#300 Fred Laluyaux: How Decision Intelligence & AI Agents Are Redefining Enterprise Operations

0:00To remain competitive in this growing digital world, companies would have to think about their decisions and the impact of their decisions across the value chain, breaking the silos, not just data, but knowledge and organizational. national side. When you look at the volume of decisions coming in and you look at the constraints that are basically driven by this human-operated model, our vision was from day one to move from people making and executing decisions supported by machines, by data, by software, to a world where machines would be able to make and execute decisions guided by people. I usually start by having you introduce yourself to listeners and explain what AERA Technologies is.

0:44And then I wanted to talk about decision intelligence broadly and then how it relates to AERA. But why don't you start by introducing yourself? Okay. So first of all, thanks for having me today, Greg. My name is Fred Ladueo. I'm the co-founder and CEO of AERA Technology. I've been working on this adventure for eight years now. Prior to that, I spent my entire career in the world of helping enterprise manage complexity with software analytics. So my career took me from Europe to the U.S., starting with data modeling, business intelligence, ERP, AI. Prior to building Aira, I built another company in a space called Anaplan.

1:34And prior to that, I was with SAP. So basically, spent the last close to 30 years really working on performance management with data and analytics. Okay. And this is all built around this idea of decision intelligence. Is that right? It is correct, yeah. And you've been pioneering decision intelligence for years, long before most people were talking about it. Can you explain what decision intelligence is and why it's becoming important? Well, with Piety or decision intelligence before it was called decision intelligence, when we launched ERA in 2017, June 19, 2017, I'll never forget that day, we said, welcome to the self-driving enterprise.

2:22Why did we do that? It goes back to 2010 when I was at ACP at the time and I wrote a paper around my vision, which was quite simple, related to the fact that, or driven by the fact that we believe that the digitization of our economy was going to drive three fundamental things. The first one is an acceleration, a continuous acceleration of business cycles. And if you go back, you know, a decade, two decades ago, and you looked at the planning and the execution cycles, you were counting in months, you were counting in quarters, you're now looking at real time. So that continuous acceleration has been proven a valid assumption to start the company.

3:04The second one was back to the digitization driving now a consumerization of our economy. And the result of that was decisions that would have to be made at a much finer grain. We used to plan and think with segments, with aggregates, with categories. And now we're thinking and executing down to the store, the SKU level, the consumer level. And the third element was vision that to remain competitive in this growing digital world, companies would have to think about their decisions and the impact of their decisions across the value chain, breaking the silos, not just data, but knowledge and organizational silos.

3:43So for me, again, that goes back 15 years ago, the combination of acceleration of business cycle, consumerization and complexification was going to drive an explosion in the volume of decisions that companies would have to make to remain competitive. And then you look at the way companies make decisions and the way they're structured. It's a large pyramid of 10, 11, 12 levels of hierarchy, right? sitting on top of a bedrock of transactional system. And people within this big pyramid are scrambling with tools, bespoke software, data analytics, workflow systems to actually make the decision. So when you look at the volume of decisions coming in and you look at the constraints that are basically led by or driven by this human-operated model, our vision was from day one to move from people making and executing decisions supported by machines, by data, by software, to a world where machines would be able to make and execute decisions guided by people.

4:50So that's really the foundation of era. This is why we started. We said, OK, we got to build the technology stack from the ground up to enable that process. People called us nuts and crazy and a bunch of funky names back then. And you fast forward eight years and decision intelligence, which is now defined as the digitization, the automation and augmentation of decision making is a category that's being, you know, it's an AI category. and the validation of the growth in the market and the interest in that category is the fact that in December or maybe January 2026, December 25, January 26, Gartner is going to release the first magic quadrant for decision intelligence platforms, which is squarely what we've helped define as a category for the last eight years.

