We are entering a new era where intelligent AI agents are transforming how businesses operate. These agents go beyond the simplistic large language models (LLMs). They can reason, take actions, and work autonomously to complete complex tasks. This on-demand webinar explores the rise of these intelligent agents—what they are, how they work, and why they matter. It examines how agents will change the way we work, from creative teams to technical professions. As agents handle more routine tasks, we (humans) will need to focus on what we do best. What skills and mindset will we need to stay relevant and thrive in this new world?
Watch the webinar to learn:
- What are AI agents and why it’s such a big deal today?
- What kind of outcomes can agents deliver — from productivity gains to elevating human potential?
- What change is needed for organisations to take advantage of agents?
Speakers
- Michael Wu, Chief AI Strategist – PROS
- Douglas Mesquita Rocha, Global VP of R&D Innovation – MARS
- Hamish Taylor, Moderator (Former CEO of Sainsbury’s Bank and Eurostar Group)
Full Transcript
Well, hello, everybody, and welcome to today’s session from dumb LLMs to intelligent agents, brought to you by CXO Sync UK in association with our sponsor and our lead contributor today, PROS.
My name is Hamish Taylor, and it’s my very great pleasure to be moderating today’s session. Now those of you that know me from previous sessions will know that I am not a technology expert by any stretch of the imagination.
My background took me from consumer goods in Procter and Gamble where I was an expert on housewives toilet cleaning habits. Through consultancy with Pricewaterhouse, I was head of brands at British Airways, best known for things like putting beds into airplanes, and then chief executive, first of all, of Eurostar, the high speed trains that link London, Paris, and Brussels, and Sainsbury’s Bank, one of the world’s first supermarket banks. Now I’ve spent the last twenty years, however, as an adviser and speaker on customer led transformations.
How does the business world transform while keeping the customer as the primary focus for the innovations that are happening? A theme that’s gonna underpin today’s discussion.
Now during my career, the world of technology has changed dramatically. Whereas in my early days, customer data was contained in a filing cabinet in the corner of the office and then later on onto something like an Excel spreadsheet or maybe I should say Lotus one two three just to show my age. Today, of course, with the explosion of available data, that wouldn’t be possible, and organizations have had to adapt accordingly. And if you move into today’s era, of course, the latest technology innovations are in the world of artificial intelligence.
Now it’s easy for somebody of my background and age to think of artificial intelligence as a single entity that you now need to learn how to use in order to enhance your business. The reality is, of course, that artificial intelligence itself doesn’t stand still and is rapidly developing and creating new possibilities and challenges for us all. So today, in the spirit of customer led transformation, we’re going to bring together two experts to discuss how AI is developing, one from the development side and one from the customer side. And we’re going to look at how we’re entering into a new era where intelligent AI agents are transforming how our businesses operate, and in particular, how we’re moving beyond simplistic sort of large language models into an era where AI agents can reason, can take actions, can work autonomously to complete complex tasks for us.
Now to discuss this, I’m gonna be joined two experts. The first one I’m gonna bring on in a minute is a familiar face to these events in Michael Wu, the chief ex strategist at PROS. And I’m also delighted to be welcoming a new face today, Douglas Rocha, who’s the global vice president of r and d and innovation at Mars. First up will be Michael in a minute, and he’s gonna introduce us to the world of AI agents, explain how they’re used, and help us understand the possibilities for the future.
Then Douglas is going to come on, and we’ll talk about some of the practical applications that he has seen today in his business. And then I’m gonna put the two of them together to discuss the future. Now a couple of things just before I bring Michael on. On your screens, you will see there is a discussion tab and a question tab.
Do join in the discussion tab if you’ve got thoughts or just wanna say hello to your friends, whatever. Do do do put any comments you’ve got onto the, discussion, tab. But, equally, please do if you have any questions to ask our experts, there’s a tab called questions there. You’ll see as well at the bottom.
Bottom right, it is on my screen. I’m assuming it’s the same on yours. Please do put any questions you’ve got on that. I’ll be monitoring on the way through, and I will turn to those questions towards the end of today’s session.
Now I’m glad to see that some of you have started to use the the chat to say hello. That’s great. So let’s get into it. My first guest today is Michael Wu. Now Michael is the chief strategist at PROS, a position he’s held for the past seven years. Prior to that, he had an academic career with UC Berkeley where he had a bachelor’s in applied math, physics, and molecular stroke cell biology.
He has a PhD in biophysics specializing in computational visual neuroscience using statistical and machine learning techniques.
And he spent ten years also as the chief scientist with Lithium Technologies.
So, Michael, let’s bring Michael onto the stage. Let’s see if he can join us.
We just seem to be missing ah, here we go. Michael, look. Great to see you again, and thank you for joining us at what must be pretty early morning in San Francisco, I’m guessing.
Yes. It’s it’s early in the morning. It’s it’s my pleasure to be here. I know.
It’s a sad thing. Whenever we bring you on, we always seem to put it at a time that can enable it makes you have to get out of your bed early, so we do apologize. But, look, it’s great to see you again and and a and a great topic to cover off today. So, I thought maybe we’d start with look. You know, I’ve heard you in our conversations describe twenty twenty five as the year of the agent. And perhaps it might be an idea to start with a brief understanding of what an AI agent is and and kinda how we got here.
