Organizations are facing a critical challenge to AI adoption: how to leverage their domain-specific knowledge to use AI in a way that delivers trustworthy results. Knowledge graphs provide the missing “truth layer” for AI that transforms probabilistic outputs into real world business acceleration.
• 🚀 AI adoption is accelerating, but most implementations fail to deliver expected business value
• 🔍 Knowledge graphs provide the essential “truth layer” for reliable AI systems
• 🔄 Pragmatic AI combines LLMs’ creative potential with knowledge graphs’ verification capabilities
Introduction
“Context is what gives meaning to pretty much everything. So to that extent, all graphs have the inherent potential to bring more knowledge or meaning because they’ve already taken the first step of acknowledging the interconnectivity and the contextual nature of information”.
We live in a world dominated by AI and large language models (LLMs), and it turns out that context and meaning are essential to getting quality results out of those. Knowledge graphs may hold the key to providing context and meaning to unlock AI’s potential, and the evidence in support of this is mounting.
On the occasion of the release of the latest Gartner Hype Cycle for Artificial Intelligence, Research VP, AI at Gartner Svetlana Sicular noted that investment in AI reached a new high, with a focus on generative AI. Yet, in most cases, this has yet to deliver the anticipated business value.
Knowledge Graphs are at the heart of Critical Enabler technologies in Gartner’s list of emerging technologies for leaders to consider as part of their strategy. Gartner recommends knowledge graphs as being critical in building and advancing GenAI models. Organizations like Amazon and Samsung are using knowledge graphs, and the market is expected to grow to $6.93 Billion by 2030, at a CAGR of 36.6%.
Gartner has been advocating for the role of knowledge graphs in AI today and the downstream effects in organizations going forward for the last few years, as neither the technology nor the vision are new. Knowledge graph technology has been around for decades, and people like Tony Seale were early to identify its potential as a truth layer for AI.
Seale, also known as “The Knowledge Graph Guy”, is the founder of the eponymous consulting firm. The above quote is taken from an extensive conversation, covering everything from knowledge graph first principles to application patterns for safe, verifiable AI, real-world experience, trends, predictions, and the way forward.
🧠 Knowledge Graphs and AI Context
• Despite significant AI investment, most organizations have yet to deliver anticipated business value
• Context and meaning are essential for quality AI results
• Knowledge graphs provide the crucial context layer for AI systems
• Knowledge graphs are critical enablers for competitive GenAI strategy
From data silos to Linked Data and Knowledge Graphs
Seale has decades of experience working with data in Tier 1 financial institutions. About ten years ago, he was working on “yet another ETL project” for a large investment bank, bringing data into a data warehouse and implementing data pipelines. This is a typical approach to serve organizational reporting and compliance needs. The problem is, it doesn’t scale, or help add context and meaning.
Then Seale came across Tim Berners Lee’s 2010 TED talk on Linked Data, and that changed everything. In 2010, Google was just getting into Knowledge Graphs, and the term had not really surfaced. But the technology was there, under the Linked Data moniker. TBL’s TED talk was enough to get Seale to grasp the 2 key principles of Linked Data, and to start experimenting with this as an alternative to ETL.
The basic idea of Linked Data is to apply the general architecture of the World Wide Web to the task of sharing structured data at global scale. What it all comes down to is using HTTP identifiers for data, so that they can be looked up, and providing information about their meaning (semantics) using standards.
What Seale understood was that if the decentralized nature of this approach could work for the web, it could work for any organization. Rather than having one central point of integration and control, which is the de facto approach of ETL projects and data warehouses, knowledge graphs enable scale through decentralization and standards.
These are the same principles that make the web work. It’s no wonder that the inventor of the web wanted to take it to the next level to go from a web of documents to a web of data. Beyond accessing data, however, this approach adds semantics to the mix. Data points as well as links among them can have specific meaning and types attached to them.
