A knowledge graph is a way of storing information as entities, such as customers, accounts, companies and products, and the named relationships between them, such as "owns", "is director of" or "guarantees". It records how things connect, not just what they are. For AI agents it answers questions that depend on those connections, which plain document search handles badly.
Why does a knowledge graph matter once agents act?
Many regulated decisions turn on relationships. Is this applicant a director of a company that already has a loan with us? Does this payee share an address with an employee? Which customers are affected if this product's terms change? An agent searching documents may find the pieces but miss the link between them. A graph holds the link directly.
That matters most when an action depends on it. An agent approving a supplier payment should know if the supplier is connected to the person who raised the invoice.
How does a knowledge graph work in practice?
Facts are stored as triples: subject, relationship, object. "Company A, has director, Person B." "Person B, lives at, Address C." The agent, or a tool it calls, queries the graph by following relationships: start at this payee, find every person connected within two steps, check whether any of them work here.
Each fact in the graph should carry where it came from and when, and its provenance class. A director relationship drawn from an official register is stronger than one a customer typed into a form. Without that, the graph turns a mixture of sources into one confident-looking network. Whether your agents need one at all is covered in do AI agents need a knowledge graph?
What does that look like in an example?
A commercial lender's credit agent prepares a facility for a trading company. The graph shows the company's two directors, and that one of them is also a director of a company in arrears with the same lender. Document search on the application would not have surfaced that; the link sits across two separate files. The agent flags the connection in the credit memo for the credit officer. It does not decide what the connection means; that stays with a person.
How is a knowledge graph different from a knowledge base?
A knowledge base is a collection of written material: procedures, product guides, policies, answers to common questions. It is organised for people and agents to read. A knowledge graph is a structured set of facts and relationships, organised for querying. A knowledge base tells the agent how to handle a transfer request; a knowledge graph tells it who is connected to whom. Most agents need the first; fewer need the second. Writing the first well is covered in how to write a knowledge base an AI agent can use.
What are the common mistakes?
Building a graph before knowing which questions it must answer, which produces a large structure nobody queries. Mixing sources of very different strength without labelling them. And letting agents write new relationships into the graph unchecked, so an agent's inference becomes a fact other agents rely on. Inferred links should be marked as inferred, with the agent that created them.
Questions people ask
Is a knowledge graph the same as a vector database?
No. A vector database finds content that is similar in meaning. A knowledge graph follows explicit, named relationships. Some systems use both: vectors to find the right entity, the graph to explore its connections.
Who should maintain a knowledge graph?
The owners of the source data, with each relationship type traced to the system that supplies it. The governed agentic operating system guide explains how the facts it holds stay stamped with their origin.