Sure. Yeah. I think so the best way to look at this is that, you know, OpenAI, you know, who developed ChargebeeT that most of us know Yeah. Has a model that describe how these GenAI evolve.
You know? Basically, they they have this five stage model. You know? GenAI basically starts, you know, with a chatbot.
You know? That’s the two thousand twenty three, you know, when we had, you know, ChattyPT.
So today, I think everyone is very familiar with this large language model.
Right?
They Yeah.
Chatibiti might be going with.
Yeah. I’m with you so far.
Yeah. I think anybody who, you know, can speak a a language, it doesn’t even have to be English. Right? I mean, can can use, you know, ChatTBT, you know, in in a in any way that they they wish.
Right? It’s just like actually talking to to someone with with a lot of knowledge already encoded in their, into their neural network. So so these are the chatbots. Right?
And that’s the first stage. And then the next stage is what, you know, they call the large reasoning model. So these are, you know, large not large language model anymore. They’re large reasoning model in the sense that they can actually reason and they could think.
Okay? Because they can reason, they could think they could think. Right? And one of the to understand, you know, what this brings to the table, you know, we have to understand that large language model, it’s great at creating text.
They could write lots of stuff for you, poems to emails to to anything you want them to write. Right? But with limitation. Right?
They’re they’re actually hallucinate. They they create things that don’t exist. Right? So so you cannot, rely on them for purely factual kind of, information.
Right? That that’s kind of, the thing. But it’s really, really good for creative work. Right?
Because if you if you write if you wanted to create something new than no one has has, has written before or create some design, right, LMs are great for that. But they’re not very good for math problems or engineering problem, coding problem that actually has a precise answer. Right? Because it’s long as, like, too creative for that.
Right? So large reasoning model is endowed with this capability to think. So now when you actually ask it some math bomb, it will actually reason it out step by step, right, and actually arrive at the answer. So so that’s large reasoning model.
Right? And that’s primarily enabled through a process what we call chain of thought, You know? And and if you ask the language model to, to think and think through the answer and give it give you the the thinking step by step, right, then, basically, you will see the effect of that. Right?
Basically, you will actually not just give you the answer if you ask some question. You actually think through, okay, the you know, all the logical stuff or how you actually arrive at the answer. Now so that’s large reasoning model. But the, you know, large model is very, very smart.
It’s it’s been shown that they actually can, you know, beat human experts, PhD level, you know, beat, you know, experts. Yeah. They they on multiple, variety of subject and domain. Right?
And It doesn’t make you feel threatened then as a person, as a PhD yourself.
Yeah.
Yeah. Yeah. I feel so small now compared to, this large using model. But the key limitation that they have is that they are passive.
You know? They actually cannot take actions, and that’s where the next stage comes in. That’s the the agents. Right?
So the key difference between, you know, agent and these l m and l r m is that they can actually use tool. Okay? They can use tool and which, you know, enables them to take action.
Right? To to now they can actually you know, if you ask, like, LOM or LRM some any question or anything or or no, He could tell you precisely how to achieve what you the question that you ask. So if you if you ask them to build your table, you can actually tell you describe everything in detail. How do you need what material do you need, you know, how to actually assemble them, and and and what do you need to to build a table.
Right? Right. But at the end of the day, you still need to build that table. Right?
You don’t have a table. Right? You still have to yeah. So the different with the agents are now the agents could, you know, access tool depending on what the tool is.
Right? So maybe the tool is an API that that could control some robotic arm. Right? Then you can actually go and actually, you know, you know, you know, you know, get the material and then actually, you know, construct the table for you.
Right? So that’s the difference. The main difference is that, like, agents can use tool. And these tool may be a robotic arm in the physical world, and it could be some digital tools, you know, APIs to some other programs for you to take action.
Right? For example, if you need to build a table, I need to buy some wood, I need to buy some paint. If you give it the access to, for example, to to to use a credit card, right, then it will be able to, you know, access that tool, you know, of of using the credit card to buy the material that’s needed to construct the table. Right?
And and then later on, you may access a different tool to control a body arm to assemble those those materials and and, you know, to turn that into a table. So that’s why agents are are so, powerful because they can actually do things for us and actually, you know, help us complete tasks, even complex tasks with multi steps and and, achieve our goals.
Wow. I mean, I was with you until you told me that it can take my credit card and start using it, but, you know, up till then, I was fine.
You didn’t have to give an access to that.
Yeah. No. I under understand the difference then. So it really is quite a a big advance then, the the the development of agents in that it, you know, it’s not just advising you what to do. It can actually take you down that journey.
Great. Well, look. Thank you. That’s given a a great picture of kind of where we are now, and I know we’re gonna talk about the future later on. But before we go any deeper into it, let’s bring Douglas on because I think, as I say, from my point of view, you know, the key to all of this is the customer.
How useful is this technology to people that are actually, you know, in in the businesses wanting to use it? So I’m gonna turn to the customer side now and and bring somebody in whose whose career has been spent at the forefront of of r and d and innovation in the consumer goods world. And for I’m a for whom I’m assuming has had a whole new world opened up by these technological advances. So Douglas was educated in chemical engineering at university in his native Brazil, followed later by a career with an MBA at, Manchester University and participation most recently, I know, in the Cambridge University chief technology officer program.