The best example of semantics in action at web scale is schema.org. Schema.org is a collaborative effort to define a standard vocabulary that’s used by 30% of all websites, and 72.6% of pages on the first page of Google. Beyond defining semantics using standards, schema.org makes annotation and integration scalable via decentralization.
🌐 Linked Data Foundations
• Linked Data applies web architecture principles to structured data sharing
• Uses HTTP identifiers so data can be looked up systematically
• Provides information about meaning (semantics) using standards
• Enables organizational scale through decentralization
• Schema.org exemplifies semantic standards at web scale
AI meets Knowledge Graphs
Schema.org is what enables the Googles of the world to build their knowledge graphs and make more sense of the web. It’s the same approach that Seale first started toying with as an under the desk project at the investment bank he was working for at the time, expecting it to fail. It didn’t.
Encouraged by initial success, Seale became a passionate knowledge graph advocate and initiated a number of related projects. He moved organizations in pursuit of his passion, and was looking into graph neural networks as a way to bootstrap the semantics and annotation needed to build knowledge graphs when the first GPT large language models were released.
Seale started experimenting with LLMs, and soon became convinced of two things. First, that LLMs are going to have a massive impact. Second, that LLMs are a perfect match for knowledge graphs. He started sharing his ideas on LinkedIn, and getting viral. Eventually, he formed his own consultancy and is now working on implementing these with a number of clients.
“All organizations are going to have to accept the reality that we’re moving to a more probabilistic world. So everyone has got to start using AI or you’re probably going to go out of business. We’re moving to this new world where things will be probabilistic and AI will be embedded in a lot of the decision making.
You might not like it or you might have whatever opinion, but it doesn’t care. It’s some force of nature that’s happening, so you might as well just get used to it. So then the question really becomes, well, how do you do that in a safe way. And in my opinion, that comes through external verification”, Seale said.
That’s the core of the approach he’s advocating for. It includes patterns with fancy names, such as Working Memory Graph and Neural-Symbolic Loop, and examples ranging from DeepSeek to the Cyc project. But before diving into these, it’s worth pausing for a moment to ground ourselves in first principles.
🤖 AI and Knowledge Graph Integration
• LLMs and knowledge graphs are complementary
• We’re moving toward a probabilistic world where AI will be embedded in decision-making
• External verification through knowledge graphs creates safer AI
• Organizations must adapt to this shift
First principles: Graphs and Knowledge Graphs
So what makes graphs different from other data structures, and what makes knowledge graphs different from other graphs? We can approach this at the implementation level or at the first principles level.
Regardless, whether we’re talking about a spreadsheet vs. a mind map, a relational database’s rows and columns vs. a graph database’s nodes and edges, or set theory vs. graph theory, there’s one thing that sets graph apart: connections as first-class citizens. But not all graphs qualify as knowledge graphs.
Both nodes and edges in a graph can be of different types. A simple graph can include nodes representing products, and edges representing a generic type of relationship between them. A bipartite graph can have two different types of nodes, representing products and customers, and edges representing which customer bought what product.
A heterogeneous graph can have all sorts of different types of nodes and edges. For example, nodes representing products and customers, and edges representing which customer bought what product and what product was reviewed by which customer.
There is utility in graphs even at the simplest possible level. Graph algorithms like path finding and centrality can be extremely useful for applications and analytics, and don’t require heterogeneous graphs.
“Once you start saying, well, actually, no, some of these nodes are different things, and the edges between them, they’re special different types of edges that mean something, then the complexity goes up. The nature of the algorithms you can run, including machine learning algorithms, changes. I think we could call that the entry level to what a knowledge graph is”, Seale noted.
Note the “entry level” part. There’s a long and complicated history here, going back to the early 00’s and the Semantic Web. It was on these ideas, standards and technical stack that the Linked Data principles built on. “Semantic Web” died out while “Knowledge Graph” caught on.
