01 — The starting point
AI alone is not enough. BPMN alone is not either.
Classic BPMN is rigid and rule-based. AI on its own is probabilistic and hard to control. Together they become regulated flexibility: the AI acts dynamically inside limits that were defined before it started.
| Classic BPMN alone | AI alone | AI + BPMN | |
|---|---|---|---|
| Way of working | Rigid and rule-based, inflexible with exceptions. | Probabilistic, a black box. | Regulated flexibility: the AI acts dynamically inside defined guardrails. |
| Governance | Definable and auditable in advance. | Hard to control, risk of hallucinations. | Hard limits, for example a four-eyes approval, that an AI cannot skip. |
| Creation and upkeep | Laborious, manual, time-consuming. | Quick drafts, but without reliable structure. | The AI drafts the model from the conversation, the person checks and decides. |
02 — One language, two readers
The notation may look old. As a language for AI it is more current than ever.
Many see BPMN as yesterday's cumbersome modelling language. But it is precise, standardised and machine-readable, which is exactly what AI agents need to complete tasks reliably.
For people: the diagram Reader 1 · human
- Who does what, in which order
- Bottlenecks and hand-overs at a glance
- The basis for agreement and sign-off
For the AI: the workflow behind it Reader 2 · AI agent
- Responsibilities (lanes), decisions (gateways), data and systems per step
- Controls and approvals that must not be skipped
- The goal: which result has to stand at the end
The visualisation is for people. The structure behind it is for the AI, so that it does not merely act but reaches the right result.
03 — The new role of BPMN
From the handbook on the intranet to the operating system for AI agents.
BPMN does not die because of AI. It moves from a tiresome documentation duty to the control layer for intelligent workflows. Four roles it takes on:
Guardrail for AI agents
The model defines when an agent works, which data it gets and when a person steps in (human in the loop).
Traceability and audit
When AI makes business-critical decisions, it must be visible where it was active. Rules for high-risk AI applications, such as the EU AI Act, ask for exactly this kind of traceability.
Check the effect before you automate
Whoever speeds up one step without knowing the process only moves the backlog to the next manual hand-over.
Process knowledge that answers
Staff do not need to read the diagram. They ask "Who approves this offer?", and the answer comes from the stored process.
04 — The prerequisite is missing
All of this only works with transparent processes. That is exactly where it fails.
AI can only automate what it knows. In most organisations the actual sequence of work is not cleanly described anywhere, and smaller and mid-sized companies feel this most.
AI automates what it sees. What is not described, it does not see.
Documented, but outdated
The process handbook describes the state at the last certification, not everyday work today.
Only in people's heads
How it really runs is known to a few experienced colleagues. When the knowledge leaves, the process leaves with it.
No staff department for it
In a mid-sized company the management itself has to do this. Modelling projects lasting weeks are simply not an option.
05 — Where this leads
Every jump in AI came from better context.
From a single prompt to rules, knowledge and actions: each step gave the AI more context. The next step is structure, the process graph. After that, the model does not describe the work any more; it runs it.
Today the model describes the work. Next, it runs it. That second step is the Executions pilot, in preparation with a small group of pilot organisations, not a released feature.
06 — How patchley closes the gap
From the conversation to a process that people and AI understand.
patchley collects process knowledge where it lives: in people's heads and in the conversation. No modelling project, no staff department, just your own words.
STANDARD FORMATS · NO LOCK-IN: BPMN 2.0 XML OPENS IN SIGNAVIO, ADONIS AND CAMUNDA MODELER · AI MODELS FROM A EUROPEAN PROVIDER, HOSTED IN THE EU