Context Graph
Last updated
A context graph captures how work happened and why decisions were made along the way, rather than only the outcome a system recorded. It preserves the coordination, precedent, and evidence that made a decision make sense.
What Context Graph actually means
Systems of record capture outcomes. A CRM knows a deal closed. An ERP knows an invoice was paid. An HR system knows someone was hired. Each of those is a fact about a result, stored after the fact, with the reasoning stripped out.
A context graph captures the part that is discarded. When a discount is approved, it preserves the coordination that made the approval possible: the back and forth with the approver, the escalation history that created urgency, the similar case from last quarter that set precedent, and the evidence reviewed before deciding. When one handler processes a claim in twenty minutes and another takes two hours for identical work, it holds the different steps, checks, and workflows that produced the gap.
This is organizational memory made visible and queryable. Not only which decisions were made, but how work actually flowed and what context made those decisions reasonable at the time. Rules describe what should happen in general. A context graph records what happened in specific cases, and what made it work.
The reason this matters now is that both people and AI agents need that context to operate reliably. An agent given a rule can follow the rule. An agent given the accumulated context of how similar cases were actually handled can handle the case in front of it, including the ones the rule never anticipated.
Examples
The reasoning behind an approval
A discount is approved. The CRM records the discount. The context graph records the three hours of discussion with the VP, the customer escalation history that created the urgency, and the comparable case from the prior quarter that set the precedent.
Why identical work takes different time
Two claims of the same type take twenty minutes and two hours. The context graph holds the different checks each handler ran, the order they ran them in, and which of those checks changed the outcome.
Context an agent can act on
An agent meets an exception its rules do not cover. Rather than escalating or guessing, it queries how comparable exceptions were actually resolved, including what evidence was gathered and who was consulted.
Frequently asked questions
Related terms
Work Ontology
A work ontology is a living, continuously updating graph of how an enterprise actually operates, built by observing work at the point of execution. Its nodes are the actors, activities, and artifacts of work, and its edges are the handoffs between them, with intent inferred for every action.
Agent Grounding
Agent grounding is connecting an AI agent to a current model of the work it is acting on, so it decides from what is true now rather than from rules written when it was built.
Execution Data
Execution data is the record of how work actually moves through an organization: the layer between the inputs a business captures and the outputs it reports. It covers workflow patterns, collaboration structures, and where work slows down.
