Agentic Enterprise - What your CEO must know about AI - Hubert Misztela

Hello everyone. It's a great pleasure to present here for AI Engineer World Fair 2025. My name is Hubert Misztela, I work for Novartis where I lead AI researchers in drug design. But today I would like to talk about a slightly different topic, which is Agentic Enterprise: what your CEO must know about AI.

So today I'm going to walk you through a few different angles of AI, um, specifically workflows, network, and some challenges, and individual perspective as well, individual employees' perspective, of how we can start thinking a little bit about the future of how AI agents might be used, and how they can impact our organization, and how we can prepare for that.

So let's get into it. I'll start with a little bit provocative statement, which is: in 3 years, maybe, your organization will be run already by AI agents.

Again, maybe provocative, but I believe it's still realistic within the pace of AI and AI agents which we haveright now. Let's define first what AI agent really is. As some of you might already know, AI agent is a large language model-based application exhibiting certain autonomy in its behavior.

AI agents are composed of language model which glues together specific things. It executes with planning, step-by-step by step fashion, usage of different tools,right, and the usage of those tools can be adjusted depending on the outcome of previous slabs and on the steps, and that's, that's called dynamic action.

Why it matters? Because it this provides us with another level of capabilities, previously almost unique to humans. But before we get to that, let me say one thing, and that is: many companies already, and every company, should be working on already on digital assets and really morphing them or evolving them to for the usage of AI agents.

So you can think about it this way: every digital asset in your organization should become a tool for AI agent, retrieval-augmented generation itself, or an agent. And these capabilities are really: reasoning for multi-step planning, adaptability for dynamic work, uh, control, persistent memory, contextual talk related to the fact that some sometimes we want to use only specific knowledge which is not available in language model, and this is where we start talking about retrieval-augmented generation, or interactive sun workspaces like sandboxes or canvas where we iteratively can work with an AI agent as a collaborator to refine specific job, like for example code or a specific piece of art.

And last but not least, capability is: computer using agents. Those can navigate graphical user interface to help us automate steps, uh, which previously, again, only humans were, were, were able to do. And sometimes, uh, robotic processes automation.

Now, let me go to

first perspective, which is process perspective, or in other words, workflow. Workflow is the ac-actually execution of a certain process. So imagine you have a process in your company and you want to execute that process. You define a workflow, which is defined by, in this case, 5 steps, sequential steps, executed from left toright.

Some of those steps might be commoditized simple, others might be specialized specialized tasks. This, uh, this design, beautifully simple but very explanatory, was, was, uh, done for the first time, I guess, by Sanjit. Recommendable blog posts, wonderful reads.

So starting with that, I wanted to point out that technology might be used in two ways. One, automation. Another one is augmentation. So automation really is that we substitute with, with the technology-specific step, and augmentation is doing that step or quicker or better.

But wait a minute, couldn't we do that before with machine learning? Yes, yes, some of the tasks we were able to automate or augment with machine learning, or maybe even without machine learning,right. So what's the difference today? So the difference today comes with agentic workflow, and that's, that's where the title of the, uh, this presentation comes from.

So why can we do itright now? We can do itright now for because of two reasons. First reason is the fact that language models, or specifically, uh, AI agents, um, can tackle broader set of tasks operating on our thoughts or cognitive steps, usually represented quite precisely in natural language.

And the second fact is that language models and AI agents can act as a glue between those tasks. So we can compose them together and we don't need a human intervention, and human intervention would be necessary only to provide the feedback in a certain step when the agent needs it.

So in that situation, what you see in pink, we call agentic workflow,right. So with this, we can automate and delegate not only one but a few steps, and that's what AI agents enables us. And I want to focus on this aspect throughout this presentation, now from a few other angles.

So first takeaway is: to start building engine building agents, you need to first identify your workflows,right. Then you need to understand the exact context around them. And that's, that's a big, um, challenge very often because, uh, more often than not, that context, which is what kind of data is used, what kind of systems are used, what kind of transformation and steps are being executed on a machine, are usually only in the head of the executor of, of the employee, and usually it might be not logged anywhere.

Some of the companies might be doing that, others don't. But very often we conduct, um, specific steps in our work without putting it anywhere because this is, this is sometimes you have to work just quickly, you need to perform, you need to invent,right, and then you start repeating that newly creative process.