5:47But how does decision intelligence differ from the agentic AI that's spreading so quickly? Well, decision intelligence leverages agentic AI and other AI and analytics capabilities with the goal of digitizing specific decisions for the enterprise. So the agentic is a tool, but the category that supports the full end-to-end process of digitizing decision is decision intelligence. So within decision intelligence, we're leveraging LLMs, we're leveraging all different kinds of capability, but it's not relying exclusively on that. If you want to make a decision intelligence a reality in your enterprise, you need to resolve four fundamental problems.

6:39The first one is the data. If you are going to make decisions digitally, you need 100 % of the information that is required for these decisions to be made to be available in a normalized data model. And that's a lot to unpile and unpack, but that's the first block. This can leverage LLM and Egentic, especially when it comes to pulling in unstructured data, which is new. Back in the early days of DI, there was no real way to leverage unstructured data. that agentic technology is allowing us to do so. The second is the intelligence. And again, there's a lot to unpack in the intelligence. First is data.

7:22Second is intelligence. Intelligence is the ability to reason. LLMs can be fantastic for that, but not always need it. You can have deterministic processes that allow some reasoning. But it's also the ability to do all the math that is required when you make a decision. You need to project. You need to predict. You need to allocate. You need to select. You need to calculate a bunch of stuff, which is not what LLMs do. So the intelligence is a combination of the reasoning and the calculation. And those are two different technologies, right? The third is the automation. When you think about decision intelligence, you don't have just a system that tells you what to do.

8:02It would not solve the problem I've explained earlier of the volume of decisions. You need a system that executes the decisions as well. So I'll give you an example. you have a, the system will predict an inventory stock out, meaning you will be ring out of stock and some customers are not going to get what they ordered. Well, then you have several options, right? Conceptually, you can make more products, you can transfer products, you can expedite, you can do a series of things. The system will calculate the different options, will select the one that's the most relevant for that specific combination of product, of customers in time, based on expected business outcome.

8:40I want my customer service level to go up. I want to make sure I get the revenue. So you select all of that. But then once you make that decision, the system is to execute your decision. It will go and create a stock transfer order, for example, in SAP. It will go and update your planning parameters. So it's the data, it's the reasoning, it's the execution, which is not an LLM process at all. And the last part, LLM play a big role as that as well, but not only in national language understanding and generation, is the engagement. How do I engage with the users? How do I learn from the users? Last year, our technology enabled our customers across different industries to execute about 25 million decisions, right?

9:26So a lot of that is automated. The system does the process I've just described on its own. But in many cases, you have exceptions. You need to go in front of a human. Then you will have your inbox, your ARA inbox, and you'd say, okay, a message for Greg from ARA saying, I've done the work. I've done the analysis. I identified a problem, I should say. I've done the work, done the analysis. I make that recommendation. But my confidence score for that recommendation is maybe below 80%. I need your input. So then you open the message, and the action you take is you, of course, need to understand the logic, the data, the source.

10:02You build the trust. But then you say, accept. reject, modify. So that action needs to happen with business operators who are working anywhere and everywhere. Could be in a warehouse, could be loading trucks. So that's where you have to have a level of engagement that's super smart. And the software will then capture your decision and execute it. So data, intelligence, automation, which is very important, and an engagement. And across those different tiers, you will leverage multiple flavors of AI, including agentic, which is changing the game for us right now. Yeah. Yeah. So decision intelligence or Aira's technology stack, the agentic part is kind of at the bottom of the stack.

10:51There is a lot more going on in context analysis and that sort of thing. Are there other systems that are doing this kind of thing? You mean systems inside our technology or other vendors out there? Other out there in the world, because I haven't heard it described this way. As you have in the market right now, you have two, like always, when a market is forming. So first of all, I would say that in the last year, what we've seen is the word decision became ubiquitous. Everybody is talking about decision. I can guarantee you that even three years ago, nobody was talking about decision. They were talking about AI on one end.

