--- title: "Knowledge Graphs" date: 2026-04-01 description: Knowledge graphs make entities, relationships, and provenance explicit. Why they can make enterprise AI more consistent, traceable, and easier to govern. author: Mario Thomas canonical: https://mariothomas.com/signals/knowledge-graphs/ --- ## What are Knowledge Graphs? {#what-it-is data-toc="What they are"} A large language model with access to accurate but flat data can summarise a document competently. It can also infer relationships among documents, but without explicit structure it must reconstruct many of them at query time. That may be adequate for some questions; where relationships are dispersed across a large collection, it can make answers less consistent, less traceable, and harder to govern. [Microsoft's GraphRAG evaluations](https://www.microsoft.com/en-us/research/blog/graphrag-unlocking-llm-discovery-on-narrative-private-data/) found substantial improvements over a conventional retrieval baseline on global sensemaking questions, where the answer depended on themes and relationships spread across a corpus. An enterprise knowledge graph goes further than a graph derived from documents: it governs the entities, definitions, relationships, and provenance on which those connections depend. The surrounding infrastructure is maturing too. ISO published GQL in 2024, standardising a language for modelling, querying, and modifying property graphs. ## Why it matters to Boards {#board-implications data-toc="Board implications"} Boards have spent a decade governing data quality, from completeness and accuracy to timeliness and lineage, and that work remains necessary. Quality and structure answer different questions: quality tells the organisation whether information is reliable; semantic structure makes its chosen meanings and relationships explicit. Some AI uses need both, particularly where systems must follow relationships across multiple sources rather than retrieve and summarise individual records. Lack of explicit structure can become a constraint, but it is not the only reason AI programmes stall. Proprietary definitions and relationships can nevertheless be difficult for a competitor to reproduce because they reflect how the organisation actually operates. Structure does not determine what every AI system must treat as true, but it gives the organisation an inspectable account of the definitions and connections on which its systems rely. The regulatory pull has stopped being voluntary. Since February 2026, the [Data (Use and Access) Act 2025](https://www.legislation.gov.uk/ukpga/2025/18/contents) has required safeguards for significant decisions taken solely by automated processing, including the right to contest a decision and obtain human intervention. Consider what answering a contest requires. A governed graph, combined with contemporaneous decision logs and versioned data, can make the evidence path more traceable: the authority under which the decision fell, the definitions and relationships used, the inputs available at the time, and the sources from which they came. That does not expose the internal reasoning of a model, and it does not by itself make the system explainable. It is one part of explainability, alongside model-specific techniques, decision records, and substantive human review. The Board question is whether the organisation can reconstruct a contested decision well enough to show what information and rules shaped it. ## Questions Boards are asking {#questions data-toc="Questions"} ### Is this a technology purchase? Not primarily. The hard work is agreement rather than software: deciding what the organisation means by customer, contract, supplier, or incident when different functions currently use different definitions. The graph database comes later. What is being built is an explicit, governed account of what the organisation knows, how it defines it, where each assertion came from, and the contexts in which it holds. The technology stores and exposes that account; it does not settle the definitions. ### We have invested heavily in data quality. Is that not enough? Quality gives the organisation reliable records. Semantic structure makes the meaning and relationships among those records explicit across contexts. Both may be needed where a system must reason across customers, contracts, products, suppliers, and policies rather than retrieve one item at a time. Accurate records with implicit relationships can be difficult to use consistently at scale, while rich relationships built on unreliable records merely make the error easier to traverse. The quality investment is the foundation; structural work extends what can be done with it. ### Why does this reach the Board rather than staying with the data function? Because the definitions are choices with consequences. The definition of customer can shape retention reporting, regulatory scope, and whom an AI system includes or excludes. Once encoded and reused across systems, those definitions influence decisions at scale. Connecting data can also create sensitivity that existed in no single source, so permissions must govern relationships and derived inferences as well as individual records. The Board does not author the definitions, but the few that materially affect revenue, risk, regulatory exposure, or customer trust warrant its visibility. ### Does a knowledge graph make our AI explainable? Not by itself. A knowledge graph can provide provenance for the information and relationships used in an answer, giving an auditor or Risk Committee an evidence path it can inspect. It does not expose the internal reasoning of a model, prove that the model used the information appropriately, or replace contemporaneous decision logs and human review. Where decisions depend on relationships spread across several systems, the graph can be an important part of explainability. It is one layer of the evidence, not the whole answer. ### Where would we start? Start with a question the Board asks repeatedly and receives only a partial or slow answer to: exposure to a jurisdiction, supplier dependence on a commodity, or concentration risk across contracts. These are relationship questions, and the relevant connections are often fragmented across tables, documents, systems, and people. Model one bounded domain, such as suppliers, jurisdictions, commodities, and contracts, then test whether the resulting graph makes repeated analysis faster, more consistent, and more traceable. The aim is to prove the value of the structural work, not to begin with an enterprise-wide graph. ### What happens if we leave this alone? The latent graph exists either way, in your systems' foreign keys and your people's heads. Left ungoverned, it forms by default: definitions set implicitly by whichever system shipped first, relationships encoded by accident, and AI systems reasoning over structure nobody chose. Meanwhile the advantage compounds elsewhere, because structural work is slow, proprietary, and cannot be bought in. A well-governed semantic layer built from proprietary operating knowledge is one of the few durable advantages available while the models themselves commoditise. ## References - **Google** (16 May 2012): [Introducing the Knowledge Graph: things, not strings](https://blog.google/products/search/introducing-knowledge-graph-things-not/). Google's 2012 announcement of the Knowledge Graph, the moment typed entities and explicit relationships entered mainstream search. - **Microsoft** (13 February 2024): [GraphRAG: A new approach for discovery using complex information](https://www.microsoft.com/en-us/research/blog/graphrag-unlocking-llm-discovery-on-narrative-private-data/). Microsoft Research introduces GraphRAG, showing graph-structured retrieval materially outperforms flat retrieval when questions require reasoning across a corpus. - **ISO** (April 2024): [ISO/IEC 39075:2024 — Information technology — Database languages — GQL](https://www.iso.org/standard/76120.html). The ISO/IEC standard for GQL, the property graph query language, published April 2024. - **Financial Reporting Council** (22 January 2024): [UK Corporate Governance Code 2024](https://www.frc.org.uk/library/standards-codes-policy/corporate-governance/uk-corporate-governance-code/). The Code that fixes the responsibilities of chair, chief executive, and Board whose execution AI now changes. - **UK Government** (2025): [Data (Use and Access) Act 2025](https://www.legislation.gov.uk/ukpga/2025/18/contents). The statute that rewrote the UK's automated decision-making regime and introduced the four safeguards. - **European Commission** (2026): [AI Act](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai). The Commission's implementation page: the Act becomes generally applicable on 2 August 2026, with staggered high-risk deadlines to 2028.