So that's a big challenge because that's buried under our heads and maybe indistributed even across the teams or systems. So that's, that's a first, uh, first thing which company wanting to make use of AI agents should focus: identifying workflows and identifying deep context around them.

So let's now try to take a closer look at the same process, at the same wor-workflow where we have a certain steps. Now imagine not one person but a group of people is executing that, uh, that workflow,right. In each step you can have one or two people in this case, and a manager.

This just tell you about the number of people,right, or the, the, the, the position they occupy,right. But does this really tell you give you the big, the big picture? I don't think so. Relying solely on the traditional roles and job titles might be not enough for, for usage of AI.

Because when you want to apply an AI agent, you need to understand exactly what kind of tools you want to use, uh, how you can plan, what are the conditions for changing the behavior. And that's where you need to get deeper.

And it's not explained by maybe documentation or maybe just the po-position, job description. But what might be very helpful is an interesting angle, and that's a network perspective. So when we take let me come back. When we take the whole network of our employees, we can start thinking about it as different archetypes or different personas.

So when you when we think about each contributor to the given workflow as a persona, then we can start getting a little bit more information. So let's, as an example, conceptual example, let's define here five different personas. One would be silent achiever, person with a, let's say, uh, lower communication but very good performance.

Then individual contributor, which would have both. Then the connector connecting non-adjacent teams in the organization, multiple teams, let's say. Multiplier, person who, who enhances the work of other team members and knowledge hub providing domain expertise or necessary information in the process.

So now, why am I saying this? Because when we define the personas in our, in our workflow,right, we can start thinking already about how AI agents might impact our workflow and our specific employees. When you think only about the job titles, that's a little bit harder.

But when you think, oh, they both have the same job title but that person behaves differently than that person, they really contribute in a different fashion. So that's what you when you can start optimizing for the agents. And one can say, okay, we have that information in our team, so that's not necessary.

But wait a minute, this presentation is mostly for the CEO, as I said. So you need to have a more detailed look at the organization at the global level to be able to understand what are the similarities and differences between the teams.

Because when you realize that, for example, um, let me speed it up a little bit. Uh, when you when you realize the, for example, that, um, there is a certain pattern repeating in two different workflows, then you can optimize for those two workflows.

So first takeaway around, around, uh, network of personas is that if you want your organization and people adapt to AI, then maybe looking at the personas might be helpful first step. And let me show you how. Let's now try to project how that same, that same workflow might look like with AI agents.

So first, as you can see, difference from the previous slide, let me come back one more time. Here we had two steps, and now these two steps merged into one step. Why? Because of one reason. Because of that silent achiever who got an access to coding assistant might become so productive that now he can he can do the work for a few people for two steps already.

On top of that, the same coding assistant might be used for other team members, and they are, again, more efficient, so they don't need to perform the work in two people don't need to, to perform the same work for, for one task, but one person per, per, per task in the, in the workflow.

Now, another agent would be agent magnifying the, the communication and the, and the impact of the multiplier persona on other, on other, uh, team members. So for example, previously that, uh, that persona was impacting only the adjacent team members, but now with the, with, with, with the help of AI agent, that can be magnified to all of the team members, and now they could become higher performance with a better communication,right.

So you see already that these things sum up together or multiply together even. Another agent would be for knowledge hub. So that probably is one of the personas which would be the most prone, again, personas, not positions. Personas most prone for substitution of an AI agent because you can, um, prepare a knowledge hub in such a way that it dynamically that's what I try to depict with a, uh, with this bigger arrow, that it doesn't support one or two specific team members but all team.

And then that person who was working on that workflow can now take this whole workflow at the bottom,right. And there is efficiency across the board which one can try to imagine. But my point here really is, um, it's not only about the workflow or the structure,right, because they will be changing very dynamically.

It's more about, um, understanding the archetypes or the profiles or the personas of the, of the employees working on different workflows. And that's where it would be much easier to start projecting and planning for the development of different agents.

Because imagine, for example, that you verify this workflow for, uh, you, you do this analysis for all of the workflows in your company and you realize, ah, we have the same issue always on the first step. And, and that, that's when you know what kind of agents you need to build for, for the optimization.

