Technical Lead & BPMN Educator·8 min read
AI agents need process context
AI agents are moving beyond chatbots. They are becoming operational actors that carry out tasks, make decisions, and hand off work to other agents or humans. To do this reliably, they need to understand business processes: which tasks exist, in what order they happen, who is responsible, what decisions need to be made, and where handoffs occur.
The question is no longer whether to give AI agents process context. The question is in what format. That choice determines how much extra infrastructure you need to build and how well AI can actually work with your processes.
BPMN is already in the training data
BPMN has been an ISO standard (ISO/IEC 19510) since 2013. Before that, the Object Management Group maintained it as a public specification since 2004. Over two decades, a massive body of material has accumulated: the specification itself, university courses, textbooks, tutorials, academic papers, Stack Overflow answers, blog posts, and countless examples of BPMN XML.
All of this material is in the training data of every major large language model. GPT-4, Claude, Gemini, Llama, Mistral: they have all ingested BPMN specifications, tutorials, and XML examples during pre-training. BPMN is not something you need to teach an LLM. It is something the LLM already knows.
The practical proof is simple. Export any .bpmn file and paste the raw XML into ChatGPT or Claude. Ask it to explain the process. It will correctly identify tasks, gateways, events, sequence flows, and participant lanes. It understands the structure because it has seen thousands of examples during training.
"Try it yourself: export any .bpmn file and paste the XML into ChatGPT or Claude. Ask it to explain the process. The results are surprisingly accurate."
Why native AI readability matters
Because BPMN is natively understood by LLMs, the format itself becomes an interface between your process knowledge and AI capabilities. An LLM can read BPMN XML the way it reads English: it understands the structure, the semantics, and the relationships between elements.
This is what makes BPMN different from proprietary formats in the AI era. You do not need a vendor to build an integration between your process data and AI. The standard is the integration. Any AI model, current or future, can work with your BPMN files without a custom connector.
In practice, this means process tools that use BPMN can build AI features on a foundation that LLMs already understand. The result is more reliable AI assistance, faster development of new capabilities, and no dependency on a single vendor's integration layer.
"Raw AI readability is the foundation, but it is not the whole story. The real value comes when AI is integrated into the modeling workflow itself: visual feedback, version control, team collaboration, governance. Pasting XML into a chat window proves the concept. Integrated tooling makes it practical at scale."
Why proprietary formats need adapters
Proprietary process formats use internal identifiers, undocumented schemas, and vendor-specific data structures. These formats were designed to work inside one tool, not to be read by external systems. LLMs have never seen these formats during training because the data is not publicly available.
The workaround is to build a connector: an MCP server, an API integration, a plugin that translates the proprietary format into something AI can read. This works, but it adds a layer of complexity. You depend on the vendor to maintain the connector. You depend on the translation being accurate. And every new AI model or agent framework needs its own integration.
With BPMN, there is no translation step. The format itself is the interface. Any LLM can read it directly. Any agent framework can parse the XML. Think of it as the difference between teaching a translator a language they have never heard and speaking a language they already know fluently.
The portability advantage
AI readability is one dimension. But choosing an open standard over a proprietary format has broader implications for how portable and future-proof your process models are.
| Criteria | BPMN (open standard) | Proprietary format |
|---|---|---|
| AI readability | Native. In LLM training data. | Requires custom connector |
| Tool portability | Import/export across any BPMN tool | Locked to one vendor |
| Process execution | Executable on standards-compliant engines | Vendor-specific only |
| Validation | Structural validators available | No external validation |
| Longevity | ISO standard, 20+ year track record | Vendor-dependent |
What AI still cannot do with BPMN
BPMN being natively readable does not mean AI is perfect at working with it. There are real limitations to be aware of:
- -Structural errors in generated BPMN. AI often produces XML where gateways are not properly merged, sequence flows are missing, or events are incorrectly typed. The output needs human validation.
- -No organizational context. AI does not know the politics, culture, or implicit knowledge behind a process. It can read the structure but not the reasons behind design decisions.
- -Cannot facilitate workshops. The hardest part of process modeling is getting stakeholders to agree. AI cannot run a workshop, mediate disagreements, or build consensus.
- -Cannot guarantee compliance. AI can flag potential issues if you provide the rules, but it cannot independently verify regulatory compliance without domain-specific knowledge.
Choosing a process tool for the AI era
If you are evaluating process modeling tools and plan to use AI with your processes, ask three questions:
Does it export standard BPMN 2.0 XML?
If the answer is no, every AI integration will require a custom connector. If the answer is yes, any LLM can read your processes directly.
Can I use my processes with any AI model?
Open standards work with any model. Proprietary formats lock you into whatever AI integration your vendor decides to build.
Will this format be readable in 10 years?
ISO standards have staying power. Proprietary formats depend on a single company continuing to support them.
Most established BPMN tools export standard BPMN 2.0 XML: Camunda (developer-focused, execution engine), SAP Signavio (enterprise governance), ARIS (enterprise architecture), bpmn.io (open-source, embeddable), and Crismo (collaborative modeling with built-in AI discovery). The BPMN export gives you native AI readability across all of them. The differences are in what you do after the export: execution, governance, collaboration, or AI-assisted modeling. Not sure which fits your needs? Take our process tool quiz for a personalized recommendation.