11:33They were talking about data systems and data lakes and analytics. But now everybody's converging toward decision. We talk about process intelligence, talk all this good stuff. Everybody's converging toward decision intelligence, decision automation. Because that's the future. That's what's obvious right now for large enterprises. So if I break down the market for you, I would say a lot of the systems of differentiation, which is if you think about the enterprise stack, if I trivialize it, you've got your system of records, your ERPs, your CRM, you've got your systems of differentiation. This is the work.

12:09This is the technology that allows people to manually process data, collaborate, to make those decisions. It could be a plan, it could be a decision to repoint sales order, to allocate inventory to this. I'm talking a lot about operations, but it goes beyond that. And then you got your system of analytics, really basically fundamentally allowing the information to be distributed across the organization. Everybody who is in the category of system of differentiation is trying to now bring that intelligence to their system. So a lot of the vertical vendors, where a planning vendor, a procurement vendor, are trying to bring that level of service, basically, of automation to their software.

12:54Then there's another category, which is the pure play decision intelligence platform. And I can talk about why I think this is so important to differentiate the two. And in that category, you'll find Palantir. You'll find Aera. You'll find very few others that have been able to build a platform end-to-end. And then you have the do-it-yourself, which is a big part of the market today, especially in the early stages where large companies who have access to a stack are trying to couple things together to make DI work for them. And that's pretty much how the market is structured. Now, who's coming into this market?

13:30I talked about the system of differentiation. The analytics vendors are trying to get into this market. The RPA vendors are trying to go to this market. The ERP vendors, everybody is trying to converge toward that new Eldorado, which is decision intelligence. Yeah, that's fascinating. And what are the core components of decision intelligence? How do they work together? Well, I mentioned that a bit earlier. I think for us, we break down four key components. The data, I mentioned it, you have to be able to crawl. So the data is bidirectional. The data is structured and unstructured. The data is a big, that's where we come from initially, right?

14:08We said we got to resolve that one problem first to enable the intelligence and the automation and the engagement later. So if you just break down the data process, our technology connects to the transactional system. It actually crawls them intelligently. It can be a data lake if you have that, but we recommend you go straight to the source, which creates a lot of dialogue with our customers because there's a lot of questions about it. How do I crawl those systems without impacting their performance? Then the data, again, that's for the structured data. A big part of what we do is work on top of SAP, of the traditional systems of record.

14:44But you also bring in external data, unstructured data. It can be crawling websites. It can be looking at PDF for claims management or emails or images. So that's data. It's all combined into an indexed and structured into what we call the decision data model, which is a normalized data model. Why do we call this decision data model? Because it's a model that has all the metrics that allow you to understand your business, calculated, refreshed in real time in the system. But it's also going to contain the memory of all the decisions that are made with the decision intelligence. So the connectivity between the transaction, the augmented metrics, and the decision data is quite a big piece of software.

15:33Then, as I mentioned, you have all your engines. So the capabilities you need, I need to model my data. So it's a multidimensional structure that allows me to use formulas to cascade stuff, to project stuff. So you need to have that capability. We have a cortex, which is basically our machine learning. You need to have machine learning. You need to be able to do statistical forecasting, all the different flavors of AI. You need to have an agentic capability with agents, agent teams, agent functions that can orchestrate a lot of that logic that can also be leveraged to understand unstructured data.

16:16And so that's another big component of that. So you've got your analytics, you've got your data modeling, you've got your machine learning and other AI capabilities. You have your agentic capabilities and all of that gets orchestrated. Then you have the ability to write back and orchestrate the decisions back into the decision memory and into the transactional system. And as I mentioned before, you need the ability to engage with the users. And the way we looked at it, and it's very important because when people think about DI today, as the market is maturing, they think about a system that's going to generate recommendations, going to provide some intelligence that allows people to individually make decisions.

17:00Well, that's great, but it creates a lot of noise. And, you know, you're just pounding the users with, hey, I'm a smart system here and I'm going to tell you all the things that you should be doing. People don't have the time for that. You need to automate the work and you need to engage with them at the right level at the right time. And one of the foundation of ERA after 25 years building software, I'm going to square a little bit here, but I wanted an end user experience that didn't suck. And, you know, I've looked at my previous companies and, you know, all those systems are required. They introduce a level of complexity that forces people to learn how the software works and get familiar with it.

