In this blog post, we recap a panel discussion from the HybridAIMS 2024 conference, which explores the impact the potential convergence of hybrid artificial intelligence and enterprise modeling could have on information systems.
Converging hybrid AI and enterprise modeling
What is the future of information systems? And how will technology like hybrid AI and enterprise modeling influence these systems?
In a panel discussion held at HybridAIMS 2024, hosted in conjunction with CAiSE ‘24, these questions were explored by leaders in the AI and Semantic Web community, such as Dr. Giancarlo Guizzardi, Dr. Peter Haase, Prof. Dr. Dimitris Karagiannis, Prof. Dr. Alessandro Oltramari, Prof. Dr. Oscar Pastor, and moderated by Prof. Dr. Emanuele Laurenzi.
Keep reading for an explainer on hybrid AI and enterprise modeling, and gain insight from the featured panelists on the future of information systems, knowledge graphs and LLMs.
Table of contents
What is hybrid artificial intelligence?
Artificial Intelligence (AI) topics are essentially inescapable in all tech environments, online and off, with its reach extending across almost all major industries. That’s because its potential for impact is huge. However, AI is merely an umbrella term that covers a myriad of sub-concepts that operate within it, and many have varying definitions of these terms.
Before we dive into the discussion, we’ll define the two prominent types of AI that were also the focus of the panel.
Symbolic AI, according to Oscar Pastor, professor at Valencia University of Technology in Spain, “is based on the idea that intelligence can be represented using semantically meaningful symbolic rules and representations of these rules.” This is also considered the traditional AI, and what was considered AI in its early days — it is reliant on logic and rules-based reasoning.
“While ‘deep learning’ – [this] is what I see what people refer to when we talk about sub-symbolic AI,” says Pastor, “is based on an idea that is different — intelligence emerges from the collective behavior of artificial neurons that are connected to each other.”
Sub-symbolic AI is machine learning-based, where output is generated from processing data in a neural network system, mimicking the neural pathways of the brain. It can handle more complex and nuanced tasks; however, its black-box nature makes it difficult to understand the reasoning for the answers or predictions it generates.
When you combine the strengths of both symbolic AI and sub-symbolic AI, the result is Hybrid AI, which offers a combination of both rules-based reasoning and data-driven learning. Dimitris Karagiannis, professor at the University of Vienna, argues that Hybrid AI is a technology and distinguishes between data-driven AI and knowledge-driven AI. The former observes the past and identify patterns in raw data, whereas the latter makes knowledge structures available to machines and therefore enables future innovation.
Yet, there is still a level of explainability that cannot be achieved without a conceptual model base.
Pastor explains, “There are two different types of knowledge. One is predictive, the sub-symbolic one. It's predictive knowledge because it's machine learning-based. It's knowledge that comes from data… this is something very concrete.
From data, we train a function and we infer valuable knowledge, but the knowledge is the result of the function, and the only thing that we can ask for to explain the answer is to look at the function because it’s the function [that says] yes or no.
The other one is explainable knowledge. Predictive knowledge [is] machine learning-based [and] explainable knowledge [is] conceptual model-based. Because we have a conceptual model, we have an ontology that is represented through a model, we can explain, [and] we can explain even predictions in the best case, but not always.”
He concludes, “What a model AI platform should have, also in enterprise modeling, [is] a clear predictive knowledge perspective and a clear explainable perspective, and they should be connected. This is the big challenge—how to connect, in the same platform, two different types of knowledge.”
What is enterprise modeling?
Karagiannis stated that enterprise modeling refers to the capturing and modeling of concepts, processes and terminology that exist within an organization. The process requires collaboration from multiple stakeholders to contribute specialized knowledge (whether domain-specific or organizational wisdom, for example) to the model, to explicitly define these concepts and the relations between them.
Enterprise modeling includes the explicit definition of terms through a vocabulary and enriching data with contextual meaning so that the data can be consumed and understood by both humans and machines.