He began his career with Unilever, which gets a boom from me as a procter and gamble man, in both Brazil and in the UK, and followed by positions in several of the good and the greater consumer good, Reckitt, L’Oreal, Henkel. And he’s currently the global vice president of r and d and innovation at Mars, based once again back in the UK. So welcome, Douglas. And where was P and G on your career list?
It’s not. And I and I differently from you or equally as you, I hope it will not gonna be Hamish. I think it’s almost kind of the dark side of the force, you know. You know how how it feels.
Would it be the side. Absolutely.
It’s a fun part of the career in P and G or Unity, but that, you know, in the first part of your career, they are seen as the big rivals. And then it’s great when you mean to eat people from the other side as you go through later on in your career.
And you really and and you can be partying to the other end of a discussion that was going on with competition.
And as we progress, we see how how similar they they can be as well.
So it’s, it’s just a Jogues Apart is a company that I also respect a lot as well.
And, and likewise, you know, I mean, that was the whole point. You know, you knew we’re our greatest rivals. You have to you have to respect them as well. Absolutely.
Well, look. I mean, as I said, to to to Michael, what I’m keen to do is bring the two of you together. So we’ve got, you know, the customer side and the development side and discussing how the you know, what the possibilities are and how they can be applied. So, look, let’s start off at the beginning, though. You know, you’ve worked in a variety of kind of world leading organizations since your career started twenty or five years ago. And over that time, you must have seen some kind of major technology driven changes in the approach to to to r and d.
First of all, it’s a pleasure to be here. Thanks a lot for the invite. And it’s, it it’s really nice for me to be learning a lot from Michael, myself as well. So I think all the conversations that you had so far, it was super enlightening.
I think to to your point, when I started my career in nineteen two thousand, the way that we innovated, the way that we are using, AI and not AI on that part on digital. It was actually almost non existent. It was very physical, linear, and normally very slow. We used to run experiments and maybe you will remember this from your time in PNG as well, Himesh. But in a very way that we gather the data, we iterate step by step. And over the last twenty five years, digital and and somehow AI have been the biggest biggest disruptors as well.
So they really shifted the way that we do our engine product development today from a way that we are primarily lab based to being really increasingly data and, driven and and and predictive.
So in the early days, even the digital tools were fairly done. We use already for for quite, number of years and decades already statistical models, of course, simple automations or knowledge databases that could really spin things up. But these all these models, they couldn’t really think.
Then after some years larger large language models they came along. They gave us new ways to interact with data but it’s still in a very static way. Basically, you ask a question, it gives you an answer but it’s useful but it’s still very limited.
What’s exciting now and also builds on all the things that Michael explained is that this shifting to intelligent agents, really cannot only that cannot only interpret but also reason all the data that we have, take action, and operate autonomously is really exciting.
For RNG specifically, this means that moving from tools that support decisions to agents that can really design and execute parts of the process, whether it is creating thousands of new ingredients as we do in our day to day or running simulations overnight or even orchestrating lab equipments, automatically. This is, again, is a huge opportunity.
And suddenly, we are not only just accelerating science but we we are augmenting, science with the systems that can work really alongside us.
These changes also means a lot to me specifically as a leader of r and d because it changes a bit my role. So my role today is no longer just about managing scientists in labs and investments from different equipments and so on. But it’s also about integrating digital, educating our talent, on, artificial intelligence, and also, of course, building trust in these systems as as we progress and as, new developments come. And really ensuring that we use these technologies in a very responsible way.
The big challenge that we have and also the the opportunity that we have is to combine creativity of our scientists and the empathy of our sciences with scale and speed that these agents, they they bring us. So just to close, I believe that in the next twenty five years, it will be really about co piloting with AI and co piloting with these agents. So it’s not just about humans versus machines, but really about these intelligent agencies as colleagues, as new associates as well that could somehow free us up from repetitive work and some work that today in the past used to take us a very long time.
Also, to solving problems that truly require, you want imagination and and judgment.
It’s an interesting I mean, I think talking about this business about, you know, what what’s actually changed and taking away, you know, tasks that that can be dealt with now, which you don’t you used to have to do yourself. And you see the same thing in the finance community when they talk about accounting tasks that they used to have to do, they no longer have to. Now that obviously opens up doors for you as you’ve described in terms of things. Absolutely.
But presumably, though, it also means you’re having to recruit on a kind of different profile to maybe the the the pure biology scientists or whatever you were recruiting in the past.
Absolutely. No. It changes completely the the the approach. I think as as you said, I think it changes completely the way that we recruit people, the way that we educate people. But also, and maybe bringing a little bit to the reality that I have today. Ai is not something that is theoretical for us. It’s already part of how we innovate, every day.
The key is that we are not just using generic tools and in in more from open markets, like char dbt and so on, but really developing applications trained on our own data, which is something that is really important for our area of work.
This allows a lot to personalize to to generating sites that are truly relevant to our categories. In in my case, currently, pet nutrition. And, ultimately, to to really move faster than is marketing how, we innovate.
And, So you’re so you’re able then to mix a combination of external expertise with your own data without giving your own away your own data to other people then?
Absolutely. Absolutely. And maybe I can elaborate in maybe three ways that we use Yeah. Please. AI today.
Because I think it helps to ground the example a little bit more. So in R and D and in innovation specifically, I see three big areas that we are also investing a lot of money. One is accelerating innovation and ideation itself. So, traditionally and, this is valid for all categories. Today, I’ll I’ll try to link a little bit more with pet nutrition that’s there at the at the work, but this is valid for home care, for for for everything.