17:40And how do I... We said we're going to move from, and it sounds marketing, but it's not. It's really part of core to our vision. Going from a people that people have to learn how to... Sorry. Moving from people that use a software that they have to learn how to work with to a software that learns how you work. a software that learns how you decide. And that's what we built. For that, we needed to create that experience just like Siri or Alexa. This is what ERA is called, ERA. Four letters, A-E-R-A. It took a long time to come up with a name that you could call in any language. And you talk to ERA.

18:21You engage with ERA in natural language. Now, when we started, we were already doing demos of that technology eight years ago. That was literally talking to ERA. but it was very scripted. It worked, but I had to engage with that. Greg, I would say, Ara, what is the evolution of my forecast this month? And if I didn't spell it properly or spoke it properly, the system would not understand. The addition to LLMs in the last two or three years has been fantastic because now I can have a conversation with my decision intelligence agent that is completely natural. I can say, what is the volume of blocked inventory between Stuttgart and Frankfurt over the last month.

19:03And I'll say, break it down by product category and then show me the ones that are impacting this customer. And I use this language, the LLMs and the agents that we built in the system that allow me to have a natural language conversation, not just using words to interact with the software and keeping the context and having composite questions. All this is super important. So that was part one. engaging as much as possible in actual language. And the second part was for us very critical is what do you do every morning? You check your inbox. You go into your email and that's what we do. So you use Aera.

19:44You can ask questions, get answered. You can run simulation. But you fundamentally start your day by going to your inbox and seeing how many messages in English or in whatever language you want properly documented as a system prepared for you. And I don't need to train people on how to use an inbox. but I have to train people on how to use a complex software with a bunch of buttons and windows and processes. So taking care of that, taking care of the workflow dynamically is all the stuff that we had to build. So the components to wrap up, data, intelligence, I broke it down between the engines and the agentic, the ability to orchestrate sometimes in a deterministic fashion, a simple process, the automation and the engagement.

20:25And this is the stack that makes DI a reality for some of the largest companies in the world today. Yeah. Now, you talked about memory and a learning loop. Is the memory offline? Is it in a graph database or something that gets updated? And the learning loop, how do you avoid overriding previous memories or previous information? I mean, this catastrophic forgetting that everybody has worked on. So it's a good, it's a super interesting topic. When we talk about, I describe the stack, I describe the components, which is the technical part. But when we talk about what does AIR do for you? Now, we have a question.

21:21You go and meet many, many companies all the time. Last week, I was doing a great conference with Hershey's, and they're deploying ERA right now. For the end users, what is this decision intelligence agent called ERA that you have access to? What does it do for you? First thing is it answers any question that you have in natural language. The second is it allows you to do simulation. What if planning, collaborate. Third is it generates recommendation for you, for your job. You do your job every day. Your job is to reduce, manage inventory or reduce waste in components. The system provides all of that.

21:56It executes the decisions. It learns. So ERA understands. ERA recommends. ERA executes, right? And then it learns from the decisions that you made. A couple of things. It just doesn't learn from the decision that I make. If you're a company like Unilever or Orshies or Dell or whoever, They have for a specific function, you'll have dozens or hundreds, sometimes thousands of operators at scale, right? So the key is that the system needs to learn not just from my decision, but from the decisions made by the entire network of people. So how do we do that? It's quite tricky to develop and code, but now it's deploys very easily.

22:38There is a couple of components. The first one is you got to build the data for the learning. The data for the learning is basically what is the context of each individual decision that is being made. So I could have backpedaled a little bit, and if you allow me, I'll do that. One of the biggest questions that I had when we launched ERA was, it's very simple. I've been working for 25 years or 20 some odd years in analytics, and the message for all the tools that we built and all the companies that we created around that was, we're going to help you make better decisions. However, the dimension of decision is not captured anywhere.