Now that we have a shared understanding of the terms used in the panel, below is a transcription of the panel discussion with questions by host, Dr. Emanuele Laurenzi, and responses from panelists, Prof. Dr. Giancarlo Guizzardi, Dr. Peter Haase, Prof. Dr. Dimitris Karagiannis, Prof. Dr. Alessandro Oltramari, and Dr. Oscar Pastor.
Question 1: “How do you envision the future of information systems and what are their advantages?”
One of the first questions the panelists were asked was about how they envision the future of information systems.
Prof. Dr. Dimitris Karagiannis began the discussion by first expressing his definition of hybrid AI and enterprise modeling. “To me, hybrid AI is a technology, and enterprise modeling is domain knowledge. I took the definition of enterprise from enterprise modeling as a human endeavor characterized by motivation (e.g., purpose, drivers), participants (e.g., actors, agents) and platforms (e.g. resources, infrastructure).”
When considering how he sees the bridging of the two, Kargiannis says “I cannot say I can bridge AI with enterprise modeling. It’s too abstract.” He says there first needs to be an agreed-upon definition and consistent understanding of these terms within the scientific community and more broadly, which is not necessarily the case. Instead, Kargiannis considers the question from a data and semantics point of view.
“If you look at these two views, whatever you pull about data is [already in the] past.” He states that we need to consider the fact that data evolves so quickly. “We have to observe the past, but in this observation, things have already happened, and to innovate the future.”
When it comes to bridging the two views, it’s quite simple, he suggests. “You put data and you get models. It’s very easy, it’s not difficult. The other way around, you put models and you get behavior—or knowledge”.
Dr. Oscar Pastor also suggests there needs to be a clear definition of what is considered “hybrid AI”, which he believes is a kind of symbolic and sub-symbolic AI. Referring to the question, Pastor says, “The role of large language models in this enterprise modeling context is a kind of high-level assistance, where we have to select the right tasks that we want to do, use the right prompts, and teach the machine to do what we want it to do.”
To Dr. Giancarlo Guizzardi, a professor at the University of Twente in the Netherlands, “Information systems are basically in a sense like digital twins because they are digital shadows. They are representations of things in the world, so they have little people, little relations, little events that move inside them, and they should keep track [of] everything that is happening outside. So it's a basic data quality feature, but they also create things in the world.”
Two core points emerged from the conversation, with the first being about trustworthiness and explainability. Guizzardi envisions a future where information systems are trustworthy, meaning they must be built on conceptual models that are reliable and accurately represent real-world concepts. To achieve this, he believes that introducing constraints will help distinguish intended interpretations from unintended interpretations. According to Guizzardi, the biggest challenge to achieving this goal will be connecting information systems that already exist. “We have to come up with this structure—[a] core conceptual model with the right constraints in a distributed and collective manner. And that's a challenge that we don't know how to solve,” he argues.
The second point that emerged was the prediction that large language models (LLMs) would play an expanded role in providing assistance, particularly through increased integration with agents and a wide range of tasks, ultimately being used for nearly everything. Pastor argues that “The role of large language models in this enterprise modeling context is in a kind of high-level assistance, where we have to select the right tasks that we want [it] to do, use the right prompts, and learn and teach the machine to do what we want it to do.”
Dr. Peter Haase, founder and Chief Scientific Officer at metaphacts argues that generative AI and specifically LLMs will serve as an interface to a very broad range of specific tasks across diverse domains, such as with target discovery or drug repurposing systems. However, he specifies that large language models will need to be combined with other modalities and technologies in order to provide the results required:
“[As] of today, large language models excel at certain things, language understanding, capturing, and general knowledge - but some things are missing. First is that grounding in really explicit knowledge in knowledge bases. Large language models are not databases or knowledge bases that you can query directly. Second, LLMs are not for storing and retrieving facts, you cannot query [them] for knowledge directly. And what you need is trustworthy, explainable answers based on the knowledge that you have in the organization. So there are deficiencies, but all these deficiencies can be complemented with the capabilities that we get from elsewhere, in particular from the knowledge representation side. That's where this symbolic–subsymbolic interface [takes place].”
