Researchers Develop AI-Powered Method for Business Process Redesign
- By John K. Waters
- 05/14/25
Researchers have developed a new method that allows users to redesign process workflows by simply chatting with an AI.
Dubbed "Conversational Process Model Redesign" (CPD), the technique combines natural language prompts with a library of process change "patterns," allowing users to tell an AI assistant things like "Add an approval step after task B" or "Merge these two review steps" — and have a fully updated process model returned, no BPMN expertise required.
The approach was detailed in a new academic paper and tested across major LLM platforms, including GPT-4o, Gemini 1.5 Pro, and Mistral.
"We're bringing low-code process design into the no-code, conversational era," the lead researcher told TechCrunch. "And we're doing it without compromising on structural correctness."
Why It Matters
Business process modelling (BPM) is core to how organizations optimize workflows — think onboarding flows, approval chains, or customer support handoffs. But most BPM tools still require expertise in complex diagramming languages like BPMN, creating a bottleneck between domain experts and the technical teams who build the models.
Enter CPD, a technique that treats natural language as the input, and structural model updates as the output. Under the hood, the system uses a multi-stage pipeline: first classifying the user's intent, then mapping it to one of 18 process change patterns, and finally generating a new process diagram.
In pilot tests with 64 users, the researchers found that eight patterns (including both standard and newly proposed ones) were executed successfully over 30% of the time. While not perfect, it's a strong proof of concept for using LLMs as interactive workflow designers.
AI + BPM: A Quiet Revolution
The study also found some surprising emergent behavior. Users often described desired changes in their own terms — using phrases like "summarize," "split into steps," or "reverse last change" — rather than the formal terms expected by BPM software.
This prompted the researchers to propose five new change patterns tailored to conversational input, along with a "rename node" primitive that surprisingly turned out to be one of the most common (yet previously unsupported) actions.
"We didn't just adapt process patterns to LLMs — we learned from user behavior what was missing in the first place," wrote one co-author.
Where It's Going
While the results are promising, the researchers acknowledge that both LLM limitations and vague user input still pose challenges. The team is now exploring hybrid architectures, where the LLM handles language interpretation but hands off execution to deterministic process engines.
They're also eyeing features like auto-suggestions, intent clarification, and pattern recommenders — essentially bringing the UX of tools like GitHub Copilot to BPM platforms.
The potential payoff? Business users who can iterate on and optimize their own workflows in minutes — without submitting a ticket to IT.
Read the full report here.
About the Author
John K. Waters is the editor in chief of a number of Converge360.com sites, with a focus on high-end development, AI and future tech. He's been writing about cutting-edge technologies and culture of Silicon Valley for more than two decades, and he's written more than a dozen books. He also co-scripted the documentary film Silicon Valley: A 100 Year Renaissance, which aired on PBS. He can be reached at [email protected].