L
Lance Thieshen
Guest
In this post, we discuss the benefits of combining the strengths of neural (LLM) and symbolic (semantic) methods for grounded AI agents.
Ontology used to be a term mostly associated with semantic web researchers and taxonomists. Over time, it became a niche discipline within enterprise data architecture. It was useful but often treated as optional. Ontologies are becoming more central to how organizations define meaning, connect data and support AI. They are quickly becoming a core component of the intelligent enterprise.
Ontology management and knowledge graph design skills are back in demand, largely driven by AI. LLMs are powerful, but they struggle to consistently understand domain-specific concepts and stay aligned to business rules. AI agents amplify this challenge because they act on that understanding. Ontologies help address these gaps by providing structure, shared meaning, context and constraints, so systems stay aligned and decisions remain sound.
As AI agents take on more responsibility for planning and executing processes, the focus in enterprise AI is shifting toward governance and risk. The real question is whether these agents can be trusted to operate in real business environments, where they interact with enterprise data, invoke tools and support decisions that must be explainable, repeatable and defensible.
Even the best language models can behave like black boxes. They may produce correct answers one moment and hallucinate or break rules the next, without any clear reasoning path. AI agents extend this problem. They are capable, but they lack a consistent understanding of business context and meaning. This leads to unpredictable behavior and mistakes. They can misinterpret goals or bypass policy constraints, which undermines reliability, makes explanation difficult and creates governance challenges.
Without deeper semantic grounding in enterprise knowledge, an AI agent is effectively guessing. It cannot reason about business rules with confidence. It needs a map.
Context has been the missing link in RAG implementations and continues to be the antidote to inconsistent results and opaque logic for AI agents. For organizations operating in highly regulated industries, where accuracy and accountability are non‑negotiable, ontologies, semantic models and transparent reasoning are essential to making AI systems work in practice.
What do we mean by an ontology-driven AI? There is a strong synergy between agents and ontologies in modern AI systems. In a layered architecture, LLMs act as reasoning engines, while agents orchestrate tasks and workflows, using protocols like MCP as the connective layer between AI and enterprise systems. Within this framework, ontologies provide the structure that ensures consistency and shared meaning across interactions.
Advanced systems use LLMs to interpret and orchestrate over knowledge bases, then rely on ontologies to structure outputs and maintain coherence. The result is more explainable AI, where reasoning can be traced through formal rules and graph-based relationships.
Ontologies provide an operational map of the enterprise and a shared understanding for agents operating across fragmented systems. Instead of forcing the agent to reconcile inconsistent data models or translate between vocabularies, the ontology defines what key concepts mean and how they relate. For example, it can define what "equipment" is, how it relates to production, maintenance and personnel and what operations it can undergo.
With a well-built ontology, the agent operates against a durable semantic model rather than trying to infer meaning on the fly. It can consistently understand how concepts relate, such as equipment linked to maintenance, certifications, locations, and personnel. By navigating a semantic graph, the agent has a structured layer of context it can reason over.
Here is how robust, domain-specific AI agents benefit from ontologies and knowledge graphs:
In traditional application development, each use case typically results in its own isolated data model, one for maintenance, another for quality analysis, and another for production scheduling. These models are built independently, and insights or structures rarely carry over between them.
An ontology changes that dynamic. The semantic model is built once and becomes a shared layer of meaning that all agents and use cases can rely on. For example, a scheduling agent can access equipment capabilities, operator qualifications, material constraints and maintenance schedules without requiring months of integration work.
This creates a consistent view of the business across systems, without forcing everything into a single physical source. As more use cases are added, the value of the model compounds. Each new use reinforces the ontology and extends its impact across the enterprise.
Agents improve only when their actions can be tied to real outcomes. If a system recommends an action, classifies a case, or routes a request, the organization needs to know whether that decision was useful, accurate and aligned with business intent.
Ontologies enable this by linking actions to results in a structured way. A recommendation can be tied to the event it influenced. A classification can be evaluated against downstream resolution. The model itself can be refined as new terms and patterns emerge in real content. This creates a traceable path from decision to outcome.
With this structure in place, agents are not just executing tasks. They are learning from the operational feedback they generate and building an evolving base of insight that informs better decisions over time.