Haase concludes by saying that the combination of LLMs and explicit knowledge representation will result in solutions that have the capability to perceive, learn, abstract and reason, making it even more important to consider what the goal of AI systems are and reflect on how technology companies can ensure that machines really act in the interest of the human.
He says, “I'm saying that not because I believe there’s an inherent risk that machines might replace humans, but because it is humans that own, design and control work systems. And it is that intent that controls the AI. Therefore I also think it’s important, on a societal level, to look at how we can make sure that what is happening is really in the interest of humans.”
Dr. Alessandro Oltramari, Senior Research Scientist at Bosch, says “Hybrid AI-reliant information systems should contain a human-machine collaboration capability, aiming at high-level decision.
Such systems can be grouped in three different categories: 1) decision support; 2) decision augmentation; 3) decision automation. 1) Machines provide some basic tools to support human decision making, such as alerts, analytics and data exploration. The decisions themselves are made entirely by humans; 2) machines play a larger and more proactive role in the decision process. They analyze the data and generate recommendations and predictions for decision-makers to review and validate. Humans can make decisions based on the machine’s suggestions, or they can work cooperatively with the machine to amend the recommendation; 3) machines perform both the decision step and the execution step autonomously. Humans have a high-level overview, monitoring the risks and any unusual activity and regularly reviewing outcomes to improve the system.”
Question 2: What projects have you worked on or are currently working on or project proposals you are engaged with that contribute to the realization of such information systems?
Currently, Karagiannis focuses on realizing the equal collaboration between machines and humans, rather than viewing them as separate entities acting autonomously. Considering the framework of AI being the technology and enterprise modeling as the knowledge source, Karagiannis is interested in understanding “how to extract knowledge, because I think physical environment and digital environment of the future will be equal trade.”
“The point is that the business layer will have human-centric knowledge captured, and then we have now scenarios to digitalize, and [we need to consider] design thinking and digital design thinking and to set it as a digital twin.
When we talk about modeling and enterprise modeling, it's not enough only to have this digitalized, I need the structure. I need to see the models that we have behind them. We have to align because we have the capabilities with these new techniques and so many technologies to align.” He says that with real-time situation modeling a service, the service can now be executed in one hour.
“The understanding is not only for the understanding for the human, it should be also for the understanding of the machine. What we have we have here generative operations and we have here data generating physical teams.”
Pastor provides two examples of applications where the work he is doing is attempting to clearly distinguish between “predictive knowledge, machine learning-oriented versus an explainable knowledge, and shy conceptual modeling-oriented explainable AI based on sound ontological background with a conceptual modeling representation.”
First, he discusses a project with a company where they are trying to decipher the language of life and understand the genome in order to manipulate it. They are looking at concrete diseases such as identifying and diagnosing colorectal cancer. “It's a very clear information system, this clinical application of an AI platform with two dimensions. The predictive one is machine learning oriented.
We have hundreds of patients with colorectal cancer and genome sequencing processes and tons of data. With this data the predictive [part] is how to train a function, and with this data, [identifying] whether a new patient has or [does not have] colorectal cancer. This is done using machine learning techniques and the platform is just the predictive.
If a doctor asks or the patient asks, why [the function is saying they have colorectal cancer]. There is no explanation for that because it's the function itself that is [giving] the result, because it is a well-trained function that is giving this information and nothing else with the explainable perspective. The explainable counterpart, that is this conceptual model, where we know that there are barriers that are related to the generation of proteins that participate in pathways whose metabolisms has been affected and generated the problem. We explain that the problem is because there are some barriers that, hopefully in the next future, we are going to be able to edit and [correct] directly in the souls of the problem. This an ai platform, but we are combining machine learning for the predictive part with the explainable part, which is conceptual model-based, and these are very nice works that we could apply it to and this is what we want to do right now.”