The Semantic Web was arguably ahead of its time. Many implementation efforts were misguided, and its proponents have not always been pragmatic. However, as Seale noted, neural networks were also considered a flop for a long time. Using URIs as identifiers and having a shared vocabulary and an agreed schema remain the hallmarks of knowledge graphs and the value they can bring.
📊 Graph Fundamentals
• Graphs differ from other data structures by treating connections as first-class citizens
• Not all graphs qualify as knowledge graphs
• Knowledge graphs add semantic meaning to nodes and edges
• URIs as identifiers and shared vocabularies are defining features of knowledge graphs
The continuous world and the discrete world
The structure and semantics that knowledge graphs bring enable things that are simply not possible with other data types or even other graphs. Seale believes every organization should be working on their own version of schema.org and using it to annotate their data, building knowledge graphs to power their AI.
Seale used DeepSeek as an example to explain the verifier approach. Like everyone else, Seale was obsessing with DeepSeek and trying to work out what it is that they did. Clever algorithms and optimizations aside, at the core of DeepSeek’s success is the fact that they used verifiable data for reinforcement learning: math and code.
“They took all the web data, like everybody is doing. But then they pulled out just the bits related to mathematics and coding. With that, you can create an external verifier.
You can look at the math or the code, then you can look at the answer at the end, and you can check whether the answer is actually correct. Then you can feed that to the LLM and ask the LLM to do that, and then check against the external formal verifier. What that is doing is, it’s adding quality control upon the probabilistic model”, Seale explained.
Seale then elaborated on what he calls the continuous world and the discrete world. In the continuous world, everything’s probabilistic, everything’s fuzzy, and that’s where these generative AI models are. One thing blends into another, and you get hallucinations. But the flip side of that, as per Seale, is that there’s something a bit like creativity there.
In the old-fashioned AI world, there’s the legend of the Cyc Project. Cyc is a hugely ambitious AI project, aiming to encode general knowledge about the world in a formal way. Seale has great respect for Cyc. However, he noted, Cyc did not and could not succeed, whereas generative AI models do – in their own way. But they come with their own set of drawbacks.
Generative AI models cannot be trusted, and that makes them inappropriate for enterprise adoption in domains such as finance, law, or medicine. For domains such as math or code, it’s possible verify results formally. What if there was a way to do that in other domains too? Seale thinks there is, and the keys are knowledge graphs and ontology.
🌓 The Continuous vs. Discrete World
• Continuous world: probabilistic, fuzzy, creative but prone to hallucinations (LLMs)
• Discrete world: logical, formal, verifiable but limited (traditional AI)
• For math or code, results can be formally verified
• Knowledge graphs and ontologies can provide verification for other domains
The rise of Ontology
We’ve mentioned schema.org, as well as the general notion of schema. Schemas are typically associated with relational databases, where they define the structure and organization of data. Graphs can have schemas, too. Schemas for knowledge graphs are called ontologies, although the word “schema” does not really do justice to ontologies.
Ontologies go beyond schemas by enabling modeling of constructs such as inheritance hierarchies or logical axioms. They can capture not just the structure and organization of data, but also things such as business rules and domain knowledge.
“The name of the game is to get as close to the semantics of the business as you possibly can. You’re trying to take the words that the business people are using within a given organization, and turn those into these formal concepts to get really specific about what they are, and then to connect concepts together in a way that they interrelate to each other with specific types of edges”, Seale explained.
Building an ontology is not easy. It requires access to domain knowledge, which is typically scattered, partly documented and understood, and disputed among experts. It also requires ontological modeling expertise and the right tools.
This is why even though ontological modeling has been around for decades, it never really gained mainstream adoption. Seale thinks that’s changing though, and there may be some circumstantial evidence in support of this.
On Google Trends, “Knowledge Graph” has seen a 3.450% increase in the last 5 years. In the Year of the Graph’s dedicated repository, references to ontology since 2022 have more than doubled both in terms of quantity and variety of sources. Plus, Seale’s own success story is becoming the poster child for knowledge graph virality.