In enterprise AI, governance comes down to whether the system is using the right concepts, respecting the right distinctions and producing outputs that can be reviewed and defended.
In legal, healthcare, compliance, and similar domains, it is not enough for an answer to sound reasonable. It must be consistent, reviewable and justifiable. This is where ontology-driven classification and extraction outperform unconstrained generation.
A semantic model supports governance in practical ways. It lets teams define the concepts that matter, express valid relationships, encode structural expectations, and validate that the model remains coherent. It creates a layer where business meaning, policy expectations and system behavior stay aligned.
This is why the role of the ontologist becomes more important in the age of AI. The ontologist is not replaced by generative AI. They act as the human-in-the-loop, ensuring semantic guardrails are correct, useful and aligned to the business.
Building and maintaining ontologies can be labor-intensive, and many organizations stall due to skill gaps and time to value. AI assistance can help reduce that effort if it is applied carefully.
A common question is whether to rely on LLMs or an enterprise classification platform. The answer depends on what you are trying to solve.
When it comes to classification and extraction, the tradeoffs are clear. LLMs can work well for smaller volumes or less critical use cases. But in domains like healthcare, legal, and compliance, where accuracy and defendability are required, an enterprise classification platform is the better choice. From my extensive experience in legal, defendability is not optional. Every outcome must be consistent and justifiable.
Cost is also a factor at scale. Using LLMs for classification introduces token-based costs that grow with both data volume and usage. As content scales or models change and require rescanning, those costs increase again. Organizations that need predictable and stable cost models often look for a more consistent approach.
Ontology generation is a different problem. Here, LLMs can add real value by accelerating model development, suggesting concepts, and helping expand coverage. When combined with expert review and governance, this can significantly reduce the effort required to build and maintain a high-quality semantic model.
To get the benefits of both, organizations should use LLMs to accelerate ontology development and an enterprise classification platform to deliver accuracy, scalability and governance in production.
Robust ontologies require teams to define concepts, boundaries, relationships, constraints, synonyms, reuse patterns and governance rules, while avoiding the temptation to over-model.
This is where Progress Semaphore helps. The Semaphore AI Model Builder assists with some of the more time-consuming parts of ontology development by suggesting narrower terms, alternative labels, and potential model structures. These suggestions are guided by user review and existing model context.
It does not replace the role of the ontologist or domain expert, but it can reduce the manual effort required to build and extend a model while keeping decisions in the hands of the team.
Semaphore also addresses a common challenge in semantic work, which is managing complexity over time. As ontologies grow, teams need structure, validation, and a controlled environment to refine and maintain the model. Semaphore provides that environment, supporting taxonomy management, ontology structures, constraints, validation, classification, extraction and reuse.
The AI Model Builder acts as a productivity aid for semantic engineering. It helps teams move faster, especially in early model development and expansion, while still relying on governance and expert oversight to ensure quality.
Looking ahead, this area is evolving further with the introduction of MCP server capabilities. This extends beyond the current AI Model Builder by enabling more direct interaction with the semantic model through natural language. Users can create and refine model structure, including custom classes, alternative label types, and relationships, as well as generate concepts and apply those relationships through prompts. With a bring your own LLM approach, organizations can control how models are developed and integrated, while keeping governance and oversight in place.
Gaps in organizational knowledge, lack of governed retrieval and a hard trust ceiling are exactly why ontologies are back.
Ontologies fill that gap by grounding AI in real business context. They capture how the business actually understands its world, provide a shared layer of meaning across systems, and make outputs easier to trust, explain and govern. This is what allows AI to move from interesting to operational in enterprise environments.
At the same time, modern platforms like Semaphore are making this more practical. AI can assist with model development and expansion, but the outcome remains governed, structured and aligned to the business.
For organizations investing in LLMs and AI agents, this is a critical shift. Success is no longer just about model capability. It is about how well those systems are grounded in the enterprise, how consistently they behave and how confidently their outputs can be used in real decisions.
This is the right time to build or refresh your domain ontologies so they can support reliable, explainable decision-making across your AI and agent pipelines.
Watch our team discuss this topic.
Continue reading...
