Last time we talked about the danger of anthropormophizing AI because it misunderstands what various AI techniques can do - and we looked at that through the lens of generative AI and LLMs because that is the current public fixation.
I've tried a few different LLMs for generating BPMN diagrams, with mixed results. Mostly, they haven't been able to generate a valid file (but I expect this will change). Sometimes, they've generated simple and sensible diagrams (Claude liked to do this in Mermaid).
And yet, I'm not so sure about the future of doing this, from a people perspective. It's certainly going to be faster, initially, than training people in writing BPMN, but I don't know that there's a sufficient corpus for having an LLM produce *good* diagrams that help people think about process and improve it.
I think we'll need a bit of help from firms like Camunda - more specialized in BPMN, and even their demo version worked quite well at the conference for the samples I gave it, which weren't rigged.
I think a. generic LLM it is too hard to constrain the output and of course it has no notion at all of correctness. A specialist firm can focus on addressing these concerns.
Regarding the future - we'll see. You raise valid concerns. You might be surprised how many good BPMN models there are out there- though like code, these good examples are vastly outnumbered by really bad examples. We once had a process-analysis product (much like a code analysis product) and some of the scores it produced were shockingly bad (the models were shockingly bad as well). we had to change the interface to accommodate such scores that we never anticipated them.
It's an interesting space, for sure. I don't doubt that some companies have a really good set of process diagrams. I'm less sure they share them publicly. (From the business audits I've participated in, diagrams make auditors very happy, particularly if they're executable diagrams.)
I think in the case of BPMN, it is also of interest to generate readable BPMN, which is not yet executable (as in , the implementation details aren't there). might be a good starting point for auditors or auditable processes.
I've tried a few different LLMs for generating BPMN diagrams, with mixed results. Mostly, they haven't been able to generate a valid file (but I expect this will change). Sometimes, they've generated simple and sensible diagrams (Claude liked to do this in Mermaid).
And yet, I'm not so sure about the future of doing this, from a people perspective. It's certainly going to be faster, initially, than training people in writing BPMN, but I don't know that there's a sufficient corpus for having an LLM produce *good* diagrams that help people think about process and improve it.
I think we'll need a bit of help from firms like Camunda - more specialized in BPMN, and even their demo version worked quite well at the conference for the samples I gave it, which weren't rigged.
I think a. generic LLM it is too hard to constrain the output and of course it has no notion at all of correctness. A specialist firm can focus on addressing these concerns.
Regarding the future - we'll see. You raise valid concerns. You might be surprised how many good BPMN models there are out there- though like code, these good examples are vastly outnumbered by really bad examples. We once had a process-analysis product (much like a code analysis product) and some of the scores it produced were shockingly bad (the models were shockingly bad as well). we had to change the interface to accommodate such scores that we never anticipated them.
It's an interesting space, for sure. I don't doubt that some companies have a really good set of process diagrams. I'm less sure they share them publicly. (From the business audits I've participated in, diagrams make auditors very happy, particularly if they're executable diagrams.)
I think in the case of BPMN, it is also of interest to generate readable BPMN, which is not yet executable (as in , the implementation details aren't there). might be a good starting point for auditors or auditable processes.