--- title: "Your AI Agent Has the Data. It Is Missing the Relationships. Let's talk about Graphs." author: "Dave Glaser" publication: "The AI Operator" published_at: "2026-08-14" canonical_url: "https://qstve.com/writing/linkedin/ai-operator/your-ai-agent-is-missing-the-relationships/" original_url: "https://www.linkedin.com/pulse/your-ai-agent-has-data-missing-relationships-dave-glaser-1v6we" content_type: "essay" primary_primitive: "context-graph" secondary_primitives: ["evidence", "operating-model"] original_content_hash: "sha256:0c0032798a20e30b23b3c5c4581fb611f84cc0d0e01e44d6d472d20199275903" provenance: "A QSTVE Publication" --- # Your AI Agent Has the Data. It Is Missing the Relationships. Let's talk about Graphs. *Why the next practical control layer is a small map connecting intent, authority, action, and evidence. We call these Graphs.* *The AI Operator | August 14, 2026* An auditor asks who authorised an agent to release a payment. Everyone can produce a record. Finance has the transaction. Identity has the user and role. The workflow platform has a run log. The approval tool has a ticket. The agent has a transcript. Yet nobody can answer the question without opening five systems and reconstructing the chain by hand. That isn't a missing-data problem, it's a missing-relationship problem. ## What's Happening This Week As AI systems gain more tools and longer-lived context, operators are discovering that lists and dashboards are excellent at showing things and rather poor at explaining how those things depend on one another. Anthropic's guidance on context engineering makes the first constraint plain: an agent’s context is finite, so simply stuffing more records into the prompt is not a durable answer. A graph, in operator language, is a map of things and the relationships between them. Amazon Web Services (AWS) describes graph data as a network of entities and relationships; Neo4j's beginner explanation puts the emphasis in the same place, the connections are represented explicitly rather than reconstructed for every question. In a payment workflow, the things might be a person, role, account, permission, limit, transaction, approval, exception, and evidence record. The relationships are verbs: the person holds the role; the role grants the permission; the permission applies to the account; the transaction used the permission; the approval authorised the exception. A spreadsheet can contain every item. The graph makes the path itself something you can inspect. ## Why This Is a CFO Problem Consider a flat permission report: Agent A, Account 42, $50,000 limit, active. Useful, but incomplete. It doesn't tell you who granted the permission, for which business purpose, whether the approver still holds the right role, which policy supplied the limit, or what evidence must exist after the action. Connect those records and different questions become possible. Who gave this agent authority? What approvals did the payment depend on? Which other workflows will fail if the account tool is unavailable? Can an auditor travel from intent to action to evidence without relying on somebody’s memory? This is where a graph-shaped operating model helps AI. It can retrieve the surrounding context instead of merely the nearest document. It can inspect dependencies before acting. It can show the authority path alongside the proposed action. During an incident, it can help identify affected workflows and responsible owners. National Institute of Standards and Technology (NIST)'s research connecting evidence graphs and attack graphs illustrates the broader principle: linked evidence and dependencies can help reconstruct a path that isolated records obscure. But a graph is not magic. A stale relationship is still stale. An unnamed owner is still absent. A link inferred by a model is not equivalent to an approved policy record. More connections can create a more convincing wrong answer unless consequential relationships carry a source, owner, and freshness rule. Nor does every team need a graph database. “Graph” describes the shape of the operating problem before it describes a technology purchase. A whiteboard, a Markdown file, or a small structured dataset may be enough to prove whether the connected view answers a useful question. ## The Operator's Log I see the same distinction in the AI stack I run. A folder can tell me which jobs, prompts, reports, and payloads exist. That inventory becomes operational only when I can follow the connections: which schedule triggered this job, which sources it was allowed to read, which permission tier applied, which preflight passed, which artifact became the handoff, and where a human decision remained mandatory. I do not need a grand knowledge graph to begin seeing those relationships. A small map is enough to expose missing owners, ambiguous authority, stale dependencies, and evidence that nobody has actually connected to the decision. The useful question is “Can we trace this consequential action through the relationships that made it legitimate?” ## Money Move Pick one consequential workflow this week and build the smallest useful operating graph: 1. Write down 5–10 important entities: people, agents, roles, tools, systems, permissions, actions, and evidence. 2. Draw the relationships between them using verbs such as owns, grants, uses, approves, produces, and depends on. 3. Mark the owner and source of truth for each consequential relationship. 4. Add a freshness rule: what event or date could make that relationship untrue? 5. Test the map with one awkward question: if this payment, tool, approval, or agent fails at 2 a.m., can you find the affected workflow, the responsible human, and the evidence needed to recover? The goal is not a complete enterprise graph. It is one operating map that answers a question your current list cannot. I would like to hear from operators building AI into finance, payments, risk, or operational workflows. Which relationship is hardest to trace today—and where has a simple connected map exposed a control gap your dashboard missed? -dg ## Sources - [AWS: What Is a Graph Database?](https://aws.amazon.com/nosql/graph/) - [Neo4j: Graph Databases—How Relationships Change Everything](https://neo4j.com/blog/developer/neo4j-graph-databases-for-beginners-2023-edition-chapter-1-relationships/) - [Anthropic: Effective Context Engineering for AI Agents](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents) - [NIST: Mapping Evidence Graphs to Attack Graphs](https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=911920)