With this, let me move to another, uh, another part of it, uh, which is the list of emerging patterns. So based on the previous, uh, analysis, you can start come to the to a few questions and conclusions. Let me start with the questions.

First question would be: which tasks really in that workflow would be better for humans to be executed, and which of them would be better to be executed by AI agents? And, um, referring back to, to, to the challenge about the context, I would say that ambiguous workflows where a lot depends upon the situation and it's not clear how to define it, obviously it's going to be better executed by humans.

And in the situation when you want to keep the subjectivity in place, that's also when you want to still have a human in place. Because maybe you don't care about the subjectivity, maybe it's a feature, not a bug.

On the other hand, you might want to remove that subjectivity, and that's when you use the agent. You might also want to use the agent, obviously, when something is repetitive and tedious and, and you want to perform it perfectly in the same fashion every single time.

So there, there, there is, there is obvious or maybe not so obvious start of the analysis. Then we see already some, some emerging patterns. And one of them is that intelligence and the domain knowledge be-be-comes commodity, which means they become cheap, easily available, and that's not the ag-edge anymore of the employees.

So that tells us that the employees and companies maybe, because if a company has an edge because of the domain knowledge and intelligence, uh, on average, let's say, uh, that, that's not going to be a, a, an, a, an advantage anymore.

So one employee or a company has to start pivoting to get multidisciplinary know-how or get much deeper into specific, into specific area. So we see so also some patterns, uh, online appearing that in-engineers, and this is what I depicted there as well, that some of the senior engineers are being dragged from the engineering teams and working directly with the product, uh, with, with the domain expert-experts on the product because they can develop much quicker with assistants and they can communicate with the domain experts much quicker thanks to domain assistant agents.

Then we can use agents as super connectors. Uh, context understanding will gain more value in contrast to intelligence or domain expertise. That's, that's the difference between domain expertise and the context. I'm not going to get into the detail, but I wanted to highlight that as well.

And AI supercharger, supercharges and, uh, supercharges doers. So people who do and perform can do even more and quicker. So the magnitude of the difference might be huge. It some people say, talk about 10 times engineer. In reality, we might be talking even about 100 times engineer.

And less of middle grounds. I think, uh, what's going to happen is we're going to see a lot of pola-polarization in the positions, in the tasks. So we're going to be have we have to we will have to go specializing.

And those specializations might be very new. So let me show you what will appear or what might appear. We can see new roles in AI mature organizations like workflow miner, as I mentioned before. There are already companies work, uh, with a great experience around that,right.

Then you can have human AI orchestrator because, for example, you might need to orchestrate that interaction between human and AI and how you do that. That would be that position and so on. I'm not going to get into the details.

You get the idea.

Um, now, the big point for, uh, or takeaway for, uh, executives, in my opinion, would be how you can enable all of your organization to start adopting, uh, or democratizing all of this. So it goes step by step.

Obviously, step by step. Obviously, first you give the access to general assistants,right. That's probably in place in all of the companies. Then you give the access to individual assistants per employee with a memory, with adjustments, with understanding the context, with personal database and so on.

That allows the employee to work closer on this context and don't repeat every single time the same thing. Then we can get to new capabilities, and that's very important. If we want to build those optimized workflows,right, or maybe redefined even workflows, we need to build those agents very quickly.

And one of the angles of building those agents very quickly is building by the employees themselves. So there is a, the key fea key skill set, skill or capability, which I believe will change the organizations tremendously, is building the agents on the spot.

And we have those capabilities across the board with, with brilliant, uh, coding assistant tools. No or, or, or no code and low code tools,right. But before employees can get to that, they need to work on a cognitive skill, which would be self-awareness of what I really think at work, what kind of steps I really perform in my head and on different systems to get to that goal which is only visible at the very end.

And that's what everybody else only sees. I, I do much more inside of that process,right, inside of that workflow. So that requires cognitive, uh, self-awareness, let's say, to translate it into specific steps and then build with no code, low code, or maybe code agent to do that for me, as a lot of data scientists already do.

So another step would be digital twins like time travelers and memory to retain the, uh, uh, organizational knowledge. So imagine you have an employee with many years of experience. If that employee leaves, all the know-how is gone,right, with the exception of some of the documents or presentations maybe.

