I haven't look at Claude workflows, but I guess now we have the option of getting an LLM to make a judgement call whether to follow a workflow, jumping on another variant or escalating to a human, perhaps with a proposal on how to change the flow.
The « unforeseeable” part in a Business Process may not simply stay in the way a given step is executed. But in the very fact that the step is actually required and required at that point of the flow.
An experienced professional may decide that the flow should be A D B C. Instead of A B C D . And perhaps sometimes is should be C D A
If A, B, C and D are the leaves, how this is solved?
Great parallel with the process modeling languages and thanks for sharing your thoughts on the deterministic vs probabilistic portion of the execution.
One additional thought is the Claude approach allows you to also define the workflow using with human language by describing a use case. With traditional modeling languages you needed the understanding of a specific formalism to define the workflow, which prevented the business users from correctly using the tool. Claude lowers that barrier, but it also risks of having the deterministic part defined by the probabilistic LLM as for bespoke workflows, you prompt and Claude tried to build the steps.
I run the engineering department for a company selling a BPMN-based workflow system for 10 years and despite our promise to bring the business user onboard in the development process, we never really managed as the business preferred to delegate to the engineers the understanding of the model notation. What we learned through all these years was the notation was only part of the story. The best engineers were leading the most successful implementations because they could manage the inherent complexity of the problem at hand. The incidental complexity introduced by the modeling notation was already relatively low and Claude will further lower it by allowing using natural human language. What we don't yet know if how well will Claude manage the inherent complexity without the structured thinking of an engineer when dealing with complex problems.
I haven't look at Claude workflows, but I guess now we have the option of getting an LLM to make a judgement call whether to follow a workflow, jumping on another variant or escalating to a human, perhaps with a proposal on how to change the flow.
I am not sure i fully understand
The « unforeseeable” part in a Business Process may not simply stay in the way a given step is executed. But in the very fact that the step is actually required and required at that point of the flow.
An experienced professional may decide that the flow should be A D B C. Instead of A B C D . And perhaps sometimes is should be C D A
If A, B, C and D are the leaves, how this is solved?
Thanks
Great parallel with the process modeling languages and thanks for sharing your thoughts on the deterministic vs probabilistic portion of the execution.
One additional thought is the Claude approach allows you to also define the workflow using with human language by describing a use case. With traditional modeling languages you needed the understanding of a specific formalism to define the workflow, which prevented the business users from correctly using the tool. Claude lowers that barrier, but it also risks of having the deterministic part defined by the probabilistic LLM as for bespoke workflows, you prompt and Claude tried to build the steps.
I run the engineering department for a company selling a BPMN-based workflow system for 10 years and despite our promise to bring the business user onboard in the development process, we never really managed as the business preferred to delegate to the engineers the understanding of the model notation. What we learned through all these years was the notation was only part of the story. The best engineers were leading the most successful implementations because they could manage the inherent complexity of the problem at hand. The incidental complexity introduced by the modeling notation was already relatively low and Claude will further lower it by allowing using natural human language. What we don't yet know if how well will Claude manage the inherent complexity without the structured thinking of an engineer when dealing with complex problems.