Additionally, Pastor shares about working with his master's students to understand how large language models work so well and the limits of their capabilities. “From the enterprise modeling perspective, the role of large language models could be to provide high-level assistance. Again, there is a new chance for conceptual modeling. Maybe these large language models are going to provide a new opportunity through this notion of prompt engineering. Maybe the new models are going to be the prompts, because now you have to interact with the machine and the model and the way in which you tell the machine through a chain of prompts.
I did an exercise with a master's student to identify the type of contribution a paper in conceptual modeling research is making. I saw a framework that categorizes classifies the contribution in three types: type one type two, type three—either an artifact, a model, either a process of how to use the model (a procedure perspective) and how to use a model (that t is the type two). Type three, that is a PC. It’s not syntactic. There's some reasoning behind it. And when we [gave] the paper to the ChatGPT 4, and applied what my students are doing in their masters—we were just training as if it were a student, telling ChatGPT how to do it well—it’s amazing that it did it much better than all the students. It did all the classifications correctly.
So the point here is [to understand] how large language models, even in these cases where there is some reasoning, works so well.”
Moving on to Guizzardi, he says, “We’ve been talking about many things that are quite semantically overloaded, right? Explanations [are] one of them, we had some discussions this morning on ontologies as well. Even the term neurosymbolic AI is quite fuzzy these days. People mean very different things by neuro-symbolic or hybrid. There are some people trying to replicate symbolic using sub-symbolic. There are people that think that this is just any combination of techniques [from] both camps.
I have projects both on using symbolic to help sub-symbolic. For example, in cybersecurity, from things like using all these graphs or helping to understand or to extract graphs from natural language descriptions of security breaches, or to use generative AI to generate software patches for software vulnerability.”
On the flip side, he also explores “how sub-symbolic can help build symbolic models.”
“I want to go back to this point of having information systems as digital twins that are manipulating theories of the real world, which are captured in models endowed with constraints. I think the question is, can we do this in a completely data-driven way, can we generate these structures and constraints, which I seriously doubt.”
About this project, he says, “Traditionally, if we look at ontologies or conceptual modeling people, you have [it in a] completely kind of top-down way. You think about the domain and you do want to launch an analysis in this case, and then you come up with this authoritative reference model of that domain. In practice, even if you do it extremely well, there are things you are going to miss. So we have to come up with ways of doing this that are both bottom-up and top-down.
What we've been doing recently is understanding the evolution of modeling languages that will help us to create better ontologies and conceptual models. Of course, you have the top-down way, because you design this language using ontological theories, foundation ontologies, for example. But then, once these things are used by many people in many domains at different points of space and time, you can build a repository of models and use data-driven approaches to understand what's going on there.
We also have a project for understanding how language changes with time. The language itself, so how people bend the rules of language to say what they want to say, we call this systematic subversion of language. To understand these patterns which are useful for modeling, using a pattern mining approach, we are using a particular type of combination of model-finding techniques, a kind of formal verification technique that generates the logical models of a logical theory, plus machine learning but not using neural networks.
Again, this machine learning is also quite vague, right? An inductive logic program is a type of machine learning not using neural networks, and what you can do there is to come from examples. With this background theory, which is your model itself, you take the model, you generate all possible interpretations, faults and quotes in a finite sense, and people judge if those interpretations are intended or not. By doing validation, they are curating a database of examples and counter-examples, so you can reverse this process in an inductive way and learn the missing constraints.
In these two examples of projects we are using, let’s say, techniques more in the sub-symbolic camp. Using interpretation to view the construction of symbolic models.”
Haase then shares his projects with metaphacts as part of Digital Science, “With Digital Science, we're trying to support researchers with different kinds of research tasks. Also different domains, and media and industry, for example pharmaceutical industry, research practices on network discovery based on different data modalities, unstructured content, public data, proprietary data and so on.”