Even for toilet cleaners like I have.
Even for toilet cleaners. I think the the case is the same. They are as I said in the beginning, I think they’re much more similar than than than we think. Yeah.
Traditionally, we’d run consumer research that costs a lot of money. We would collect lab data. We would test hypothesis, one by one. Today, basically, what we do is that we build AI tools on our proprietary data that was developed with years and years of different tests and so on From formulations, from different consumer feedbacks. And this allows us to generate ideas much more quickly because we can use a lot of all the the ideas and the data that we created in the past. So, for example, instead of weeks and weeks of manual analysis of this data that a scientist would have to do, by hand, we can ask, basically, an AI agent to screen thousands of past formulations, link them with consumer acceptance acceptance data, for example, and propose new directions really overnight. This was the work that used to take, months, and now it’s it can be done really in two days.
One second area that I’m particularly passionate about, and I’ve been deploying that a lot recently, both in Mars and in my previous role in L’Oreal and Henkel as well, is the definition of technology roadmaps.
We can use AI not just for product ideas, but also to shape long term, research and development strategy. So AI can help us scan scientific literature, patents, start up activity, investment flows from different companies. And really this gives us a live view of where technology is moving. Because basically we can gather all this data and make a reason out of this data much quicker. This means that the roadmaps that we create today, they are not no longer static PowerPoints as as I am sure that you used to see in the past, Hamish. But they are really dynamic and continuously refreshed with the changes that we observe. And this allows us to see both risks and opportunities almost in real time.
The third point it would be, really future consumer needs. So one of the most powerful applications that we have today is demand for sites. So combining behavioral data, social listening that we can have from from ecommerce like Amazon and so on. In early trend, signals with AI, we can really spot these shifts in consumer. And it’s much earlier than we used to do. This allows not only to respond faster but really to personalize innovation to consumers.
And, and this is the real value that we can add from that. So it’s really to tailor made all these, these observations to the business context on on on that moment.
So maybe just to conclude, for me the key is not really AI replacing scientists and marketers but really augmenting their capacity.
Really by building these personalized AI agents and to train with our own data and I think this is crucial. We are making R and D much faster, sharper, and really much more predictive than it was in the past. And, the future, as I alluded in the beginning as well, is about people and AI working side by side. Humans bringing creativity and judgment and intelligence systems really bringing speed, scale, and and, foresight.
Great. So your your your r and d colleagues don’t need to worry about their jobs at the moment then? They just have to change Yeah.
That that’s that’s my job. It’s it’s to protect my job as well.
Brilliant. No. Let’s bring the two of you together for for for a few minutes. I mean, I I guess, Michael, one of the things Douglas has talked about there is this idea of using the company’s data and and and combining that with expertise. So how does LLM learn one company’s data? Excuse me.
Sure. I think that that’s a a very interesting question because, you know, we all know, you know, from, you know, news or or out there that training a large language model is very, very expensive. Right? So to retrain a language model to a company’s specific data, it is almost kinda cost prohibited.
So how can a large language model learn a company’s data? Right? And this turns out that there there’s a trick that you can use. This is called retrieval augmented generation. It’s called RAG, r a g.
And what that does is that, essentially, you what you do is that you augment. Right? It’s it’s called retrieval augmented generation because you’re using what we call information retrieval technology, such as a search engine, to augment that generative AI. Okay?
So you’re combining a search engine with a with a Gen AI now, the the large language model. So what what does is that, like, now, it gives the language model the ability to to read a lot of content that might be relevant to the questions that you’re, asking me. Right? So the way this work is that, like, when you ask just l m, some questions specific to your company, right, it probably will not give you the right answer.
You will hallucinate something out that’s is sometime by chance you’ll get it right, and sometimes you’ll get it wrong. And you you you know, in many of the use cases, you can’t take that chance. Right?
So, so now with retrieval augmented generation with Rag is that when you ask a question, it doesn’t keep the question to the language part of the right away. It give it to the search engine first. Okay? Okay. What does search engine’s gonna do? The search engine’s gonna basically, you know, search through your internal, document or or or knowledge base or some kind of repository, right, to find the relevant documents that might answer your question. And then what it does is that you tell the language model, read all these document that I find and answer the question based on what you have just read.
Okay? So it’s two step. Right? There’s a you when you ask a question, the question first goes to an information retrieval search engine.
Right? You find the relevant document. Now you have all this relevant document, right, that that could potentially answer the question. Right?
Now you ask a language model to read all this and answer the question based on what you have read. Right? So now it’s it has basically read your document and understand what you’re talking about and then use that to kinda answer your questions. So now it’s actually much specific to your company, because the answer is is based on a document that your company has provided it.
Right?
So Gotcha. That’s how those people of planet generation work. And that’s a a typical way, how a company can make L and learn a company’s data. Right? You you especially to make it anytime you ask them a question to ask them to to read a bunch of documents first, that’s that’s internal of the new company.
Yeah. Gotcha. Gotcha. It’s actually getting the right information together and that before it looks. Okay.
Douglas, Michael’s talked about the next stages up and in terms of PhD smart, you know, level smart LRMs and whatever. Do you have any sort of use cases or anything that leverages the super smart aspects that are, you know, the latest developments that are coming through at the moment?
