BPMN is dead, long live BPMN

We've all seen how tools like ChatGPT can handle a variety of business tasks, automating nearly everything. And it’s true — generative AI really can do a wide range of tasks that humans do currently. So, why bother designing and running your business processes with automation standards that were defined 20 years ago in a different technology era? Why not let your business users work directly with AI to do it all?

The short answer? Because those standards are still the strongest, most efficient ways to orchestrate AI agents, maintain governance, and execute at enterprise scale. 

In this article, we'll show you how BPMN provides the common language agentic AI needs to communicate, why specialized engines execute processes more efficiently than AI can, and how enterprises are already running millions of AI-orchestrated workflows at scale.

Key takeaways

• BPMN serves as the lingua franca for agentic AI, enabling AI agents to communicate about processes unambiguously while remaining explainable to humans for governance and oversight, just as it previously aligned human stakeholders.

• AI agents can define and orchestrate BPMN workflows, but organizations should run them on specialized BPMN engines that are faster, more efficient, and proven at enterprise scale.

• Platforms like Flowable support multi-agent orchestration through utility agents, document agents, knowledge agents, external AI agent integrations, and orchestrator agents working together through standard BPMN constructs with vendor-agnostic integration.

• Case Management Model and Notation handles unpredictable, adaptive scenarios better than rigid sequences by providing sentries and stages that let AI agents activate conditionally based on context rather than following predetermined paths.

What is BPMN?

The Business Process Model and Notation (BPMN) was introduced in 2004 as a standardized language for modeling business processes that all stakeholders can understand — from business users to developers to the systems that execute those processes.

Why was this so important? Because each group involved in defining the processes had different needs and expectations. Relying solely on a business user's description of a task or activity wouldn't address the broader needs of the organization.

For example, a sales team might design a customer onboarding process that seems simple: collect information, set up an account, and grant access. But without IT's security protocols, finance's credit checks, and compliance's data handling requirements, that “simple” process creates serious operational and legal risks.

BPMN is a simple way to ensure that all perspectives are considered and aligned when defining processes that cross departmental boundaries.

BPMN in the age of agentic AI

Just like BPMN was required to create alignment among different stakeholders, we are now finding that agentic AI needs that same solution.

We're now seeing systems capable of reasoning over multiple steps and interacting with a network of AI "agents." These advanced systems, known as agentic AI or large action models (LAMs), use specialized AI agents trained for different business perspectives, like HR, privacy, compliance, and IT, that could eventually replace the human roles involved in building business processes.

But how are these AI agents going to interact unambiguously?

The answer is BPMN because, in actuality, there's little difference between humans and AI agents needing to communicate about processes unambiguously. Plus, crucially, BPMN remains explainable to non-technical humans, so people can validate what the AI agents are doing — not just execute black-box workflows. The way forward, then, is to make your AI agents generate explainable BPMN definitions of your business processes, possibly with external intelligent services contributing using the same common representation.

Orchestration patterns: How BPMN manages multi-agent AI workflows in practice

So how does this actually work? In Flowable's implementation, AI agents are defined using agent models: standardized abstractions that represent anything from a simple single-call to a large language model to a complex, long-lived agent participating across multiple workflow steps. These agent models integrate into BPMN and Case Management Model and Notation (CMMN) using dedicated agent tasks, following the same notational standards already familiar to process designers.

The Flowable platform supports six types of agents that handle different orchestration needs:

• Utility agents execute simple LLM prompts with defined input and output, either structured or unstructured.

• Document agents classify documents and extract structured data, often used as part of an orchestrator agent.

• Knowledge agents use knowledge base models to retrieve contextual information at runtime.

• Orchestrator agents coordinate multiple agents and external APIs, typically invoked from a CMMN case model.

• External agents connect to third-party agents such as Salesforce Agentforce or Azure AI Foundry.

• A2A agents connect to third-party agents using the Agent2Agent (A2A) protocol.