📚 Ontology’s Growing Importance
• Ontologies go beyond schemas by creating formal concepts from business terminology
• Goal: Accurately capture business semantics and relationships
• Requires domain knowledge and ontological modeling expertise
• Google Trends shows “Knowledge Graph” has grown 3,450% in 5 years
The Neural-Symbolic Loop
There is a two-way relationship knowledge graphs and ontologies can have with LLMs. LLMs can aid ontology development and knowledge graph population. Seale reported having a good experience using LLMs for this, but your mileage here may vary. In any case, such tools are meant to support the experts, not fully automate the task.
Where it gets really interesting though is the other way round: ontologies and knowledge graphs acting as a verifier, essentially a truth layer, for LLMs. Seale calls this the Working Memory Graph pattern for LLMs.
In the Working Memory Graph, the ontology distills domain knowledge and the knowledge graph serves as a database specific – and private – to the organization. LLMs act as mediators and add the exploration and creativity part, also granting access to unstructured knowledge. This can be general knowledge distilled in the LLM, or domain-specific knowledge via RAG.
The Working Memory Graph is part of a bigger pattern that Seale calls the Neural-Symbolic Loop. In this, the Working Memory Graph acts as the verifier for domains where verification is needed. The idea is to make possible for every domain what is possible for math or code: to verify the correctness of results generated by LLMs.
Clearly, that’s much harder to achieve in domains beyond math or code. The effort and the expertise required to build ontologies and knowledge graphs remain considerable, and results may not be as clear-cut. But it can be done, and Seale is convinced that’s the best way to a truth layer for AI.
🔄 The Neural Symbolic Loop Approach
• Start with domain knowledge specific to your organization
• Develop an ontology to formalize this knowledge
• Build a knowledge graph as your organization’s private database
• Use LLMs as mediators for exploration and creativity
• Apply the Working Memory Graph as a verification layer
• Create a continuous loop of verification and improvement
The Pragmatic AI Approach: From Theory to Implementation
“AI is off like a rocket. There’s pretty much nothing anybody can do to stop that. That’s happening anyway. So in any organization, you’re going to be in a situation where you’re going to be able to import this general intelligence. It’s smart at the moment, maybe not super smart, but it’s getting there in the next 5 to 10 years.
You’ve got this short window. What you need to do is take AI to the context of our organization, and concentrate on the bottom of the AI iceberg, which is the data. So you need to take the power that you’ve got in the models that you have in your hands right now, and focus that back upon the data that you’ve got in there.
You need to clean up and consolidate the data so that it’s in a state to be an effective external verifier. You need to be aware of what information is worth $0,001, and what information only you have, and what is the value that you’re adding. You need to do that now, because that’s the only game in town as far as I see”, Seale said.
Seale also shared a number for predictions for 2025: the data crunch, knowledge graphs as a foundation for data fabric, GraphRAG via ontologies, and the approximation of formal reasoning by reasoning LLMs. For an in-depth conversation on these, check out the podcast episode. Overall, Seale thinks AI is massively overhyped in the short term, yet massively underhyped in the long term.
Seale is working on applying knowledge graphs and ontologies to organizations that are ready for it. But that comes at a cost, and it can’t scale to everyone. Plus, no consultancy is ever going to be able to do all the education or the foundational data work that’s needed for you.
The Pragmatic AI approach bridges this gap by educating on data first principles, management, governance, modeling and data science. Then, domain knowledge unique to organizations can be leveraged to build AI systems on a foundation of trustworthy, verified data.
Pragmatic AI Training
The Pragmatic AI course provides executives, managers, entrepreneurs, consultants, and creatives with the foundational knowledge and practical expertise needed to build AI systems that deliver real business value. Start with the basics, and get a head start on creating your organization’s truth layer and gain a competitive advantage in the AI era.