By now, the word is out that knowledge graphs increase the accuracy, reliability, and overall usefulness of enterprise AI. They provide the deterministic structure that keeps GenAI outputs and AI agent orchestration grounded in your actual business logic. But, as anyone who has built one knows, curating and maintaining taxonomies, ontologies, and the knowledge graphs they enable takes considerble effort.
The problem is that you can't buy a tailored enterprise ontology off the shelf. Capturing true knowledge requires mapping specific personas, semantic relationships, and unique business context. Traditionally, this means specialist taxonomists have to spend months manually researching and curating concepts, creating a labor-intensive bottleneck for enterprise knowledge management.
So, our innovation team asked, what if we could help organizations get a head start by automating the process? Read on to learn more about the steps involved or listen to Squirro CEO Dave Clarke talk about himself in the webinar recording shared here (feel free to skip ahead to minute 26).
Instead of throwing a generic prompt at a large language model (LLM) and hoping for a usable taxonomy, our innovation team modeled the exact steps of a human ontologist into a structured, deterministic graph using business process model and notation (BPMN).
They built an AI agentic framework that walks the user through the process, asks the right questions, captures human-in-the-loop feedback, and does the heavy lifting of graph creation.
Here is an under-the-hood look at how they engineered an agentic AI knowledge graph builder from scratch and what it hints at about the future of enterprise AI governance.
The Challenge: Agentic AI Knowledge Graph Creation
Raw data only becomes true enterprise knowledge when you explicitly map the relationships and business context connecting it. To make your enterprise AI effective, you have to give it a precise map of your specific business world – an industry-standard Resource Description Framework (RDF) ontology.
Traditionally, building this map requires specialists in ontology engineering to spend months researching and curating concepts. This manual effort can be a massive bottleneck for organizations trying to tailor their enterprise AI architecture to their unique enterprise knowledge.
We wanted to see if an AI agent could handle this heavy lifting. Prompting a standard LLM to "build a taxonomy" usually generates a basic textual outline rather than a machine-readable ontology model. Left to its own devices, the LLM hallucinates relationships and breaks the rigid structural rules required for a functioning ontology. To execute this agentic workflow reliably, we needed a deterministic framework to guide the agent's actions.
The Solution: BPMN as an Ontology Model
To provide this deterministic framework, we turned to a proven industry standard: the BPMN business process management notation ontology.
When you look at a process workflow diagram, it is essentially a graph – a series of nodes connected by specific relationships and triggers. We took the exact step-by-step processes required to build an ontology and modeled them as a graph using BPMN. We defined every process stage, escalation point, and validation node as an abstract schema.
By bringing that model into a graph and converting it into an RDF ontology model, the workflow became entirely transparent as a knowledge graph. Directly inside those graph nodes, we embedded specific AI prompt templates. The BPMN model acts as a strict workflow state machine. It dictates exactly which internal systems the agent can query, the specific data formats the outputs must take, and the precise conditions that trigger a human escalation. The graph controls the journey, while the large language model executes the micro-tasks at each stop.
The Execution: The Three-Tier Taxonomy Test
With the deterministic process model built, we added a conversational front end to the AI agent orchestration platform so a non-expert user could interact with it.
Next, to validate the approach, we asked the agent to execute an automated taxonomy generation workflow for a three-tier clothing business from scratch. Because the agent strictly followed the semantic graph, it walked the user through a guided, step-by-step conversation rather than spitting out a generic text response.
At each node, the agent used generative AI models to execute specific tasks:
- It generated preferred labels in English and Spanish, and drafted precise definitions for the concepts.
- It paused at mandatory human-in-the-loop trigger points to ask the user for guidance (e.g., instructing the agent to go deeper into specific areas, but explicitly exclude "baby clothing").
Probabilistic Intelligence, Deterministic Output
Within minutes, this conversational agent generated a fully functional, industry-standard RDF ontology.
This architecture establishes a continuous feedback loop. As the generative model categorizes new information, it feeds that structured data back into the semantic graph. The deterministic layer grows richer, providing highly accurate context for the next automated task.
Achieving this level of autonomous output highlights a fundamental architectural truth: Reliable autonomous AI agents sit at the exact intersection of generative AI and knowledge graphs. Designing an AI agent framework where GenAI sits inside a rigid, graph-based process model delivers capabilities that prompting alone cannot match.
Large language models are probabilistic natural language systems. When you use an enterprise knowledge graph to dictate the process flow, you constrain that probabilistic nature. You get the speed of generative AI combined with the precision of deterministic business logic.
If your organization is trying to move complex workflows out of the sandbox and into production, you'll need an enterprise AI architecture that supports actual enterprise AI governance.
Build an AI Foundation You Can Trust
Ready to build a foundation for accurate, deterministic generative ai? Download our free white paper, Closing the Gap in Generative AI Accuracy, to learn how to eliminate hallucinations, improve decision-making, and unlock the true ROI of your AI investments with an enterprise taxonomy, ontology, and knowledge graph.