23:18If you look at a system of information today of any large company, you say, where do you capture your decisions? Where are they memorized? Nowhere. Now, it is in some specific sectors. If you look at automatic trading, they do. They know that. They've done it. They're ahead. But if you look at the world of operations and sales and finance in an enterprise, you don't capture that. And I'm not just talking about strategic decisions, which is, do I go into this market? No, I'm talking about the decisions every day. I have that many sales order. I've got that much inventory. I've got to match this.

23:55Those are micro decisions that are very important. I need to ship my goods a certain way. I need all this stuff that the thousands of decisions that you make every day, they are not memorized. You rely on people to drive that process and the decision stays in their head. Now, they follow guidelines, they follow rules, they follow everything. But the first thing we said is like, how do I digitize a decision? And our approach was to go A-B testing. Like if I come to you, Craig, with a recommendation fully documented that you can understand. When you press accept the recommendation, now I'm digitizing your decision, right?

24:32So what is the decision memory? It's the context of the recommendation. So the business context, what are the metrics that have been driving the recommendations to be generated? It could be on schedule. It could be on demand. It depends on event. Then I capture your decision. I capture when the recommendation was made, when you made the decision to accept. Maybe you rejected it. If you rejected it, the system should prompt you and say, why? Select from the list or enter a little text. Well, Fred, this is the reason why in East Cleveland at that time of the year, we have something that wasn't captured in the software.

25:11Great. Or maybe you ignore it. Why you ignore it? Or maybe you modify the recommendation. That needs to be captured. It's memorized. All right. And then what the system does. So now I've got the context of the decision, the expected outcome. I expect that to generate X amount of dollars in saving. I expect to generate that much improvement in service level. I expect to retain. We have a fantastic customer called Western Governor University. It's one of the largest or the largest digital university in the United States. And they'll be speaking at our conference in a month. They're leveraging decision intelligence to help with student retention.

25:49So I'm not talking about logistics. I'm not talking about operations. I'm talking about leveraging data, make generic recommendations to help retain students. and the same logic, I need to have the context of the decision, the recommendation, the decision itself, and then what is the expected outcome. And as the system runs and refreshes every day, multiple times a day, we have a client who refreshes the model every 15 minutes. The system will see and identify whether the expected outcome was reached or not, and if not, to what degree. And that's when you bring a feature that we build called auto-decision learning that allows you to calculate over time when there is enough data, when the data is rich enough, it will calculate the confidence score for future recommendation.

26:37So now you look at your inbox and you have a recommendation to rebalance inventory from a DC to another DC. It will explain the root cause, the logic, the expected outcome, and it will say, ERA is confident at 92 % that if you say yes, the outcome will be reached. And the system continually learns. It's not a number. Maybe if you had run the same scenario three months later, maybe the confidence score would have been 95%. I don't know. So that's one part of the learning. Now, there is another dimension, which is super interesting. With Agentic, if you're leveraging agents and you're prompting the decision logic, which is we're doing now with some of our most advanced customers.

27:24This is brand new technology. It's the ability to really use agentic to run the decision logic. Now, there is a closer loop, faster loop for the learning, which we're experimenting with, that will allow me to understand from if you said no to that recommendation and you gave me a verbal explanation five times in a row, I can now start detecting faster than from the data whether there is an adjustment that could be made to the logic. But that needs to be also very much supervised. The first model that I described, which is auto-decision learning, doesn't require human supervision. It calculates.

28:01It's math, right? This one, you'll see the prompts coming up saying, well, Craig believes that this recommendation needs to be improved, the logic needs to be improved, or the threshold needs to be changed. Based on his input, I can identify a pattern and then I can go manually and adjust or not automatically adjust the decision logic, the prompts. That's new. That's starting where we've got one customer fully live on that now. And that's going to create a decision learning process that will be faster, but will require more adult supervision, so to speak. Yeah. I'm interested in the mechanism of this memory and learning loop.