The first project is around explainable research tasks. “A research task might involve, ‘OK I want to find relevant publications for my research question’, then maybe you want to ask specific questions: ‘What's in this publication? Are they really grounded in actual scientific knowledge?’, and you may want to summarize things and extract data from publications to analyze that. Also, you don’t only want to know what is known today, but with scientific discourse and discovery, you want to generate a hypothesis and see what what is not known. Try to discover new things, but always be rounded in existing [rules?]. These are typically quite complex workflows and we want to give some specific examples of such workflows and how we can utilize LLMs and teach them and train them, with new capabilities,” says Haase.
“One key capability that we're working with is [the] ability of the LLM to access structured knowledge with an approach called SPARQL. It has the ability to take natural language queries and translate those natural language queries into structured queries over knowledge graphs. SPARQL is the query language we are using in that context. Models like GPT-4 already have good knowledge of SPARQL, what they don’t know, obviously, is the knowledge graph that you're working with, and providing that background knowledge, knowledge as context, is partly done with prompt engineering.
We’ve also been tying that process with the actual knowledge graph and entity resolution, at particular steps, to where we know, ‘OK, so that's language [...], and what we’re referring to is that particular entity in the knowledge graph. And that works well. We’re not the only ones doing that, there are competitors who are doing similar things, and there’s quite a bit of research on that, [which] shows that doing this translation, using ontologies—so really having a knowledge graph rather than just a database—leads to considerably better results. That is one step in the direction of extending the large language model with the ability to talk to knowledge graphs, to access, explicitly represent.”
The other example Haase shares relates to the enterprise modeling side. “We wanted to explore how [to] use LLMs with a modeling partner, in particular—can we teach an LLM to design or develop ontologies? We published a paper about this and presented the result of that work, [which] originally came out of a hackathon with students. We took an ontology engineering methodology called Neon, already developed in the mid-2000s, and tried to see if we could implement this methodology using large language models. The motivation is that knowledge engineering—building the knowledge graph is very costly—and figuring out if we can reduce cost by automating—using gen-AI to lower cost.
We built an automated pipeline that implements that methodology, as the methodology consists of a number of steps, and with each of these steps we supported it with specific LLM capabilities—so prompts. [You] typically start with generating competency questions. ‘How can we extract concepts or relations, how can we maximize them, and how can we evaluate the ontology using the methodology?’ We supported each of these steps with LLMs and made a pipeline that can be run in a completely automated manner but with relevance for all products. We also wanted to look at assignment workflows, and see how ontology engineers can assist it.
The learning here was that it’s possible to implement ontologies through such pipelines. And a really important outcome was that using a methodology in that process really leads to improved results. If you teach eloquence, this methodology leads to better results than just saying ‘OK, generate an ontology for a particular domain.’”
“I see that really as one example of how we can teach LLMs or agents new capabilities,” says Haase. “The next step here, and what [we’re] looking at is, can we support the human, the ontology engineer, in the process of ontology engineering? What I see later on, also as part of the vision, is that we need the agent to be able to do certain things and to develop an ontology for its purposes. So if the agent is tasked with a particular research question, and that might require ontology, to, for example, integrate data from relevant sources, can [we] teach an agent how to do that? And there are methodologies described out there, [we] just have to learn the methodologies.”
Lastly, Oltramari goes on to share his project with Bösch. “We are working on decision augmentation systems at Bosch in a variety of domains. In one project we supported emission calibration experts at Bosch.
In a nutshell, the use case is the following: Original Equipment Manufacturer (OEM) brands provides a combustion engine to our engineers, who are tasked to assess if those engines—in different scenarios, tested in the field or in the lab—comply with the emission certification standards across the world. There are hundreds depending on which region you are considering (EU, US, China, etc.), and this very quickly becomes a data intensive problem, especially when we consider the regular modifications made to legislative norms on pollutant concentrations. If you think about modern car engines, they are equipped with a lot of sensors, and even when you look at the concentrations of pollutants, there are many contextual factors that influence that.
