Until recently, every BPMN diagram in my workflow cost between three and five hours. Stakeholder interviews had to be re-read, elements manually placed, swimlanes aligned, gateways wired, and notation checked against the specification before anything could be called review-ready. It was detail work — and detail work at that scale compounds fast when you are modeling multiple processes per sprint.
That changed when I built a custom Claude skill specifically for BPMN 2.0 generation. The same diagram now takes fifteen minutes. The output is not a rough sketch — it is a structurally valid, fully annotated BPMN 2.0 diagram that opens directly in draw.io and is 90–95% ready to present.
This is not a conceptual article about what AI might one day do for process modeling. This is a practical account of a workflow change that is running in production, along with the technical reasons the generic tooling failed and what it took to build something that works.
What is BPMN 2.0 and Why Enterprises Rely On It
Business Process Model and Notation (BPMN) is an ISO-standardized graphical language for describing business processes. Version 2.0, maintained by the Object Management Group, defines a precise vocabulary of flow objects, connecting objects, swimlanes, and artifacts that any compliant tool — draw.io, Signavio, Camunda, IBM BPM, and dozens of others — must render consistently.
For enterprises, BPMN serves multiple functions simultaneously. It acts as documentation for auditors and regulators, a specification for development teams building workflow automation, a communication artifact between business and IT, and a baseline for process improvement initiatives. The notation's precision is the point: a task rectangle means exactly one thing, an exclusive gateway means exactly one thing, and a compensation boundary event means exactly one thing — regardless of which stakeholder is reading the diagram.
Kampik et al. identified in the KI journal (2025) that "the entry bar for successfully running a BPM initiative is high" — requiring simultaneous mastery of BPMN tooling and deep organizational process knowledge. This combination is rare, which is precisely what makes AI augmentation so valuable here: the AI handles notation precision while the analyst contributes organizational context.
The notation covers five core element categories: flow objects (events, activities, gateways), data objects, connecting objects (sequence flows, message flows, associations), swimlanes (pools and lanes), and artifacts (groups, text annotations). BPMN 2.0 also introduced an execution semantics layer — meaning the same diagram can describe a process visually for business stakeholders and serve as an execution specification for a BPM engine like Camunda or Flowable. That dual role raises the cost of errors considerably.
The Old Problem: Why BPMN Creation Was So Expensive
Anyone who has modeled processes professionally knows the workflow: attend a discovery session or read through transcripts, then spend the next several hours translating what you heard into a structured diagram. The bottleneck is not comprehension — it is translation fidelity and notation compliance.
Where the hours actually go
The first 30–60 minutes typically involve organizing input material: cross-referencing notes, resolving ambiguities, and mapping out a rough process spine. Then comes layout work — deciding swimlane allocation, placing events and tasks, connecting flows. This is where most of the time disappears. BPMN has strong conventions for element placement that are not technically enforced but matter enormously for readability: left-to-right flow, consistent vertical lane positioning, minimal crossing sequence flows, gateway alignment.
Then comes notation review: are the event types correct (start, intermediate, end — and within those, message, timer, error, signal, compensation)? Are exclusive and parallel gateways used correctly? Are boundary events attached to the right activities? Is data flow represented separately from control flow? These checks are mechanical but unforgiving — a diagram with notation errors will confuse automation engineers and undermine stakeholder trust.
"Such tools may drastically change the way how processes are created in the near future."Köpke & Safan, BPMN-Chatbot study, BPM 2024 Conference, University of Klagenfurt
The final phase is iteration: presenting a draft, collecting feedback, updating the diagram, repeat. If the first draft has structural issues — wrong lane assignments, missing intermediate events, implicit parallel branches left as sequential flows — iteration cycles multiply.
For a single moderate-complexity process with two or three swimlanes, six to twelve tasks, and a couple of decision points, three to five hours is not an unusual total. For organizations modeling dozens of processes per quarter, that is a meaningful budget line.
The Solution: A Custom AI Skill Built on the BPMN 2.0 Specification
Why the stock draw.io MCP failed
The natural first instinct was to use the draw.io MCP (Model Context Protocol) integration that ships with Claude. It can generate draw.io XML — but the results for BPMN were consistently poor. The core issue is that the stock MCP has no embedded knowledge of BPMN 2.0 notation rules. It produces generic shapes: rectangles connected by arrows. They may look like flowcharts, but they are not BPMN. Event types are wrong or missing. Gateways use generic diamond shapes without the correct BPMN markers (X for exclusive, + for parallel, O for inclusive). There are no proper pool and lane containers. Boundary events do not attach correctly. The diagrams fail basic BPMN notation review immediately.
