What changes
What AI changes about process modelling
Modelling used to be a translation job: someone who knows the notation listens to someone who knows the process, and draws. AI removes the translator, not the person who knows the process. The comparison below is between two ways of using AI, not between AI and a modeller.
| Asking a chatbot for a process description | A generated diagram on a canvas | |
|---|---|---|
| What you get | Text: a numbered list of steps, sometimes with roles in brackets. Good for a first outline. | A BPMN 2.0 diagram with lanes for the roles, gateways for the decisions, systems, documents and control steps. |
| Where it is right or wrong | Hard to see. A plausible paragraph reads the same whether the sequence is correct or not. | Visible at a glance. A step in the wrong lane or a decision without a "no" path is obvious to the people who run the process. |
| Correcting it | Ask again and get a new text. Comparing two versions means reading both in full. | Edit on the canvas or describe the change. Proposals are shown before they land; every version stays. |
| What happens next | Copy the text into a modelling tool and draw it after all. | Export as BPMN 2.0 XML, which opens in standard BPMN tools, or as SVG and PNG, and build the analyses (XLSX) and the deck (PowerPoint) on the same diagram. |
| What neither can do | Know what nobody said. | Know what nobody said. The diagram contains the process as described; the people in the room complete it. |
In between
From the conversation to BPMN 2.0: what happens in between
The description is a sequence of sentences about people, decisions, systems and checks. The diagram is the same information in a notation that has a symbol for each of those. The step in between is mapping one to the other, and asking back where the mapping is not possible.
- 01
Roles become lanes
"The handler registers the claim, then the underwriter checks coverage": two roles, two lanes. A role that is mentioned once as "someone in finance" becomes a lane worth naming precisely.
- 02
Decisions become gateways
"If it is covered we pay, otherwise we decline with reasons": an exclusive gateway with two labelled paths. Where the rule behind the decision is missing, that is the question to put to the room.
- 03
Systems and documents become elements
"We book it in the claims system and attach the policy file": the system is attached to the step, the policy file becomes a document element. Later, the analyses know which step touches which system.
- 04
Checks become control steps
"Before payout a second person approves": a four-eyes check is a control step, marked so that compliance analyses, audits and later automation can find every control in the process.
With responses switched on, patchley asks follow-up questions while you describe. Ask it what is unclear or missing in the draft, a decision without a rule, a step without an owner, a path that never ends, and it points it out. You decide what goes into the diagram.
Two readers
One diagram, read by people and by machines
A lane for an AI agent is modelled like any other lane: a role with steps, surrounded by the control steps where a person decides. Designing it is a To-Be exercise; running it is the Executions pilot, in preparation.
The figure shows a fictional offer approval with three lanes, one of them for an AI agent. For the people in the room the diagram is a picture: who does what, in which order, where a decision is made. For a machine, the same file is a structure: lanes with names, steps with systems, a gateway with two conditions and a control step that must not be skipped.
That is why the output is standard BPMN 2.0 and not a drawing. The generated diagram is useful for agreeing a process with people today, and the same file describes the process precisely enough to be the guardrail when parts of it are later handed to an agent. The page on processes before AI agents makes that argument in full.
Control
The analyst proposes, you decide
Generating a diagram once is easy. Keeping it the organisation's diagram, rather than the AI's, is the part that matters. The division of labour is fixed: the analyst proposes, every proposal is shown before it is applied, and the people who own the process decide.
What the analyst proposes patchley
- The first draft from the conversation or the document
- Changes you describe in plain words: "the check comes before registration"
- Questions where the description has gaps
- Updated analyses after the diagram changed
What you decide the process owners
- Whether a proposal lands: apply it or reject it, after seeing it
- Everything you change on the canvas yourself; it is never overwritten
- Which version is signed off; every version stays
- What leaves patchley, and where it goes
Every change is shown before it lands
Whether you describe a change to the analyst or ask it to resolve an open question, the result is a proposal on the canvas, not a changed diagram. You see what would move, be added or be removed, and then apply or reject it. That holds for the diagram, for the analyses and for the deck.
Versions, hand edits and what never gets overwritten
Your own edits are never overwritten by a later proposal, and every version of the diagram stays available. When the process owners sign off a version, that is the one the analyses and the deck refer to, and the earlier drafts remain there to show how it got there.
Limits
Limits: where you still need a human
A generated diagram is a faithful notation of what was said. It is not a substitute for the people who know the process, and it is not a sign-off. Four limits worth knowing before you start.
What nobody mentions is not in the diagram
The analyst does not invent steps, rules or roles. If the exception path was never described, it is missing, and the only way to find it is to ask. The process capture page lists the questions that bring exceptions out.
The human signs off
A diagram becomes the organisation's process when the people who run it confirm it. patchley can draw, ask and propose; it cannot know whether the description was right.
A description can be wrong
People describe the process as they believe it runs. Two interviewees can give two sequences; the diagram shows the difference, and the people resolve it.
Describing is not running
Today the model describes the work. Running it, with people, systems and AI agents per step, is the Executions pilot, in preparation, not a released feature.
Standards
Standards and portability: BPMN 2.0 XML
The diagram you agreed is yours to take anywhere. The path from the first description to a file in your own tool has four steps today and a fifth in preparation.
STANDARD FORMAT · NO LOCK-IN
What the export contains
The BPMN 2.0 XML carries the process and its diagram interchange: pools and lanes, tasks, gateways and their paths, start and end events, the systems and documents attached to steps, and the control-step marking. Open it in a standard BPMN tool and you see the same diagram, ready to be taken further there. Examples of generated diagrams, with the .bpmn file on request, are on the examples page.
The application runs in a data centre in Germany. The AI models that read the conversation and draft the diagram come from a European provider, hosted in the EU. Details on the security and data page.