AI-generated BPMN workflows

Asking an LLM to write raw BPMN 2.0 XML is a bad time — a hallucinated element ID or a mismatched sequence flow reference is easy to produce and hard for the model to catch. @bpmnkit/core's fluent builder gives the model a much smaller, type-checked surface to work with instead: a chain of method calls that can only construct valid BPMN.

How it works

An LLM (or a deterministic planning step) produces a compact intermediate description of the process — which tasks, which gateways, which branches — small enough to fit in a single prompt. That description is compiled into a real ProcessBuilder chain:

import { Bpmn } from "@bpmnkit/core";

const xml = Bpmn.export(
  Bpmn.createProcess("support-ticket")
    .startEvent("start", { name: "Ticket Created" })
    .exclusiveGateway("gw", { name: "Severity?" })
    .branch("high", (b) =>
      b.condition("= severity = \"high\"")
        .serviceTask("escalate", { taskType: "escalate-to-oncall" })
        .endEvent("end-escalated"),
    )
    .branch("normal", (b) =>
      b.defaultFlow()
        .userTask("triage", { name: "Triage Ticket" })
        .endEvent("end-triaged"),
    )
    .withAutoLayout()
    .build()
);

Because every call is type-checked, a model that tries to reference a task that doesn't exist, or attach a gateway with no branches, fails at generation time rather than producing silently-broken XML.

Where this fits

This is the same approach used for AI-assisted design inside BPMN Kit's own editor and CLI (casen) — see the AI guide in the docs for the full pipeline, including the plan format and deploy-readiness linting that runs before a generated process reaches Camunda 8.

Frequently asked