Using an incorrect diagram as a process specification has downstream costs that far exceed the time "saved" by using generic tooling. Development teams working from non-compliant BPMN misunderstand process intent. Automation engineers cannot map the diagram to a process engine. Auditors flag non-standard documentation.
What the custom skill does differently
The custom Claude skill was built from the ground up against the BPMN 2.0 specification. It encodes the complete element taxonomy — all event types (15 start event variants, 42 intermediate event variants, 13 end event variants), all gateway types with correct marker notation, task types (user, service, send, receive, business rule, script, manual, call activity), and the full swimlane model.
Critically, the skill generates syntactically valid draw.io XML that uses the BPMN shape library correctly — not generic shapes with BPMN labels applied post-hoc. The output opens in draw.io and renders as proper BPMN immediately, with correct element rendering, because the XML references the actual BPMN shape identifiers that draw.io's BPMN library expects.
Researchers at the University of Klagenfurt (Köpke & Safan, BPM 2024) demonstrated a BPMN-focused AI tool achieving 95% correctness while using 94% fewer tokens than competing approaches — evidence that domain-specific training and instruction is far more efficient than general-purpose LLM prompting for BPMN tasks.
Step-by-Step Workflow
The complete workflow from raw input to presentation-ready diagram takes four steps and roughly fifteen minutes in total.
Collect and prepare input material
Gather any combination of meeting notes, interview transcripts, requirement documents, process descriptions, or email threads. No special formatting is required — the skill handles unstructured text. Typical input is 200–800 words describing a process from discovery sessions.
Invoke the BPMN skill with context
Pass the input material to the custom Claude skill with a brief context note: the number of swimlanes expected, the domain (procurement, HR, IT, etc.), and any explicit requirements such as compliance checkpoints or SLA-linked timer events. The skill interprets the input and structures the process logic.
Receive and open the draw.io XML
The skill outputs valid draw.io XML. Open draw.io (desktop or web at app.diagrams.net), create a new diagram, and use Extras → Edit Diagram to paste the XML directly. The diagram renders immediately with correct BPMN notation, swimlanes, and flow connections.
Review and fine-tune (5–10 minutes)
90–95% of diagrams require only minor adjustments: layout polish for visual clarity, adding domain-specific subprocess detail, or applying corporate style guidelines. Structural corrections — wrong lane assignments, missing logic branches — are rare because the skill encodes BPMN structural rules explicitly.
What the Skill Delivers: Full BPMN 2.0 Coverage
The custom skill does not produce a simplified subset of BPMN. It covers the complete specification for business process diagrams (as opposed to choreography or conversation diagrams, which have separate use cases). This is what you get in every generated output:
- Correct pool and lane containers with proper label positioning and hierarchy
- All start event types: None, Message, Timer, Conditional, Signal — with correct filled/unfilled circle notation
- All end event types: None, Message, Error, Escalation, Compensation, Signal, Terminate — with correct thick-border and fill conventions
- Intermediate events (catching and throwing): Timer, Message, Error, Signal, Compensation, Link — attached to sequence flows or as boundary events on activities
- Exclusive (XOR) gateways with X marker and properly labeled outgoing conditional flows
- Parallel (AND) gateways with + marker for synchronizing/forking parallel branches
- Inclusive (OR) gateways with O marker where multiple paths may be active simultaneously
- Event-based gateways for race conditions between incoming messages or timer events
- All task types: User Task (person icon), Service Task (gear icon), Send/Receive Tasks (envelope icons), Business Rule Task, Script Task
- Sub-processes (collapsed and expanded) with correct BPMN marker on the activity frame
- Sequence flows with correct condition notation on gateway outgoing flows
- Default flows marked with a tick-slash on the source gateway
- Data objects and data stores with correct association connections to activities
- Text annotations linked to specific elements via association connectors
- Correct flow direction: left-to-right with top-down swimlane ordering following BPMN readability conventions
Researchers at Ulm University (Hörner, Möller, Reichert — Business & Information Systems Engineering, February 2026) evaluated BPMNGen, an AI BPMN generation system, with 67 participants and found that AI-generated models "can deliver expert-level modeling quality with considerably less time and effort" for simple and moderate processes, and were rated "equally interpretable and cognitively manageable" as expert-created models. The majority of business process modeling tasks in enterprise settings fall squarely in this range.
