In recent years, deep neural networks have dominated the artificial intelligence landscape thanks to their ability to learn complex patterns from large volumes of data. However, their internal workings often remain opaque, giving rise to the “black box” problem: it is difficult to understand why a network makes certain decisions. To address this need, XAI (eXplainable AI) systems have emerged, enabling the analysis and partial explanation of neural network behavior.
Another strategy to improve explainability involves designing smaller networks that are easier to interpret, or combining neural networks with symbolic systems in hybrid pipelines. In this way, a complex task is broken down into more manageable sub-problems, handled by lighter models with a higher degree of interpretability. This approach is crucial in sectors such as defense or pharmaceuticals, where reliability and decision traceability are non-negotiable.
The Role of LLMs and Their Limitations
The advent of LLMs has radically changed expectations around AI. Their ability to tackle complex and diverse tasks has opened the prospect of delegating entire workflows to LLMs, including analytical tasks traditionally performed by human teams.
However, concrete limitations soon became apparent. Generative models are expensive to run, consume significant resources, and present well-known issues of “hallucinations”—incorrect or fabricated responses—as well as risks related to sensitive data leakage or the exposure of proprietary know-how. Attempting to entrust a generative LLM with a structured analytical task or the management of large volumes of internal documents has proven inefficient and often unreliable.
Specialized Agents and RPA 2.0
To overcome these limitations, organizations are increasingly turning to small hybrid agents combinations of code and LLMs designed to perform a single task in a repetitive and optimized manner. This strategy is, in effect, a modern reinterpretation of RPA (Robotic Process Automation), now made more accessible and intelligent through integration with language models.
Agents can automate specific workflows, such as extracting information from structured databases, without delegating the entire process to an LLM. This reduces costs, complexity, and the risk of errors, while still leveraging the semantic interpretation capabilities of generative models.
Interpretable AI Systems for Industry
Another benefit introduced by LLMs is the possibility of envisioning a connection between structured data such as that stored in enterprise databases (ERP, MES, CMMS, CRM, etc.) and the free text contained in technical documents or PLM systems. While well-established practices exist for extracting, merging, and storing data from databases, technical text represents a far more complex domain: it contains engineering jargon, formulas, parameters, tables, and legal references.
To handle this type of data, traditional language models are not sufficient. What is needed is neurosymbolic AI – AI built on rules, logic, and explicit knowledge representations to provide genuine engineering understanding and enable systems to reason like an engineer. This is where ontologies become fundamental: as discussed in previous articles, ontologies formalize classes of concepts (entities) and the relationships between them, allowing knowledge to be organized in a structured way that is independent of the form in which it is described.
Discover our AI software for the automatic extraction of information from technical documents.
Intelligent Agents Based on Domain Ontologies
Thanks to ontologies, an agent can query different databases and understand, for example, that the class [user] represents the same concept in a CRM, an ERP, or a MES, even though it may have different properties in each system. The agent therefore no longer searches as it once did by “keyword” (few results), through semantic search (a lot of noise), or via an LLM (costly and imprecise) but instead accesses a labeled meta-text containing [entity] tags and [relationship] tags.
This meta-text is the result of processing performed by a synthesis engine essentially another agent (imagine a small, specialized LLM) optimized to search for, capture, extract, and catalog entities and relationships. The ontological engine “chews through” the documentation only once, extracts [entities] and [relationships], encodes texts with explicit references to the identified entities and relationships, and then releases an index that all other agents can access. Once the engine has identified, labeled, and indexed all [users], agents will query the released index to perform their respective tasks.
Each new document written by engineering, marketing, or legal teams is therefore processed and added to the corporate knowledge base, keeping information continuously up to date and immediately accessible for different workflows. This approach combines automation capabilities, semantic intelligence, and centralized knowledge management, ensuring efficiency, accuracy, and scalability.
Symbolic Artificial Intelligence
The renewed interest in ontological models or neurosymbolic systems is not a passing trend, but a concrete response to a real need: overcoming the limitations of generative models and deep neural networks. By combining specialized agents, ontologies, and enterprise databases, companies can create intelligent, explainable, and reliable workflows. Ontologies, in particular, make it possible to formalize concepts and relationships in a universal way, enabling coherent and structured access to heterogeneous information.
In an era of powerful but costly LLMs sometimes unreliable for complex analytical tasks integrating ontologies with LLMs represents the key to more effective and secure enterprise knowledge management.
Request a demo of our AI tools.
