BPMN With AI: Draft and Review Process Diagrams

Use AI with BPMN to turn a verified description of work into a draft diagram, then check every path with the people responsible for that process. BPMN means Business Process Model and Notation, a standard visual language for showing activities, events, and routing decisions. AI can help organize the description; it cannot establish whether the business actually works that way.

fenbs keeps the agreed process rules, open questions, and follow-up tasks outside the diagramming session. That helps a later connected assistant understand why a branch exists instead of simplifying away a necessary exception.

What steps should you follow to draft a process with AI?

1. Choose one process boundary. Name the triggering event, the expected outcome, and the people or systems involved.
2. Describe the work in plain language with a process owner. Separate the current process from proposed improvements.
3. Ask AI to list activities, decisions, exceptions, and unanswered questions before drawing anything.
4. Review that list, then request a BPMN draft using the confirmed information. Import or recreate it in a BPMN-aware modeling tool.
5. Check the notation and walk through representative cases with the people who perform the work.
6. Record the approved model, its rules, and unresolved changes. Treat any executable deployment as a separate reviewed step.

Suppose a small distributor wants to document a return request. Start with request received and end with return closed. Listing warehouse work, customer messages, and refund decisions is more useful than asking an assistant to invent a best-practice returns process from nothing.

What information should you collect before drawing?

Ask what starts the work, who acts next, and what evidence lets the work continue. For the distributor, a return might require an order reference and photographs. Find out who checks those details and what happens when they are missing.

Get the actual decision rules from their owner. If the assistant does not know the refund threshold, tell it to mark the threshold as unknown. A plausible invented value makes the diagram look complete while quietly changing the policy.

• Trigger: the event that creates a new case.
• Inputs: the information or documents needed at each stage.
• Responsibility: who performs the work and who may approve an exception.
• Conditions: the rules that determine which path follows.
• Exceptions: missing information, rejection, timeout, cancellation, and rework.
• Completion: the outcomes that count as closed and the evidence retained.

Keep current behavior separate from improvement ideas. A useful AI brain preserves that distinction between fact and proposal. Put the confirmed process rules in fenbs project context, so later drafting begins with what the team actually agreed.

Which BPMN concepts must the draft get right?

The OMG BPMN specification↗ (opens in a new tab) is the primary reference for the notation. A diagram with boxes and arrows is not automatically a correct BPMN model. Use a modeling tool that understands the notation and have someone familiar with BPMN review unfamiliar constructs.

For a first draft, focus on activities, events, flows, and gateways. Camunda BPMN documentation↗ (opens in a new tab) explains that sequence flows show ordering, an exclusive gateway selects one path, and a parallel gateway activates multiple paths. These differences change the process meaning, so they deserve more attention than diagram layout.

Do not request executable XML until you know whether execution is needed. XML is structured text a modeling tool can read. Being able to import that text proves something about its format, not that the process reflects reality or is safe to run.

What mistakes should you look for in an AI-generated model?

Look for invented approvals, missing rejection routes, unexplained loops, and paths that never reach a meaningful outcome. Compare the draft line by line with the confirmed process description. Ask the assistant to identify which source statement supports each activity and condition.

Check for changes hidden inside reasonable wording. Replacing staff approves refund with system issues refund is a substantive policy and implementation change. Likewise, adding a deadline is not harmless diagram cleanup when the business has never agreed to it.

Run the modeling tool checks and resolve the issues it reports. Then keep semantic review separate: a valid diagram can still describe the wrong responsibility or omit a real exception. AI review can provide a second list of questions, but the process owner must resolve the business meaning.

How do you review the model with the people doing the work?

Walk through cases aloud instead of asking whether the diagram looks right. Follow an ordinary accepted return, a rejected return, a missing order reference, and a customer who does not reply. At each step ask: who acts, what do they need, and how do they know they can continue?

Use anonymized or invented example cases rather than exposing customer records. Write down disagreements as questions with an owner. If warehouse and support teams describe different handoffs, the diagram has revealed a decision to make; the assistant should not quietly choose a winner.

An AI approval workflow helps distinguish a draft from an accepted process. A short agile workflow can organize the corrections without turning each diagram revision into an unbounded redesign.

How to do this with fenbs

1. Create a free account and add a project for the process, including its trigger and completion boundary.
2. Add context notes for confirmed steps and a glossary for terms the teams use differently. Link to the diagram in its modeling tool.
3. Create tasks for unanswered questions and review corrections. Record approved process choices in Decisions and rules, with the person who decided.
4. Connect a supported assistant through the integration guides and ask it to read the context before drafting or revising.
5. Save Where we left off with the latest reviewed model, unresolved branches, and the next walkthrough needed.

fenbs does not draw BPMN diagrams or execute the process. If you only need a simple informal checklist that one person owns, a BPMN model and a separate shared record may add more upkeep than value.

What should you publish after the review?

Publish the accepted model with its purpose, owner, review date, and known limits. State whether it documents current work or a proposed future process. Keep implementation changes and unresolved policy questions visible, rather than hiding them inside a final-looking diagram.

For gathering and checking material before modeling, see AI research workflows. The AI at work hub covers other practical workflows. Revisit this model when an actual rule or responsibility changes, and preserve the reason so the next revision starts from an accurate account.