In 1855, Daniel McCallum needed to coordinate roughly 200 trains and 4,700 employees across the New York & Erie Railroad. He drew a chart. It was a practical document, born from the specific problem of knowing who was responsible for what across hundreds of miles of track in a country still figuring out how to organize things that moved faster than any person could follow. The chart worked. By the mid-twentieth century, every corporation had one.
What happened next was gradual enough that nobody marked the transition. The chart was drawn to coordinate. People started using it to navigate. Managers consulted it to understand who decided what. New hires studied it to locate influence. Consultants used it to diagnose dysfunction. The representation became the thing being represented.
Organizational researchers spotted the slippage early. Krackhardt and Hanson wrote about "the company behind the chart," the informal networks where actual decisions traveled. But observing the gap and measuring it required different instruments. A study of Microsoft's internal email network analyzed 95.5 million employee-to-employee messages across 88 teams and 12 divisions. Communication did cluster around formal structure, with within-team email density running roughly 45 times higher than between-team density. The chart captured real signal. It also missed the cross-cutting flows, the informal escalation paths, the conversations that shaped outcomes without appearing in any reporting line. Partially right is a difficult thing to calibrate against.
Process mining offered a way to formalize what everyone sensed. The IEEE Task Force on Process Mining published its manifesto articulating the principles: discover actual workflows from event logs rather than from interviews or documentation. The field gave organizations a methodology for comparing how work was described with how work was performed. The divergence it surfaced wasn't limited to corporate hierarchies. Wherever prescribed processes met practiced ones, the gap appeared.
The sharpest documented instance comes from healthcare. When a team at Uniklinik Aachen built a normative process model from Germany's COVID-19 clinical guidelines and compared it against 187 ICU patient records, the model-log fitness came back at 0.69 for the first pandemic wave. Activities the guidelines prescribed as sequential appeared out of order. Steps the model required were skipped entirely. The guideline described one process. The ICU operated another. Neither was fictional. The same dynamic plays out in enterprise workflows, where process documentation and daily practice coexist as parallel realities that rarely get reconciled.
This is the landscape automated systems now enter. They receive the documented version: the process diagram, the API spec, the stated approval chain. These artifacts are the org chart's descendants, coordination tools that have quietly hardened into navigational maps. The actual workflow, shaped by years of workarounds and institutional memory, lives in event logs and in the heads of people who've learned which steps matter and which ones everyone skips.
Process mining gave human analysts a way to see the divergence. Agents encounter it as something more immediate. They follow the documented path and find the territory has been rearranged by people who never updated the map.