Then agentic AI happened, and something unexpected followed. The teams shipping AI into operations discovered that they needed all the things BPM people had been doing for twenty years.
-
They needed a way to model a process so that business and engineering could read the same thing.
BPMN. -
They needed durable execution that survived restarts, retries and the occasional model timeout.
Workflow engines. -
They needed explicit decision logic that wasn’t buried in a prompt.
DMN and rules engines, or their modern equivalents. -
They needed escalation paths and human-in-the-loop checkpoints - not as afterthoughts but as first-class citizens of the orchestration.
Same as ever. -
They needed audit trails that explained what happened and why, in terms the regulator and the business could both understand.
Same as ever.
The story we’re now telling clients is that BPM is the substrate, and AI is the new capability that sits inside the workflow. An agent that classifies a document is a step in a process. An agent that drafts a response is a step in a process. An agent that decides whether to settle or refer is a step in a process, and like any step that matters, it has inputs, outputs, success criteria and exception paths.
Treat the agent as a software component that participates in a workflow, not as a magic black box that replaces it, and three things become easier. You can measure quality at the step level rather than at the system level. You can swap models and providers without rewriting the orchestration. And the operations team, the people who actually have to trust this thing every day, can read what’s going on.
If BPM has felt unfashionable to you, this is a good moment to revisit it. It’s the discipline that lets agentic AI be useful in regulated operations. You don’t need to call it BPM. But you need to do it.