Yeah. Some some cases, but also, I will elaborate a little bit on the way forward as well because and, again, bringing to the specifically to the scenario that I that I lead today on on Mars Pet Nutrition.
Mars is a company not not everybody knows, but we on our division of pet care, we own hospital pet hospitals. We own, diagnostic clinics and we own pet nutrition. So we have an ecosystem that produces, as you may imagine, a lot of data. And today, we are starting to to use more and more AI to really make, all these science that we generate more accessible and predictive. So one example that we have of a tool that we create we created and we launched to the consumer already, specifically in the US. Michael may see it. If if you have a pet, you can give it a go.
Michael. So it’s, under the the brain, Greenies, we launched the risk, what we it’s called a canine dental check.
That basically we know from experience that eighty percent of dogs, they suffer from gum disease. But most owners, they basically don’t know it. So we trained an algorithm with hundreds of thousands of images of dog mouth. So now an owner can basically snap a picture and instantly get results and advice through our brand, through their website directly.
We also built, in a different brand, an AI model that predicts, for example, chronic kidney disease in cats up to two years earlier than traditional diagnosis. So this gives PEG vets and owners precious time to to to intervene.
But linking this with the theme of today that is really LRMs.
The two examples that I mentioned, they are really powerful events, but they are still much more not yet on LRM. They are more still more more like pattern, recognitions than the lithic.
The next wave that we are working on are really using and advancing these large reasoning models in diff, is is is a bit a bit of a different approach. So they can actually solve the problem straight away, weighing trade offs across health, nutrition, sustainability. So all the ecosystem that I said that we have with Mars and cost as well. Almost like a digital of team, PhDs as you said, Hamish.
And, maybe some future cases on that. One, and these are areas that we are we are still working on it. So they’re not existing, but these directions that I can mention. So for example, personalized nutrition.
Imagine combining genetics, microbiome, wearables, and veterinary data that we have from our ecosystem.
So an agent can design a tailored diet for each pet pet, really balancing health and lifestyle needs for each each of the pets. Another one, it would be novel ingredient discovery.
An LRM that could, for example, scan liter literature, patents, supply chains to propose new protein or fiber, sources that really meet the nutrition and sustainability requires all requirements all at once.
And, and a third, for example, predictive health ecosystems. So how could, we simulate long term health outcomes for for pets, and for humans as well to some extent, under different diets, really helping us to prevent disease rather than just, treat it. So the opportunity is really enormous and unlimited. But success, as I said and I alluded a few times, depends on the data quality, transparency, and trust. So the owners and the and the veterinarians as well, they need to understand and believe in the science that we bring out of this.
So again, for me, the the future is not about replacing, them, but really reasoning agents, augmenting, the capacities for both scientists and, and vets, as well.
What what interests me, actually, when you say that is is what one of the other things I pick up from what you’re saying is that by using external data as well as external data or whatever, you’ve actually almost increased the collaboration between stakeholders.
I mean, you you know, you’re crossing over into the veterinary world. They’re presumably crossing over into your world a bit more. And and it means that the whole kind of supply chain for the health of a dog, if you like Absolutely. It Seems to be coming, you know, much closer together, the different elements.
Absolutely. Absolutely.
Good. Michael, Douglas has talked about, the the use of of of of of LRM moving in towards LRMs for him. What what do what do you see that most companies use LRMs for?
Yeah. I think so as I mentioned before, right, one of the greatest strength of this LRM is the ability to reason and solve power. Right? So I think the, most company, you know, using this, LRM, you know, I mean, they’re they’re using them for for coding, you know, certainly to write code to solve problem, to analyze data complex data.
I think I think a simple LLM, a large language model, can write some very simple quote. Right? For example, a SQL query or something like that. You could do some simple code.
But if you have you know, if you wanted to build an application, right, an entire app. Right? For example, that like Douglas mentioned, in building those app where the owner can stamp a photo of the dog’s mouth and then actually get, you know, a diagnosis or or something, some kind of even a treatment plan. Right?
I think that’s a really complex coding project. Right? You need to analyze image. You need to install it.
You need to do a lot of things. Right? So to write all those code, right, is probably too difficult for LLM. Right?
You will basically generate a whole bunch of code that probably, you know, doesn’t work. Right? So a large reasoning model has been surprisingly effective at writing code because they could break down the problem into smaller problem and do that, you know, successively, however many level you want. Right?
You could break a big problem into small problem and then break the small problem into even small problem. And then eventually, they they will say, alright. This this is the problem that I know how to solve. I could write code for that.
And you pitch them all up and then into an entire application. Right? So this is, the really amazing capabilities of this LRM because Yes. Capabilities.
But, Mike, Michael, if if an LRM is so smart, as you say, as smart as a PhD and could do all these wonderful things, why do we need agents then? I mean, have you got an example of where an agent can perform a task for my business that wouldn’t have been possible if I remain wedded to the use of, you know, LMMs and and and our LRM?
Yeah. Yeah. I I think that that’s a good question because, you know yeah. And as as I mentioned before, the main difference between these LLM and LRM an agent, is that LLM and LRM, they’re passive.
They cannot take action. Right? So so, like, for let me just give an example to make that very vivid. Right?
