From Process Data to AI Agents: How a Knowledge Graph Is Created

There is usually a 12-month gap between “processes being documented in some way” and “an AI agent acting on them.”

With a structured approach, it takes four weeks. Not as a promise—but as the result of a different starting point: processes are built from the ground up as a machine-readable structure.

Step 1: Raw Data — What the Company Already Knows

Every company already has process knowledge—unstructured and scattered across system logs (ERP, CRM, ticket systems), existing documentation (manuals, wikis, SOPs), and key individuals who possess that tacit knowledge.

In the knowledge graph approach, it is extracted automatically—through process mining and structured short interviews.

Time commitment: 3–5 hours per week, 1–2 weeks.

Result: a complete raw material.

Step 2: Extraction — Turning Raw Data into Entities

Entities are identified: processes, roles, systems, rules, exceptions—each as its own node.

The following characteristics are recorded: duration, frequency, and responsibility.

Conflicts become visible: If system data shows that a step is skipped in 40% of cases, even though the manual describes it as mandatory—that is flagged. Not resolved, just made visible. That is information.

Result: Clean entities with properties.

Step 3: Linking — Entities Become a Knowledge Graph

A knowledge graph is a network of relationships:

Thousands of connections. No one reads this graph—machines query it in milliseconds.

Result: a machine-readable, interconnected model of the company.

Step 4: Activation — The graph becomes usable for AI agents

Query Layer: Agents query the graph: “What rules apply to suppliers with a value over 200k?” → structured response.

Context Injection: Relevant context is automatically injected. The agent does not receive everything—only what is relevant to its current task.

Continuous Learning: Corrections are incorporated. The graph learns along with the company—without manual maintenance.

Result: An agent that acts based on corporate knowledge—not on assumptions.

The Entire Journey

• Weeks 1–2: Collect raw data
• Weeks 2–3: Extraction
• Weeks 3–4: Networking (Knowledge Graph)
• Starting in Week 4: Activation (AI Agents)

After Week 4, the graph isn't finished—it's operational. It's constantly improving.

The difference from traditional BPM: The result isn't a document that becomes outdated. It's a system that learns.