For multi-agent collaboration, BPMN handles straightforward linear workflows where the sequence is known upfront. Multiple agents can work together across internally configured and external systems, all orchestrated through standard BPMN constructs. The key is that these agents remain vendor‑agnostic; organizations can integrate with AWS Bedrock, Azure AI Foundry, Salesforce Agentforce, or other AI services without vendor lock-in. Flowable supports multiple model providers and external agent platforms (for example, Salesforce Agentforce and Azure‑based agents) via external and A2A agents, rather than tying you to a single vendor ecosystem.

What makes this approach powerful is that all agent invocations are logged and auditable, with behavior governed by the same process and case models that control other automated steps. The orchestration is explicit, visible in the model, and follows the governance patterns enterprises already use for mission‑critical workflows.

This approach isn't just theoretical. All the way back in 2017, we demonstrated an example of AI trained on decisions in a process that gets dynamically converted to an explainable, executable, standard Decision Model and Notation (DMN) — proving that AI and automation standards have been working together long before agentic AI became mainstream.

Why not just use AI to execute BPMN?

You could use AI to execute BPMN, but specialized BPMN engines are faster, more efficient, and far more cost-effective than using AI for execution.

Once you have your AI-generated BPMN, you could pass it to another AI agent that's been trained on how to execute BPMN, calling it each time a new instance of the process is needed. Technically, this works.

But it's a waste of resources. Both training and running AI agents are computationally expensive and relatively slow. When tried and tested, hyper-efficient, and highly scalable BPMN engines already exist, using AI for execution makes no sense.

It's far more efficient, consistent, and proven to run your BPMN on these specialized engines. Today's BPMN systems should become the process runtimes for AI — handling execution while AI focuses on generation and orchestration. The same applies to the other key business automation standards: CMMN and DMN.

Flowable customers run millions of BPMN and CMMN instances each day using these specialized engines, which is just further proof that this approach works at enterprise scale.

When CMMN outperforms BPMN for AI orchestration

While BPMN excels at linear workflows, complex AI scenarios often need something more adaptive. That's where CMMN is the better choice.

BPMN works beautifully when you know the sequence upfront (Step A leads to Step B, which leads to Step C). But real business scenarios aren't always that predictable. Case management, as provided by CMMN, allows a way of expressing rich business automation that isn't just about going through a sequence of steps to get to an endpoint. It allows a way of defining overall, end-to-end automation, with a 360° view of what may be appropriate at any point in the lifecycle of something — whether it’s a person, a situation, a document, a project, or anything around which automation can be envisaged.

CMMN has the concept of sentries (triggers) and stages (contexts), among other things. You can use contexts to define when certain AI agents are (or are not) appropriate, then have them activate given the right trigger. Depending on the results of those agents, additional AI agents or BPMN processes might then be triggered. All the while, human interaction or intervention can be included as part of the overall intelligent business automation.

The Flowable Platform is purpose-built for this kind of complex enterprise automation. It integrates BPMN, CMMN, and DMN engines alongside AI orchestration capabilities, giving you the complete toolkit to build enterprise-grade intelligent automation. Whether you need linear workflows, dynamic case management, or complex multi-agent orchestration, it's all available in one governed, auditable platform.

Ready to see how BPMN and agentic AI work together? Try Flowable for free and start building AI-powered processes with the same standards enterprises trust for mission-critical workflows.

BPMN and agentic AI FAQs

Why do AI agents still need BPMN if they can reason and plan autonomously?

AI agents need to communicate with each other unambiguously about processes, just like human stakeholders do. BPMN provides a common language while keeping workflows explainable and auditable.

What makes BPMN engines more efficient than having AI agents execute processes directly?

AI execution is computationally expensive and slow. Specialized BPMN engines are hyper-efficient and proven at enterprise scale, and handle millions of process instances daily, making them far more cost-effective for execution.

What role does CMMN play in managing AI agents compared to traditional BPMN workflows?

BPMN handles linear workflows where the sequence is known upfront. CMMN uses sentries (triggers) and stages (contexts) to manage adaptive scenarios where AI agents activate conditionally based on evolving context, making them better for unpredictable, dynamic situations.