AI and BPM

AI touches BPM in four ways. The first two help people do BPM; the last two put AI inside the processes.

1. LLMs for process work (discovery and analysis)

LLMs can turn text into structure: draft a process model from a procedure document or interview notes, summarise or explain a model, answer questions about a process, write test cases. Vidgof, Bachhofner and Mendling (BPM Forum 2023) surveyed opportunities across the BPM lifecycle and six research directions; a separate technical report examined LLMs for process-mining. Treat output as a first draft that needs human validation in the bpm-lifecycle discovery phase.

2. Process mining combined with AI

Mining gives AI real event data to learn from (prediction of delays, next-best action) and AI helps interpret mining results.

3. Agents executing or assisting process steps

An agent can take a step that rules cannot (read an email, classify a claim, draft a reply). The design question is how much freedom: a fully free agent is hard to audit; a fixed flow cannot adapt.

4. BPMN as the contract for agent workflows

Orchestration engines now use BPMN to bound agents. camunda documents an AI Agent connector that runs inside a BPMN ad-hoc sub-process: the LLM chooses the next tool, and the engine executes the matching BPMN activity, returns the result and repeats until the LLM gives a final answer. The engine keeps control of variables, retries, incidents and human tasks; MCP tools can be added through an MCP client connector. In this pattern:

Concern	Provided by
Which tool next	LLM
Executing the tool, retries, timeouts	Process engine
Approvals and escalation	User tasks / human-in-the-loop
Audit trail	Engine history
Rules that must be exact	DMN decision tables

This is the BPM answer to “agents are unpredictable”: a deterministic skeleton with agentic freedom in explicit, logged compartments. uipath positions a similar idea with its Maestro orchestration. For agent frameworks see agent-orchestration-overview and ai-agentic-systems; for the automation tools see ai-automation.

Risks

Hallucinated models, data leaving the company, hard-to-explain decisions, and agents given more authority than the process owner approved (process-architecture-and-governance).

Related

hyperautomation-and-intelligent-automation, rpa, BPM hub

Sources

• Vidgof, Bachhofner, Mendling, “Large Language Models for Business Process Management: Opportunities and Challenges”, arXiv 2304.04309: https://arxiv.org/abs/2304.04309
• Berti et al., “Leveraging Large Language Models (LLMs) for Process Mining (Technical Report)”: https://arxiv.org/abs/2307.12701
• Camunda docs, AI agents: https://docs.camunda.io/docs/next/components/agentic-orchestration/ai-agents/
• Camunda docs, AI Agent connector: https://docs.camunda.io/docs/components/connectors/out-of-the-box-connectors/agentic-ai-aiagent/