Ontology used to be a term mostly associated with semantic web researchers and taxonomists. Over time, it became a niche discipline within enterprise data architecture. It was useful but often treated as optional. Ontologies are becoming more central to how organizations define meaning, connect data and support AI. They are quickly becoming a core component of the intelligent enterprise.
Ontology management and knowledge graph design skills are back in demand, largely driven by AI. LLMs are powerful, but they struggle to consistently understand domain-specific concepts and stay aligned to business rules. AI agents amplify this challenge because they act on that understanding. Ontologies help address these gaps by providing structure, shared meaning, context and constraints, so systems stay aligned and decisions remain sound.
Can You Really Trust Your AI Agent?
As AI agents take on more responsibility for planning and executing processes, the focus in enterprise AI is shifting toward governance and risk. The real question is whether these agents can be trusted to operate in real business environments, where they interact with enterprise data, invoke tools and support decisions that must be explainable, repeatable and defensible.
Even the best language models can behave like black boxes. They may produce correct answers one moment and hallucinate or break rules the next, without any clear reasoning path. AI agents extend this problem. They are capable, but they lack a consistent understanding of business context and meaning. This leads to unpredictable behavior and mistakes. They can misinterpret goals or bypass policy constraints, which undermines reliability, makes explanation difficult and creates governance challenges.
Without deeper semantic grounding in enterprise knowledge, an AI agent is effectively guessing. It cannot reason about business rules with confidence. It needs a map.
Context has been the missing link in RAG implementations and continues to be the antidote to inconsistent results and opaque logic for AI agents. For organizations operating in highly regulated industries, where accuracy and accountability are non‑negotiable, ontologies, semantic models and transparent reasoning are essential to making AI systems work in practice.
Ontology-Driven AI Agents
What do we mean by an ontology-driven AI? There is a strong synergy between agents and ontologies in modern AI systems. In a layered architecture, LLMs act as reasoning engines, while agents orchestrate tasks and workflows, using protocols like MCP as the connective layer between AI and enterprise systems. Within this framework, ontologies provide the structure that ensures consistency and shared meaning across interactions.
Advanced systems use LLMs to interpret and orchestrate over knowledge bases, then rely on ontologies to structure outputs and maintain coherence. The result is more explainable AI, where reasoning can be traced through formal rules and graph-based relationships.
Ontologies provide an operational map of the enterprise and a shared understanding for agents operating across fragmented systems. Instead of forcing the agent to reconcile inconsistent data models or translate between vocabularies, the ontology defines what key concepts mean and how they relate. For example, it can define what "equipment" is, how it relates to production, maintenance and personnel and what operations it can undergo.
With a well-built ontology, the agent operates against a durable semantic model rather than trying to infer meaning on the fly. It can consistently understand how concepts relate, such as equipment linked to maintenance, certifications, locations, and personnel. By navigating a semantic graph, the agent has a structured layer of context it can reason over.
Here is how robust, domain-specific AI agents benefit from ontologies and knowledge graphs:
A Unified Semantic Model
In traditional application development, each use case typically results in its own isolated data model, one for maintenance, another for quality analysis, and another for production scheduling. These models are built independently, and insights or structures rarely carry over between them.
An ontology changes that dynamic. The semantic model is built once and becomes a shared layer of meaning that all agents and use cases can rely on. For example, a scheduling agent can access equipment capabilities, operator qualifications, material constraints and maintenance schedules without requiring months of integration work.
This creates a consistent view of the business across systems, without forcing everything into a single physical source. As more use cases are added, the value of the model compounds. Each new use reinforces the ontology and extends its impact across the enterprise.
Progressive Learning
Agents improve only when their actions can be tied to real outcomes. If a system recommends an action, classifies a case, or routes a request, the organization needs to know whether that decision was useful, accurate and aligned with business intent.
Ontologies enable this by linking actions to results in a structured way. A recommendation can be tied to the event it influenced. A classification can be evaluated against downstream resolution. The model itself can be refined as new terms and patterns emerge in real content. This creates a traceable path from decision to outcome.
With this structure in place, agents are not just executing tasks. They are learning from the operational feedback they generate and building an evolving base of insight that informs better decisions over time.