If you ask, LM or LRM to to help you create some kind of, the email marketing campaign, right, it could tell you through excruciating excruciating detail all the stuff you need to do to to achieve that. For example, maybe the first step is to use to analyze your CRM data to identify and segment your your target audience. And then for each segment, you need to create some kind of, personalized email for each segment, you know, to to to target. And you have to send them the send the email, execute that profile.
Right? And then, you know, get response data and then analyze last step maybe to analyze performance. Right? Now today, right, it will not be able to do all that for you.
Right? It could tell you how to do it, but you have to do for example, like, you know, the first step to analyze CRM data to to supplement the audience. Right? How does it get access to CRM?
Yeah. It doesn’t have access to that. Right? The the language model, unless you give it the a tool, an API to allow it to access the CR, you will not be able to get the audience data and set it for you.
Right? So creating a personalized email from the second, maybe the l m can do that by itself. Okay? Now the next step, sending the email.
How does a a l m send the email? It cannot if it doesn’t have a email kinda, a server, right, how does it able to send the email to each one of the the the customer, you know, or reach, you know, that you have in each segment. Right? So you have to give it access to another tool.
Right? In in in addition to the access to CRM, that tool, you need to give it to access to a email tool or email server. Right? So that’s another tool you need to give it access to.
So you can send the email out. Right? And then the the fourth step is to get the response data. Right?
So maybe how do you how does the CRM how how does a a a LM or or LRM get those data? I think you can’t. Right? So if you have the you got access to some yet another tool.
Right? For example, the marketing content management system that tracks the the email response. Right?
And so you will be able to get those response data. Right? And then, to analyze the performance in the last step, right, maybe you could have a LRM to help you write some code to analyze those data. Right?
So notice that this agent, right, there there are many things that you cannot do by itself. It will need access to different tools. Right? One of the tools that you need to have access to is the ability to, to access a CIM and and retrieve data from there.
Another tool that you need to access is the ability to send email. Right? And another tool is the ability to get response data from some kind of content management system.
So without these access to these tool, the LLM or the LRM can only tell you what to do, but you have to go and do it. Right? When he’s creating all this email for you, you have to somehow cut and paste those and send it out. Right?
Few views on. So so it’s it’s not very effective. Right? But now, I need the agents to be able to automate all that.
Gotcha. Gotcha. So so, Douglas, okay. Let’s turn back to you for a minute. Okay. So you’ve heard what Michael said about agents and what they could do that that you couldn’t do previously with the LLM, LRM, etcetera.
So, you know, there’s all these there’s agents out there. Do you have a view on what the top agents would be that you would employ if you were given a choice tomorrow or or or if you can’t actually name me as agents? What what would you have them do for you that you can’t view at the moment? What what what’s the exciting future that Michael’s painting look like in your in your world?
No. That’s, that’s a very, very interesting question, you know. And, I think that if I could have all the agents in the world really working for me as part of my team, I focus really on the ones that could take away repetitive work, which is a bit basic, but also to connect complex dots and and free, free humans, free scientists to focus on creativity and leadership.
So I see maybe three top agents, let’s say. One, it would be an agent that’s more like a, let’s say, a discovery agent. So an agent that could continuously scan, global science, patents, startups, clinical data, all these data that we have from different source, and not only summarizes it like we do already today. I know it is becoming already kind of basic with Copilot Chargebee and so on. But even, really, an agent that could reason all this information about which are the breakthroughs that are most relevant for, in my case, today, pediatrician, but for any area of science.
That will definitely save enormous, time and help us really focus on the science that really matters and then making these things reality.
Another one would be maybe, personalization agents. So, one that could integrate thinking about our ecosystem as I alluded before. Could integrate the information that we have from genetics, from microbiome, from lifestyle, veterinary data, really to create personalized nutrition plans for each pet. So today, we know how to do pieces of this, but it’s very manual and fragmented. So it’s not reasoned.
So an an autonomous reasoning agent that could make that, it would be in a seamless way, in a scalable way, it would be great. And, maybe finally, as a as a leader, I I personally would love to have an agent that could take strategy, break into action, plans and coordinate resources and flags risks at all time at re in real time as well. Because today, as you both know very well, we often spend a huge effort translating vision into execution. This is it still takes lots of reasoning.
And an agent that could handle that orchestration could be really transformative. So, in short, I really would like, agents that really can extend the science, that we create today, personalize at scale, and really take away the heavy lifting of the execution.
No. I I I love that. And, I mean, what I like actually is the way you describe also that when you’re taking away those tasks, those routine tasks or or other tasks that they know that that that that that need to be get done in the business, you you know, your answer is not, and therefore, I can get a great cost saving. What you actually said was, and therefore, you know, I can concentrate on the people and the leadership and that side of things. And I think that’s a challenge in a lot of areas, finance and and other areas as well, where there’s a tendency to think that when machine when when when AI generally can take away tasks, the result is a cost saving. Maybe maybe the result is better leadership, better cooperation through the Exactly. Etcetera.
Absolutely. Absolutely. Look, Michael, you’ve meant he’s Douglas has mentioned two or three areas. He’s talked about the bit about people in leadership and taking away tasks.
He’s talked about joining complex dots. He’s talked about taking a vision and strategy through to action or whatever. How does that fit with your view of of of of the tasks that could be automated in the future? And and then maybe you could talk a little bit also about the things that get in the way of that at the moment.