We built a system that can summarize data coming from an engine and combine that with the advanced, hands-on knowledge that lives in the expert mind. Then we created a knowledge graph (KG), and implemented rules that reflect the heuristics that the expert employs in their daily job. We formalized the KG—and combine that with more traditional machine learning techniques that can digest the data and reason over the behaviour of the engine.
We have recently started a project in industrial manufacturing, which manifests some of the key aspects mentioned above. A manufacturing facility with its different lines, stations, different tasks being executed is a data-intensive environment. Operators interacting with those machine have very specific production targets to meet, such as reducing cycling time of a station or improving quality or quantity of components being manufactured. Similar to emission calibration, in this context the problem is to combine real-time data at scale with the expertise of different operators, the features of the tasks they execute, the compliance with standards (global or specific of a company), etc.”
Question 3: How do you see the convergence of Enterprise Modeling and Hybrid Artificial Intelligence (Symbolic AI + Sub-Symbolic AI) contributing to such information systems?
The final question posed to the panelists asks how they envision the convergence of enterprise modeling and hybrid AI (combining symbolic and sub-symbolic AI) contributing to the development of information systems. This time around there was more diversity in the answers.
“The last question here is the most difficult one,” begins Karagiannis, “Because it’s something we have to predict. And predictions without assumptions are almost impossible in my opinion.”
Before answering the question, Karagiannis states, “I thought about getting started with this transparency because that is the invariant transparency of my presentation. We have a huge added value if we [understand] the information system and the future information system as an ecosystem. Optimizing one function is not solving the problem of information systems.”
“Therefore, we have to find out how we can deal with this situation—how we can bridge this request together. How can we get this domain knowledge, the knowledge sources which only the enterprise has, and get this innovation, and also to observe the past?
We identify patterns that happened in the past, and then we have to find how to optimize. I think we can do that by bringing also the second issue. This knowledge-driven AI.
I hope at least that the deterministic part will come from this innovation because I'm saying the deterministic, because I assume that we know what we like to produce, what we like to do, how our models should work. Of course, we have to have the feedback, we have to ask, we have to evaluate and so on.”
While Karagiannis hopes that a future workshop will explore beyond simply optimizing the function of the AI, and instead goes in the direction of “how to bring them together and set up the future formation system that's an equal system, in which the requests are more
than to optimize more than one request via ChatGPT or to make a prediction.”
“The convergence of enterprise modeling and hybrid artificial intelligence, this conjunction between symbolic AI and sub-symbolic AI, could be and should be and will be done through the combination of this predictive knowledge,” says Pastor. He believes that large language models will play a crucial role in combining machine learning-oriented knowledge and conceptual model explainable knowledge.
While referencing the skepticism around LLMs–such as an argument that they are not driven by logic but merely “works of words”, akin to smooth talkers who sound convincing but lack any substance—Pastor argues that LLMs actually capture human logic that is hidden in the language. The hallucinations that happen with LLMs are not just a case of being an AI problem, but a human problem, since humans are the ones who are deciding what is true or not.
Pastor is curious about the future significance of LLMs. He says, “I’m not saying that large language models are going to substitute human knowledge, but I wonder—what is the limit? Because they work so well in some cases where they were not supposed to work so well.” He also notes that LLMs have only become prominent in the last 1-2 years and their evolution has been rapid.
Guizzardi adds another perspective, that AI lacks true semantic understanding and is therefore not even smart enough to ‘hallucinate’. He suggests that having humans in the loop will be inevitable and necessary. “It’s not smart enough to hallucinate. When people hallucinate, they understand the semantic content of their hallucination, there is a semantic model through the hallucination, and [chatGPT] cannot do that in my opinion.
There is no semantics inside computers. Those semantics, it's just wordplay. There are formal semantics, a kind of mapping from symbolic structures to a semantic domain, which is a conceptualization of the world. They can only do that because there are people in the loop,” says Guizzardi.