Comparison: Manual vs. draw.io MCP vs. Custom AI Skill
| Criterion | Manual (draw.io) | Stock draw.io MCP | Custom AI Skill |
|---|---|---|---|
| Time per diagram | 3–5 hours | 30–60 min (+ correction) | ~15 minutes |
| BPMN 2.0 notation compliance | High (if expert) | Low — generic shapes | High — spec-based |
| draw.io compatibility | Native | Partial — may need reshaping | Native XML output |
| Correct event types | Yes (manually set) | No — mostly plain circles | Yes — full event taxonomy |
| Correct gateway markers | Yes (manually set) | No — unlabeled diamonds | Yes — X, +, O, pentagon |
| Swimlane/pool structure | Yes | Often absent or informal | Yes — proper pool/lane XML |
| Requires BPMN expertise from user | Yes — high bar | Moderate | No — notes are sufficient input |
| Output ready to present | Yes | No — requires rework | 90–95% without edits |
| Scales to multiple diagrams/week | Costly — linear time | Better, but quality issues compound | Efficient — consistent quality |
| Engine-executable XML (Camunda/Flowable) | With additional work | No | Draw.io XML — requires BPMN 2.0 export for execution |
ROI Calculator: What Does This Actually Save You?
The numbers become compelling quickly when applied to realistic enterprise workloads. Use the calculator below to estimate monthly and annual savings based on your team's situation.
Process Modeling ROI Calculator
Calculation assumes 15 min per diagram with AI skill vs. the entered manual average. Does not include indirect savings from faster iteration cycles and reduced rework.
Live Example: Purchase Approval Process
The following is a representative output from the custom skill — a Purchase Approval Process with three swimlanes. This is what the skill generates from a typical set of meeting notes describing a procurement workflow.
Purchase Approval Process — AI-generated BPMN 2.0 example with three swimlanes. Purple: standard sequence flow across lanes; green path: approved → Place Order → End; red path: rejected → Send Rejection Notice → End. Dashed arrows indicate cross-lane message/sequence flows.
This diagram illustrates the key structural properties the skill consistently produces: correct pool/lane hierarchy, proper exclusive gateway with labeled conditional outgoing flows, differentiated end events (green success, red terminate), and task type icons (user task for human activities, send task for the notification step). In practice, generated diagrams of this complexity arrive in under a minute of generation time.
Hermes Agent Platform: Enterprise-Grade Integration
The custom BPMN skill described here is one component of a broader automation architecture. The Hermes Agent platform provides the infrastructure to deploy skills like this into existing enterprise toolchains — connecting them to the document repositories, collaboration tools, and BPM suites where process work actually happens.
In practice this means the skill can be triggered from a Confluence page, a SharePoint document library, or a Jira ticket — pulling source material automatically, generating the BPMN output, and writing the draw.io file back to the appropriate location. Teams working in Camunda, Signavio, or Bizagi can receive outputs formatted for direct import rather than manual paste.
For organizations running process improvement programs or compliance-driven process documentation at scale — modeling dozens to hundreds of processes per year — the compounding effect is significant. Every process analyst on the team operates at the output rate of the fastest expert. Onboarding time for BPMN tooling drops substantially. Process documentation stays current because the cost of updating a diagram approaches the cost of updating a text document.
Deploy a Custom BPMN Skill in Your Organization
Whether you need a standalone BPMN generation workflow or full integration with your existing BPM toolchain, we can build and deploy it within your enterprise environment. Schedule a free architecture consultation to discuss your specific process modeling workload.
Book Free ConsultationFrequently Asked Questions
References
¹ Avolution Software / AutomatedDreams.com (2024). Business process modeling adoption statistics.
² Köpke, J. & Safan, M. (2024). BPMN-Chatbot: Generating BPMN Models Using Conversational AI. University of Klagenfurt. Proceedings of BPM 2024.
³ TechTarget (February 2026). Review of generative AI integration across top BPM platform vendors.
⁴ Hörner, M., Möller, C., & Reichert, M. (2026). BPMNGen: AI-generated vs. expert-created BPMN models — a comparative study (N=67). Business & Information Systems Engineering, February 2026.
⁵ Kampik, T. et al. (2025). On the entry bar for AI-assisted BPM initiatives. KI – Künstliche Intelligenz, 2025.