What are the what’s the what’s stopping people jumping to this brand new world of the the mass adoption of agents already?
Yeah. I I think that is, another great question. And I think the so agent, as you can see, is very, very powerful because they can take action. So they can take action to help you accomplish goals, Right?
Accomplish complete task and accomplish your goal. Right? So I think there, I definitely could see that in the very near future, all of these agents, that Douglas have envisioned will be built, will be possible. You know?
I think that we we are, on the way there now. So two thousand twenty five is the year of agent because this is a year that a lot of companies are building agents. We PROS also are building lots of agents on our own. You know, but I think there there is some, I would say, bottlenecks for mass adoptions.
I think one is is primarily, the because this Gen AI, right, had developed so fast, there’s not a standard kind of communication protocols for these, what we call agents, so far. Right? There are some there’s actually quite a few that’s been developed. Right?
So the the most popular one is is something called the MCP, model contact protocol, and that’s actually developed by Anfarpic, the company who, developed, Cloud. Right? And, so and another one is the a to a, protocol by Google that they developed by Google. Right?
So these protocols enables, LLM or LRM to understand and discover what tools they have access to. Right? Remember, the agent is only as good as the tools they have access to. Right?
But how do you know how does the LM know what tool does it have? Right? For example, the the example that I gave earlier about designing an email campaign.
So how does a what does an l m or l m knows what it had it it can send email. Right? It has to know you have to you know, what’s the protocol? What’s the you know, what how should they send this email that they generated to, the particular email server that’s underneath, you know, in your firewall to send out those email to the client.
Right? So we need to know what tools are available and how to use those tool. Right? So that’s actually done through the MCP server.
And then but keep in mind that no agent is actually, you know, all powerful and can can do everything. Right? So sometimes agents need to get help from other agents as well. Right?
So they need to know what other agents can do. Right? And that’s actually done through the agent a protocol, agent to agent protocol, to discover what capability another agent can can bring to the table to help you accomplish the goal that you want to accomplish. Right?
So these protocols are, you know, being developed as we speak. So, you know, obviously, you know, they did not complete none of them. For example, if you wanted to access another agent developed by another company, you know, for example, maybe they wanna charge you a small fee for using their agent. Right?
But today, none of these protocols is able to handle payment yet. You know? So these are, you know so and maybe there are several agents that can actually do similar, job. Right?
And, you know, how do I know which one to pick? Right? So none of these protocols actually handles reputation, the ratings reviews. Right?
You know? So, their performance history. Right? So none of them so yeah. So right now, these, communication protocols are very, I would say, limited.
Yeah. And then the second, I would say, is that the security and governance and privacy concerns, you know, when you have a chat agent talking to outside agents. And then this last one is, integration challenges. So these are some of the challenges, that we see.
Over time, these will be solved, you know, because I said, you know, this year is the year of agent. A lot of people are building agents today, and they are sorting this out. And we already have some agents, built, you know, on the the PROS platform. Yeah.
So I I have no doubt that in the near future, all the agents that, you know, very near future, that all the agents that Douglas have envisioned will be out there.
Good stuff. That’s a nice exciting future. Now we’re rapidly running out of time. We’ve got about we can do it.
We can do another five minutes or so. So I’d quite like to just, if we can briefly, have a look at a couple of the questions that have come through. But before I do that, there’s been an awful lot to cover in the forty four minutes or whatever that we’ve had so far, and and I’m sure people would like to learn more about this. And if you would like to learn more about some of the exciting developments that Pros is making in the kind of eigenic I AI space, then follow the link that’s in the chat.
Okay? In the chat, there is a link that will allow you to get access to how you can find out more about what’s going on in this space at the present moment because, obviously, in a fifty odd minute, you know, webinar, we’re not gonna be able to do that. But I did promise people we would come back to the questions. We’re gonna have to keep it very, very brief.
But I did we’ve got three questions that have come up here that I wanted to just quickly, quickly put their put put to you guys. I think probably they’re more in that, Michael, I suspect these. Well, you may have an appointment. So if I can ask for, like, really rapid fire quick answers on these ones.
First one comes through from Michael Hilbert who says, do you think customers can actually tell when AI is influencing the price they are offered? And if so, does it put them off?
Actually, you know, so can they tell? I think probably, you know, yes. Because a lot of the pricing, you know, are actually done by AI today. You know, I think I think especially for large enterprise, you know, that with, very often these pricing are done by AI already.
I think whether they will put them off in fact, I think we have actually research that suggests that, you know, customer trust AI pricing more than they trust human pricing because they know it’s subjective. It’s based on data. It’s actually people actually it’s actually you know, it’s them more acceptance rather than turning them off.
You know, somebody is trying to pad their margin. Yeah. Ash, there is actually a the if you look on back on the library of the PROS webinars, there was one we did talking about pricing, a while back, actually. And we did talk about this and the links that you know, it’s not dissimilar to what used to be done manually or without AI when, you know, things like airline pricing models or or or holidays, you know, time out or whatever. So it’s not that different. It’s just automated, really.
Right. Here’s one dear to my heart from Ali, who’s asked, is it possible to build an agent within Microsoft Outlook that automatically categorizes your incoming emails based on their content and attached documents.
I mean, you can do it by uploading to GPT, he says. But is there is there a way you think of building an agent that would allow Microsoft Outlook to be easier for us to use when we come back from holiday with, you know, a hundred and fifty emails sitting in the inbox?