But now imagine you have an agent representing that employee and the knowledge and the previous expertise, and you can ask questions and you can time travel back, not like your lost history or, or somewhere,right. You can just come back and explicitly reconstruct that and ask questions.

Um, another level would be agents with multiple employees. So agents serving not only one employee but the whole team or the whole workflow. And then, of course, we're going to get to the swarm of agents, which we already have software ready.

And all of this is really enabled only now because we are getting to the mature point of the software, software stack around LLMs and AI agents.

So with that, let me go to another element, uh, which is the angle of, uh, challenges and opportunities. We all know about the, well, about the challenges, which is, uh, groundedness, uh, hallucinations, guardrails, security, and everything else. I wanted to discuss briefly here the bottom two of them, which is ethics and adoption.

Let's start with the adoption. So on theright, you see a table from DeepMind from, uh, two years ago, 2020, 23, where they discuss different levels of AGI or intelligence. And when you read it carefully, you're going to notice that this part of the table is when human drives interaction, and that bottom part of the table is where AI starts driving interaction or the specific execution.

And that might be an adoption problem. Are we really ready for not delegating but really giving the whole task to the, to the agent and letting the agent delegate subtasks to us? That's a big adoption culture question which, uh, which, which we need to resolve before we get, get further,right.

Of course, first we need to start, start applying simple agents. But the question is, are we really ready socially? And then the second aspect is the ethics of autonomy. So with that, we are, we are getting a few, a few, um, interesting but challenging questions,right.

Um, so it, and it's not even about, about, about, uh, traditional philosophical questions, uh, uh, like whether agents should act as A or B in a, in a certain situation. But it's rather first, we want the AI agents to help us and support us as value keepers,right.

But what kind of values do, do we want those agents to be stored? That has to be explicitly defined. We humans don't agree with everything, but we need to define something for the agents. That's going to be first challenge.

And the second challenge, which is even I think more interesting, but because that's, that's a new one, which is about the interaction and conflict between human and AI. How we're going to resolve that with another AI agent, justice agent, maybe.

These are the things which, which we will need to figure out. Uh, let me wrap it up this part of the presentation with the quote by John C. Lennox that the greater the machine's freedom, the more it will need a moral standards.

So interesting times because we need to start engaging more into moral, into ethics and philosophy, not morality, ethics and philosophy to operate our technology.

So

before, before I close with the summary, I wanted to point to a list of interesting opportunities. I'm not going to read them all, um, but I think, uh, this is a very interesting list to take a closer look for every executive.

And one of them, for example, would be to, um, simulate synthetic reality before doing activity in real life in the marketing, in the lab, or somewhere else. Or maybe, uh, understanding proactively loophole-like houses regulatory or, uh, finding M&A, M&A, I mean mergers and acquisition, acquisition opportunities and, and kind of scout for them with an AI agent and so on.

So this is all very exciting, all very interesting. I think, um, we will need even more discussions around philosophy about ethics, about, uh, social aspects of, of AI and humans and psychology. Um, and this is, uh, a little attempt to, to, to do that.

Although there is a long list of very interesting topic, technical topics which I'm not going to talk, uh, touch, uh, today. So a brief summary of what we discussed is, uh, agents. So I took quick technological perspective of, of agents.

And then we discussed, uh, the AI agents adoption from workflow perspective. Then we discussed it from the network perspective. And we touched also the challenges and opportunities. So executive summary, oops, that should be appearing as the last, but let me start with that.

If your strategy, everybody talks about the strategy these days online. And if your strategy still makes sense when you swap AI agent with some other word like machine learning or, or data, and it still makes sense after swapping, that means that strategy is already outdated.

Then I told you that agents make a difference because of new capabilities. Detailed workflows understanding is crucial for bigger benefit from the agent. We need to democratize the agents to fully accelerate AI agent re-resolution. Looking at the network of personas, which is a human aspect, is very important.

An organization must be ready for rapid transformation and the whole org charts might change drastically.

Also, a mindset shift is, is essential in terms of, um, redefining and being ready for different interaction with the technology, which might be on some occasions even leading us. And last, the ethics is, uh, still the crucial part of, of what we do, uh, in our lives.

So let me thank you with that. It's a wonderful conference. Thank you for being part of it. Let the context be with you. Take care.