Embedded Governance Model
In enterprise AI, governance comes down to whether the system is using the right concepts, respecting the right distinctions and producing outputs that can be reviewed and defended.
In legal, healthcare, compliance, and similar domains, it is not enough for an answer to sound reasonable. It must be consistent, reviewable and justifiable. This is where ontology-driven classification and extraction outperform unconstrained generation.
A semantic model supports governance in practical ways. It lets teams define the concepts that matter, express valid relationships, encode structural expectations, and validate that the model remains coherent. It creates a layer where business meaning, policy expectations and system behavior stay aligned.
This is why the role of the ontologist becomes more important in the age of AI. The ontologist is not replaced by generative AI. They act as the human-in-the-loop, ensuring semantic guardrails are correct, useful and aligned to the business.
Should You Use Agents for Ontology Generation?
Building and maintaining ontologies can be labor-intensive, and many organizations stall due to skill gaps and time to value. AI assistance can help reduce that effort if it is applied carefully.
A common question is whether to rely on LLMs or an enterprise classification platform. The answer depends on what you are trying to solve.
When it comes to classification and extraction, the tradeoffs are clear. LLMs can work well for smaller volumes or less critical use cases. But in domains like healthcare, legal, and compliance, where accuracy and defendability are required, an enterprise classification platform is the better choice. From my extensive experience in legal, defendability is not optional. Every outcome must be consistent and justifiable.
Cost is also a factor at scale. Using LLMs for classification introduces token-based costs that grow with both data volume and usage. As content scales or models change and require rescanning, those costs increase again. Organizations that need predictable and stable cost models often look for a more consistent approach.
Ontology generation is a different problem. Here, LLMs can add real value by accelerating model development, suggesting concepts, and helping expand coverage. When combined with expert review and governance, this can significantly reduce the effort required to build and maintain a high-quality semantic model.
To get the benefits of both, organizations should use LLMs to accelerate ontology development and an enterprise classification platform to deliver accuracy, scalability and governance in production.
Accelerating Robust Ontology Generation with Semaphore
Robust ontologies require teams to define concepts, boundaries, relationships, constraints, synonyms, reuse patterns and governance rules, while avoiding the temptation to over-model.
This is where Progress Semaphore helps. The Semaphore AI Model Builder assists with some of the more time-consuming parts of ontology development by suggesting narrower terms, alternative labels, and potential model structures. These suggestions are guided by user review and existing model context.
It does not replace the role of the ontologist or domain expert, but it can reduce the manual effort required to build and extend a model while keeping decisions in the hands of the team.
Semaphore also addresses a common challenge in semantic work, which is managing complexity over time. As ontologies grow, teams need structure, validation, and a controlled environment to refine and maintain the model. Semaphore provides that environment, supporting taxonomy management, ontology structures, constraints, validation, classification, extraction and reuse.
The AI Model Builder acts as a productivity aid for semantic engineering. It helps teams move faster, especially in early model development and expansion, while still relying on governance and expert oversight to ensure quality.
Looking ahead, this area is evolving further with the introduction of MCP server capabilities. This extends beyond the current AI Model Builder by enabling more direct interaction with the semantic model through natural language. Users can create and refine model structure, including custom classes, alternative label types, and relationships, as well as generate concepts and apply those relationships through prompts. With a bring your own LLM approach, organizations can control how models are developed and integrated, while keeping governance and oversight in place.
Conclusion
Gaps in organizational knowledge, lack of governed retrieval and a hard trust ceiling are exactly why ontologies are back.
Ontologies fill that gap by grounding AI in real business context. They capture how the business actually understands its world, provide a shared layer of meaning across systems, and make outputs easier to trust, explain and govern. This is what allows AI to move from interesting to operational in enterprise environments.
At the same time, modern platforms like Semaphore are making this more practical. AI can assist with model development and expansion, but the outcome remains governed, structured and aligned to the business.
For organizations investing in LLMs and AI agents, this is a critical shift. Success is no longer just about model capability. It is about how well those systems are grounded in the enterprise, how consistently they behave and how confidently their outputs can be used in real decisions.
This is the right time to build or refresh your domain ontologies so they can support reliable, explainable decision-making across your AI and agent pipelines.
Watch our team discuss this topic.
Continue reading...