Introducing ARIS Script for Process Improvement with AI (OpenRouter Claude/ChatGPT)

Hi Tim, 

First of all, that is absolutely outstanding job! Thank you!

I'm playing all day long )) especially cool when you use the same model multiple times, and it becomes better and better )). In my case, it did not want to work on SR24, but on SR27, it is okay.

Would you please explain why this LLM (Claude 3 Haiku) was selected, if you can share? 

As a small idea, I'd not change objects directly in DB (users could cause the chaos very quickly), but instead generate an Excel:
object type | current name | AI suggested name | GUID
and then import back the changed data with the second script. It is safer.

But anyway, the script is an atomic bomb!
 

Hi Alexander,
Great to hear that someone finds it useful! I’ve only tested it on SR 27 so far. Initially, I also thought about using Excel as an additional layer, so I really like your approach. In the end, though, I aimed to automate as much as possible.

As for why I used Claude 3 Haiku—I simply asked the LLM to choose a model 🙂. From my experience, Claude handles BPMN quite well.

Regarding costs: my total API key budget was $1.00. I used only $0.006 so far, with $0.994 remaining. The cost per process model is approximately $0.000253 (depending on size and tokens), which seems quite reasonable.

It could also be interesting to offer the LLM-improved terms as a separate language option 🙂.
I still have some ideas I’d like to explore—like integrating direct translation support or enabling uploads of PDFs or JPGs of the model so the LLM can analyze or summarize the process visually. 

PS: I would be nice if you could share your Excel version!

Hi Tim,
Deal ) when I have an Excel one, I'll share )
I'm not sure regarding translation. I've been involved in the translation project, and I can see the "product" produced by different LLMs, and the quality of the translation is currently inferior. 

But with graphics, that would be an interesting idea! It is necessary to test with Claud, but ChatGPT understands it very well, so for e.g. process description, it should work.

Tim

Do you think this would work with Microsoft Open AI - already existing API token in our company?  Believe Open AI can access Claude LLM's. 

Adapted your script @Tim for chatgpt but it does not seem to even start..... has anyone got the script to work?

Yeah, it works with GPT

Just a note.
For now, we have found issues while playing with GPTs (from OpenAI, maybe others would behave differently).

The main issue, I'd say, is the limitation of the prompt window. Even an advanced model as 4o (128K tokens) cannot 'digest' the task with one diagram. Saying one diagram, I mean the following structure:

"steps": [
    {
      "R": "В1",
      "A": "К1",
      "C": ["1", "2"],
      "I": ["1", "2"],
      "Step": "Some action",
      "Input": ["1", "2"],
      "Output": ["Р1"],
      "Systems": ["HRM", "Excel"],
      "stepattributes": {
        "description_definition": "1",
        "avg_time": "1",
        "guid": "1291215b-3539-11f0-3e60-d25d3332e3d4",
        "processing_code": "01",
        "remark_example": "1"
      }
    }
  ],
  "instruction": "create step description"
} 

So, when the diagram is big (e.g., 20+ steps) or the attribute description is big (e.g., 250+), then it returns nothing and admits the context_length_exceeded. Yes, there is a way to chop the model into pieces and feed it with these chunks, but it is also a questionable, e.g., in this case, creating an overall process description will be impossible in this case. So, I'm not sure about expecting AI to process a complicated request, e.g., something related to E2E analysis. Well, I'm not sure for the moment. 

Anyway, the genie is out of the bottle, so it is only a matter of finding the balance between scripts, AI prompts, AI capabilities, and business outcomes.