How to Create High-Quality
BPMN Models with AI

Artificial intelligence is becoming an increasingly important tool for business analysts. Today, AI can help analyze regulatory documents, structure requirements, prepare process descriptions, and even create BPMN diagrams. However, process modeling requires more than simply generating a visually appealing diagram. A BPMN model must accurately reflect how a business operates, comply with the BPMN standard, and follow the organization's internal modeling conventions.

To understand how modern AI tools can help solve this challenge, let's first look at how large language models work and then explore a practical approach to BPMN modeling using MCP and StormBPMN.

How AI Processes Information

Today, artificial intelligence is most commonly associated with Large Language Models (LLMs). These models are trained on massive collections of text and learn patterns and relationships between words, phrases, and concepts. When a user asks a question, the model does not retrieve a prewritten answer from a database — it predicts the most likely continuation of the text based on the provided context.

The way an LLM works is similar to a highly advanced version of predictive text. The difference is scale: instead of predicting the next word, the model leverages a vast amount of knowledge and context. It also works with tokens rather than individual words or letters — semantic units of text. This is why LLMs excel at understanding and generating content but may occasionally struggle with simple tasks such as counting letters in a word.

To process information, a model must convert it into numerical form. This process is known as vectorization and typically works as follows:

1. A document is split into smaller semantic fragments called chunks.
2. Each chunk is converted into a numerical representation called an embedding.
3. The resulting embeddings are stored in a vector database.

When a user submits a query, the query is also converted into an embedding. The system then searches the vector database for the most semantically relevant fragments and provides them to the model as additional context.

This capability is particularly important for business applications, where analysts work with policies, procedures, regulations, and internal standards rather than relying on general information from the internet.

How to Improve AI Output Quality

One of the best-known limitations of language models is hallucination — situations where the AI confidently provides incorrect information. This happens because the model is always trying to produce an answer. If it lacks sufficient data, it will still attempt to generate the most probable response, which can sometimes lead to errors.

Hallucinations do not mean that the model is poorly designed. They are a natural consequence of how language models operate. Using AI effectively therefore requires not only generating results but also implementing quality control mechanisms.

In practice, AI output quality depends on three key factors:

Another important factor is the context window — the amount of information a model can retain within a conversation. If too much information is provided at once, some details may be lost. For this reason, large documents and complex processes are often best handled in stages.

From Chatbots to Agents and MCP

Early language models could only answer questions. The next stage of development introduced Function Calling — a mechanism that allows external systems to provide models with actions they can perform, such as retrieving database records, opening documents, or creating objects in an application.

This capability led to the emergence of AI agents. Unlike traditional chatbots, agents can create action plans, use tools, evaluate intermediate results, and adjust their behavior when necessary. This is particularly valuable for business process modeling, which involves not only generating text but also working with reference data, documents, modeling standards, and enterprise architecture assets.

As integrations became more common, the industry needed a standardized way to connect language models with external systems. To address this challenge, Anthropic introduced the Model Context Protocol (MCP), which provides a unified interface for integrating tools and information systems with AI models.

Using AI for BPMN Process Modeling

Modern language models can generate BPMN diagrams from textual process descriptions. However, the limitations of a standalone chatbot become apparent very quickly. The model has no knowledge of corporate reference data, cannot access existing roles or documents, does not understand internal modeling conventions, and cannot validate the quality of the resulting process model.

As a result, analysts often receive a draft diagram that requires significant manual rework. The solution is to integrate the language model directly with a process modeling platform.

For example, StormBPMN addresses this challenge through its own MCP server. It allows organizations to connect both cloud-based and local language models and use them to create BPMN diagrams directly within the modeling environment.

The key difference from a traditional chatbot is that the model operates within a corporate context rather than in isolation. It can:

• use existing roles and documents from reference libraries;
• incorporate enterprise architecture assets;
• follow company-specific modeling standards and conventions;
• create complete BPMN models rather than simple XML files or diagram images.

Virtually any type of textual information can serve as input for process modeling:

• regulatory documents, policies and procedures, laws and industry standards;
• PDF and Word files;
• meeting and interview transcripts;
• business chats and correspondence;
• data from other information systems.

If a document is particularly large or contains a lot of secondary information, it is often helpful to ask the model to produce a structured summary before generating the BPMN model.

Automated BPMN Model Quality Assurance

Even when a model has access to complete context, quality control remains essential. StormBPMN includes an automated BPMN quality assessment mechanism that functions similarly to a software development linter. Organizations can define their own validation rules, including:

• which BPMN elements are allowed;
• which constructs are considered errors;
• how swimlanes should be used;
• what structural requirements a process model must satisfy.

After generation, the BPMN model is automatically evaluated against these rules. If the score falls below a predefined threshold, the model receives feedback about the detected issues and can revise the process before running the validation again.

This creates a multi-layered quality assurance system:

Such an approach helps eliminate many common errors before human review even begins.

Practical Example

Consider a typical scenario: an analyst needs to create a BPMN model based on a regulatory document consisting of several dozen pages, such as a public procurement law.

Traditionally, the analyst would first study the document, identify roles, documents, events, and tasks, and only then begin building the diagram. With AI assistance, much of this work can be automated.

The model can:

• identify process participants;
• determine input and output documents;
• locate decision points and process stages;
• generate the process structure;
• link process elements to specific sections of the source document.

Collaboration and Access Management

In enterprise environments, access control is critical. An MCP agent inherits the permissions of the user on whose behalf it operates. If a user is allowed to create documents or roles within the system, the agent can perform those actions as well. If the user's permissions are restricted, the agent is restricted in exactly the same way.

This approach enables organizations to adopt AI while preserving their existing security model and governance framework.

For organizations with strict data protection requirements, local language models can also be deployed. In this configuration, all information remains within the corporate environment and is never transmitted to external services. Although local models may not always match the performance of leading cloud-based systems, they are often more than capable of handling business process modeling tasks.

Best Practices

To achieve the best possible BPMN modeling results with AI, consider the following recommendations.

Conclusion

Artificial intelligence can already dramatically accelerate BPMN modeling. However, the greatest benefits come not from the language model itself, but from its integration with corporate data, reference libraries, and quality assurance mechanisms.

The combination of LLMs, RAG, MCP integration, and automated validation transforms AI from a text-generation tool into a practical assistant for business analysts.