Let me give you one very interesting example, again, from our lab. We fed an evolutionary-based system, augmented with large language models, the papers published at a conference. We took ICML, but you could take any conference and any domain. We fed it the history of the papers: title, abstract, and a summary of the body.

We had the system map those papers in terms of how creative they are. Imagine a two-dimensional map with these papers on it. You will quickly see clusters of papers. Someone came up with a cool idea, and then other people build on that. So there's a cluster of papers here and a cluster of papers there. Every once in a while, you see a new paper, out in the wilds, that is totally out there. It brings a very novel, interesting idea, and then a cluster starts building around that through time—every year, as we hold this conference.

One of the interesting things you can do now is predict what papers will come out next year. So, at ICML 2025, what are the papers, and what clusters are going to grow? What clusters are going to shrink? And in what areas, in that space of possible papers, are you going to get very interesting papers?

What's even more interesting is that you can task the large language model to guess the titles and abstracts of those net-new papers. That gives you a mix of a population-based approach, evolutionary techniques, and large language models, allowing us to brainstorm around a topic or domain and predict what might be interesting in that domain.

Michael Krigsman: Why does this approach work particularly well for problems such as conversion rate optimization on a website or robots walking?

Babak Hodjat: Conversion rate is very interesting. You have a built-in measure. You have a website. It's configured in a certain way, and you know very quickly whether it's working because if people click and buy, or click and spend time, that's positive reinforcement for you.

If you modify that page, you very quickly know whether you made it better or worse. In other words, it's fitter or less fit. That gives you possibilities. We hear a lot about A/B testing. "Let's divert a percentage of our users to this new page that we've designed to see if it does better." And if it's better, we'll take that.

Now, the problem with that is that you only have two pages at any given time: your incumbent page and the new page. That doesn't give you a lot of leeway to make improvements, try new improvements, or keep up with a changing world. Things might change, and your user base might have different preferences.

So, one thing you can do is create variations of your current successful page and redirect to that population of possible pages—redirect some of your user base. Depending on how well the pages do, you direct more users to the higher-performing page and fewer to the others. The ones that are doing better, you can mix. There might be elements from different pages that are doing better, and you can mix those. Perhaps the resulting new page will be even more interesting.

It's almost like running an evolutionary process live, with humans assessing how well you're doing because they're ultimately the ones who determine whether your conversion goes up or down. It works very, very well. It improves conversions. The company I was at before joining Cognizant, Sentient, had a company that branched out of it called Evolved.ai, which does exactly that. They were able to show not only that they can improve conversion rates significantly but also that they can leave this system running live. When certain times of the year arrive, or certain events happen, and the preferences of their users change, the system adapts. It's almost like the design of the page changes and morphs based on the preferences of the users.**

Michael Krigsman: What does this have to do with robots walking? On your lab's website, you have a demo showing the use of these algorithms. By the way, as you're discussing this, Lisbeth Shaw on Twitter is asking, "What are evolutionary algorithms, and what are the practical use cases beyond genetics?"

Babak Hodjat: Let's say you want to teach a robot to walk, and it's encountering unknown terrain all the time. There are laws of physics that, initially, it doesn't know about. So one thing you could do is—like the agent we were talking about—have it interact with its environment: try different things. Don't try to walk initially, but just try to sense its environment and the physics of that environment. Then, build a model of that environment. Train in that model. It's much less costly.

You know, if you fall and break your neck in a model, you're not really breaking your neck. In the real, physical world, you might be breaking the neck of that robot, and that's costly.

That's what we were trying to illustrate: there is a version of evolutionary computation called "evolutionary surrogate-assisted prescriptions." It basically says, "Look, you should always create a model of your environment."

Remember how the agent is operating, doing stuff in its environment? This system says, "Create a model of that environment." Track the reactions of that environment to what you do. Based on that track, create a model. Train a model. Instead of trying things out in the real world all the time, try things out against the model. That's cheaper. You can optimize against that model in your head and then start walking.

When you look at that walking robot, we have ways to visualize what its model is telling it. At first, the environment that the model pictures looks very hallucinatory.

It looks like, "Oh, the trees are levitating," or "the sky is on the ground," or whatever. Because it hasn't had a lot of experience trying out what's out there. But as it's trying to learn to walk, it's also trying to learn the physics of its environment. Gradually, you see the environment stabilizing into the environment that we see from the outside. That's because every once in a while, it tries stuff that it knows to work in its model against the real world. Based on that feedback, it adjusts its model.

What's interesting is that it's actually solving a harder problem because the world it has in its head can be pretty fluid. For example, if a tree levitates, and you need to walk around it, that's much harder than walking around a stationary tree. So if you learn how to walk in a world where trees levitate, you're definitely going to be able to walk in a world where trees are stationary! That's one of the interesting properties of this approach: it makes the robot much more robust.

It's also future-proof. You can suddenly change the environment on the system, and it very quickly adjusts its internal model and is able to operate. So there is an evolutionary technique that allows us to do that quite rapidly. It's very effective and cost-effective as well.

Michael Krigsman: Can you provide any advice or thoughts on enterprise AI adoption and the hurdles organizations face? And quickly, please.

Babak Hodjat: Businesses have gone for the low-risk, obvious use cases of generative AI. Without knowing it, I think they are already "agentyfying" things. They've gone in and said, "Hey, I want something like ChatGPT, but I want it to work against my proprietary data." So they've created these agents that can, for example, retrieve proprietary data. Then maybe they have another app, and they want an agent to operate that app. So they've created these single agents in isolation.

I think the next logical step is to have these agents talk to each other, link them to each other so you don't have to force the user to choose which agent to talk to. You can talk to one agent, and they sort it out amongst themselves. I feel like that's where we're going, but don't let it happen organically. You need to take steps to make this human-centric and responsible.

I think it also has the potential to be disruptive, so you need to think ahead about what this will do to your workforce and your processes.

So yeah, I think I'm an optimist. I think there's a bright future with agent-based enterprises coming our way. I just think that we need to go at it with open eyes.

Michael Krigsman: Babak Hodjat, CTO for AI at Cognizant. You've given us a lot to think about, and thank you so much for taking the time to be with us today. I'm grateful.

Babak Hodjat: Of course, I enjoyed it a lot. Very good questions. Thank you.

Michael Krigsman: Everybody listening, thank you for watching. Before you go, subscribe to our newsletter, and subscribe to our YouTube channel. Check out CXOTalk.com. We have phenomenal shows coming up with amazing people between now and the end of the year. Check it out! Join us!

Huge thanks again to Babak for being with us. Everybody, have a great day, and we will see you soon.