Guizzardi also says that there is an issue of conceptual choices in data handling and differences in interpretation. When solving this massive semantic interoperability problem in enterprise ecosystems, he believes it’s still a semantic problem—preserving intended models. The problem of connecting different models in information systems is binary, like understanding language. What is said is either understood or not. Guizzardi says that understanding is a deflationary operation, zeroing into one interpretation, not a probabilistic or trial-and-error function, especially for critical applications. “Semantic structure [separates] the wheat from the chaff and in a certain sense, separating unintended interpretations from intended interpretations”.
Guizzardi wonders about the possibility of creating these semantic structures in critical domains, in the distributed context of enterprise ecosystems, and in a completely buttoned-up way: “You make the conceptual choice of which data to get right, which data [to] measure, what that data means—so the problem of data semantics, the problem of data integration, is a semantic interoperability problem, and the problem of telling signal apart from noise cannot be solved without having a theory about the domain itself.”
Haase poses that “there [are] semantics in these systems, similar or higher than in the average human. And this will increase over time.” He suggests that the symbolic and sub-symbolic approaches will not be merged, but that instead of a merged approach, there will be an interface, and LLMs will be at the center of the interface. The main reason being that the emergence of neural networks is not yet well understood, he believes that we do not understand how a neural system works and are simply trying to observe phenomena. We may understand on a basic level probability distributions and token prediction, but we do not understand emergent intelligence.
“It will interface with other capabilities, for example, in a knowledge base as a symbolic component. The neural type LLMs will learn how to utilize knowledge bases, and how to access structured knowledge, and they will also learn how to do other things in the real world.”
LLMs have the potential for the capability “to understand the resources that they have at their disposal. What kind of knowledge they have—this may be structured knowledge, data, [or] texts—[and] they will need to understand what they have at their disposal in terms of knowledge, but also in terms of tools,” says Haase.
“For example, [LLMs] will be able to use an external reasoner or a consistency check in the same way as we as humans do. [Similar to how] we also would never be able to do a consistency check on a large knowledge base on our own, we use tools to do that.”
While current integration methods are still “crude”, Haase says there is great potential for teaching new capabilities, through steps like fine-tuning.
Oltramari says, “Neuro-symbolic methods are instrumental to engineer such hybrid AI-reliant information systems for human-machine collaboration. In our example, such methods enable the fusion between sensor-based data patterns that can be extracted using machine learning techniques with human knowledge, which can be elicited from experts and distilled in knowledge graphs.
Building these methods represent a real challenge, especially when explainability of the resulting system is considered: this requires a common level of knowledge representation and reasoning shared by humans and machines and, at this time, this convergence is far from being solved.”
While opinions may differ on the significance and weight of their roles, it’s clear that both LLMs and humans will play crucial parts in the convergence of hybrid AI and enterprise modeling for information systems. There is a shared understanding that connecting these two could potentially lead to significant capabilities.
Summary
The explainability and trustworthiness in the output of these systems are paramount, especially as enterprises scale and data needs evolve. Connecting hybrid AI with enterprise modeling in some form ensures that tasks and answers generated are rooted in a single source of truth—which, at the moment, is a human-operated process but in the future could also evolve into an automated LLM process. What is certain, is that hybrid AI and enterprise modeling hold great potential to impact the future of information systems.
We thank all the panelists for their participation and insightful contributions to this discussion!
The next HybridAIMs workshop will take place in Vienna, Austria from June 16-17, 2025. If you wish to submit a research paper or short paper for presentation, you can find the submission requirements here. The deadline to submit a paper is March 5, 2025.
Semantic knowledge modeling with metaphactory
metaphactory is an industry-leading enterprise knowledge graph platform transforming data into consumable, contextual and actionable knowledge through semantic knowledge modeling and knowledge discovery.
Our low-code, FAIR Data platform simplifies capturing and organizing domain expertise in explicit semantic models, extracting insights from your data and sharing knowledge with relevant stakeholders, as well as powering AI-driven solutions across the enterprise.
Speak with an expert to learn more about how metaphactory can support your organization’s information system needs!
