AI workflows - Camunda: BPMN as a tool for creating & orchestrating AI workflows

This is a guest post for the Computer Weekly Developer Network written by Daniel Meyer in his capacity as CTO of Camunda.

Camunda is known for its open source platform that automates business processes by helping users model, execute and monitor them.

Meyer writes as follows…

The rise of agentic AI has amplified the need for organisations to manage their AI workflows effectively.

Agentic AI adds a layer of intelligent decision-making and autonomy, increasing the scope and efficiency of automated processes. In an ideal scenario, AI agents can make execution decisions within the context of pre-defined workflows. They also enrich process intelligence through AI-driven insights and automation.

Yet these benefits can only be fully realised when underpinned by strong process orchestration.

Without it, scaling AI investments and operationalising them across an organisation becomes difficult. Today, large organisations already manage an average of about 50 endpoints to execute tasks that are part of a process in their business.

This will only increase as AI adoption accelerates.

Process orchestration is key to managing this complexity by connecting and coordinating all the people, systems (including AI agents) and devices within a business process.

Visualising a chain of thought

Despite the hype, many people remain cautious about introducing AI agents into their organisation’s workflows. At a fundamental level, AI agents require three key core elements. 

  1. A clear purpose and set of rules to govern how they operate.
  2. A prompt to initiate action.
  3. Access to the right systems and tools to fulfil the request.

Many builders create AI agents as static, synchronous systems that function effectively as a ‘black box’. The drawback is that this approach offers little visibility into how agents perform, is difficult to scale, and can be hard to maintain. In the long run, it reduces AI workflow agility and weakens governance.

Understanding how an AI agent performed on a task means reviewing its reasoning process in the “best-case scenario”, as this typically shows which tools it chose to use and why. However, it could lead to you having to investigate the output logs of the various systems it may have triggered.

In practice, however, it’s impractical and time-consuming: the outputs are text files that must be interpreted, offering limited visibility into what actually happened. Worse still, there’s no guarantee the agent’s reasoning is fully accurate!

Enter BPMN…

This is why process orchestration and BPMN (Business Process Model and Notation) should be part of an organisation’s AI workflow strategy.

BPMN has been around since the early 2000s and is the standard way for building and visualising executable business processes. In the context of AI agents, a BPMN approach means applying that same structured, visual way of mapping workflows to AI logic. Instead of hard coding the tools or actions inside the agent, you model them as tasks or sequence of tasks within a process, that an agent can access and trigger as needed.

BPMN provides a visual and auditable orchestration layer that governs how and where AI agents operate within a process. It also ensures that AI-driven decisions remain transparent and controllable, helping teams blend traditional automation with AI in a governed, understandable way.

By applying a BPMN approach to agent design, organisations can separate agent logic from the tools it uses.

This greatly improves the maintainability of AI workflows. Process designers can add or remove tools without needing to rebuild or update the agent itself, while IT teams and developers can focus on configuring and integrating the right tooling. The result is far greater agility, allowing organisations to adapt their services and systems over time. For example, if a developer wants to test a new LLM, they no longer need to create a new agent from scratch. They can simply swap out the model and keep everything else intact.

AI workflows will only grow more complex as we move into an era of machine-to-machine communication. Initiatives like Anthropic’s Model Context Protocol (MCP) and Google’s Agent2Agent (A2A) are already working to standardise how agents interact across different platforms.

To keep pace with this shift towards more composable, cross-platform AI workflows, organisations cannot afford to ignore BPMN – it will be essential for maintaining agility and control.