I I think they yeah. The answer is definitely a yes. I think Microsoft themselves are actually building some of these agents as well, you know, through their Copilot, ecosystem. So they’re actually, you know, building it through Copilot and and so Copilot will basically be available in their email client, their PowerPoint, their Microsoft Word, the entire Office suite to help them, do whatever task that they want to automate. Great. Yeah.
Good. Good. Thank you. And, again, very quick answer on this last one. Patrick’s asked, when introducing data driven pricing models, back to pricing again, what tends to be the biggest challenge in getting sales and finance the teams aligned?
Yeah. I think this, change management is probably the most, challenging. You know? I think one is that, you know, the ability to make these pricing more transparent so people could, understand what’s going on.
Right? People don’t trust, something that AI just give it to. This is the how we should price. Right?
Because these people are not just novice. They they have worked in the company for a long time. They know how pricing works. Right?
I have always offered this price, and it works for me. Right? But why are you telling me to change the price? Right?
So so very often, these, pricing algorithm need to provide what we call explainability Yep. So that people could understand what has changed, right, between, you know, last time and this time that, that would drive this change in price.
And that explainable AI is actually crucial to the adoption of Thank you.
Right. Last two minutes, guys. Last three minutes. Okay? Let’s close with a little bit of advice from each of you on how to get started on the journey to enhancing our use of AI and, specifically, AI agents. So, Douglas, have you got a piece of advice to those who have maybe not come as far down the track as you have about how to get started and really understanding the future possibilities?
I think mine links quite well with the previous question that on this point on creative management. You know? My clear device would be, start small, start personal and start kind of responsive, you know?
Right.
Don’t really necessarily wait for the perfect enterprise solution from Microsoft or so to to come. Really pick one area where you think that, AI agents can really add value with the data that you already have and build from there.
Make sure your teams are working with these tools. They really understand them. They really trust them and they really feel empowered by them. I think this is really key. And always keep, responsibility front and and center. You know? So transparency, ethics, and human oversight are really non negotiables in my opinion.
And in my opinion, if you do that, you will not only capture the benefits of AI much faster, but you also build the foundation of trust that is essential for scaling this across your organization.
Fantastic. Thank you very much. Start small and work up. And and and I love as I said before, I love the idea of of just, you know, the people bit is still absolutely critical in all in all of this. Exactly.
Michael, a a a last thought from you in terms of of what what people should think about, who want to go go go down this track a bit further than they are at the moment.
Sure. I think, you know, it depends on what stage you are. Right? I think if you are in the early stage, definitely take advantage of ILM and use retrieval augmented generation, right, to start making this AI, you know, know how to answer question that’s relevant to your company.
Right? And then if you are in the second stage where you are using this log reasoning model, right, have it, you know, write code for you because, actually, you know, you know, a lot of I mean, it’s it’s been, you know, a a challenge for a company to to hire enough engineer to help them code up because every department uses some kind of software, some kind of, you know, and they just simply they would love to have an engineer, but they cannot do that. But now they with this LRM, they basically can have an engineer on their side. Right?
And if you are on, you know, moving forward to this gigantic, world, then definitely, you know, experiment with actually using agent to start automating your work. And I think that will give you productivity that’s not just like ten x, but hundred x, thousand x. Right? Yeah.
It’s it’s yeah. It’s exponential. It’s fantastic.
Look. Super. That is really good. Unfortunately, we have run out of time, which we always seem to do on these sessions.
First of all, a big thank you to Douglas for joining us. As I said at the beginning, you know, a lot of the work that I do is in this whole area of what I call customer led transformations. And it’s great to bring together a developer and a customer talking together here about the developments, what’s coming over the horizon, how they might be used, or whatever. So, Douglas, thank you so much for joining us today.
We really appreciate your input. It’s been really, really enlightening. And I’ll even forgive you for working for Unity for one point in your career.
That’s a big pleasure.
Yeah. Not not believe me, that’s a big that’s a big move. No. Thank you. Thank you, Douglas.
Michael, as ever, it’s been an absolute joy. I tell you what, honestly, the way you’ve been educating me over these seminars is fantastic. I’ve learned so much from you, and it’s really, really interesting to see what, you know, what might be coming up over the horizon in the future. So as ever, thank you so much for for your contribution to this.
Thank you, PROS. As ever, you know, you you you’re picking on some really great topics, and you really are educating us. So thank you, PROS, for doing that.
Please do look at the link in the chat, as I said, if you want to find out a little bit more.
Anil in the background from PROS, thank you for your organization as ever and to the CXO sync team. Most of all, thank you to you, the listeners. I see you’re pretty much all still with us. One or two had to leap off at the end as we got slightly over time.
But, you know, you’re still with us. So, hopefully, there was some useful stuff in there. And thank you very much for for listening and and joining us. And please look out for future sessions from PROS.
I know they’ve got some very important topics to cover, and I’m sure we’ll get into in much more in-depth on some of these topics.
My name is Hamish Taylor. It’s been great fun to meet these two experts and have a chat with them today, and I’m off to look up how I can now use AI agents to sort out the biggest problem that I face at the moment, which is my Goldpanda catalog. So, thanks very much, everybody, and I very much look forward to seeing you again at a future session. So for now, thank you, and goodbye.
Thank you.