AutoBPMN.AI: Conversational Process Modeling and Automation

Abstract

We demonstrate the text-based modeling and redesign of processes in the process execution environment of the Cloud Process Execution Engine (CPEE) based on iterative conversations with an LLM such as gemini or gpt. The input can be text and an empty model as well as text and an existing model. The output is a newly created or redesigned process model in executable format, i.e., based on abstract syntax trees and displayed in BPMN-like fashion, exploiting the automatic layout capabilities of the CPEE. The tool is demonstrated for diferent scenarios such as the creation and redesign of the process models, as well as the subsequent process model execution.

1. Introduction

Provide sources of information

Create model T2M

Transfo+rmation ConMvoedrsealintiognal +

Redesign model

Provide changes

Export model for communication

Configure model

for execution moAdenlaclyrezeation ~

Over the last 20 years, significant eforts have been devoted to automatically derive business process models from natural language text [ 2 ]. However, the resulting models remain at a conceptual level [ 3, 4 ] (see Fig. 1, (1) and (2)). For their further utilization these models have to be transformed across languages in diferent layers, which can be expensive, error-prone, and dificult. Hence, modeling processes directly at the execution layer may be desirable, but requires to consider execution properties such as including data objects necessary for determining how often a loop should be executed and thinking about the functionality of certain tasks. As process modeling can be already a challenge for domain experts who are often non-technical users, modeling at execution level might be too complex.

Thus, in this paper, we present a tool, bridging conceptual and executable layers, which allows domain experts to create and redesign process models directly in an execution environment as shown in Fig. 1, (3). The UI relies on natural text input in conversations with a chatbot (conversational process modeling and redesign) [ 5, 6 ]. The input comprises process descriptions, guidelines, and legal documents. The output is a process model which has been refined by the domain experts and is per definition sound and executable from a control flow perspective. It can then be used in several ways, including augmentation with data flow and endpoints for execution, transformation back to text, and simulation and analysis. Additionally, by logging how the process model is changed and based on which interactions with the users these changes have been achieved it is possible to analyze the process of process modeling [ 7 ].

2. Conversational Modeling and Automation with CPEE

In this section, we discuss the main characteristics and innovations of the LLM-based CPEE extension, which utilizes the concept of conversational process modeling. Since the primary goal of conversational modeling is to create and improve process models and descriptions through iterative exchange of question and answers between domain experts and chatbot, we demonstrate its application with two scenarios: (1) process model creation and (2) process model redesign.

Scenario 1) Process Model Creation can be driven by the use of documents and textual sources containing information about the process—such as regulatory documents or summaries based on domain expert interviews—or it can start from scratch, without any supplementary materials. Thus, Process model creation can be represented by two use cases.

Use Case A) Create Process Model “Out of the Blue” creates a process model (c) for a specific use case without further input, cf. (a) text input and (b) selected LLM in Fig. 2. For more information about process model and possible steps after model generation see Conversational Modeling and Automation with CPEE Use Case B) (c) Created process model

Use Case B) Create Process Model Based on Text (T2M Transformation): Though Use Case A) might provide a starting point for some applications, a more guided creation of process models might be advisable in order to control the quality of the output [ 5 ]. One option is to work with more specified textual input such as process descriptions and/or regulatory documents. In Fig. 2 a gaming process description1 is inserted (a), processed by LLM gemini 2 Flash (b), and the corresponding process model is created (snippet shown in (c)).

Scenario 2) Redesign Process Model: Use cases for Scenario 1) have illustrated the one-shot creation of process models out of the blue or from process descriptions. However, as the quality of the output process model is of utmost importance for many real-world domains, the iterative refinement of the models through conversational process model redesign [ 8, 6 ] is crucial. In particular, this can foster the interaction of domain experts with the model. Conceptually, conversational redesign is based in exploiting the established process change patterns [ 9 ] in matching them with textual user redesign requests by the LLM in order to achieve sound process models [ 8, 6 ]. Consider the abstract example depicted in Fig. 3. The process model is created by textual redesign requests step by step: i) add task A; ii) add decision after task A with two branches, one branch contains task B, the other branch task C; iii) add task D after B and C; iv) delete task B. The resulting process models after each step i)–iv) are shown in Fig. 3. The process of redesigning the model is stored in the process execution log resembling the process of process modeling [ 7 ].

s l e d o m s s e c o r p d e t a e r C

Task A a2

Task A a2 exclusive Yes

Task B a5 No

Task C a6

Task A a2 exclusive Yes

Task B a5 No

Task C a6

Task D a7

Task A a2 exclusive No

Task C a6

Task D a7

resUitupn (i) add task A t(bairis)aknacdAhdwedsite;hcotinwsieoonbraafntecrh (ainiid)aCdd task D after tasks B contains task B, the other branch task C (iv) delete task B Conceptual Framework. The framework encompasses five steps ( a to e ) depicted in Fig. 4. To start with, a text describing how the model should look/change 1 , the Selected LLM 2 , and the Current Model 3 if it exists is sent to the LLM Service. Next, an appropriate scenario has to be identified a (i.e., either model creation or model redesign). Depending on the scenario, 3 is either converted to Mermaid2 representation b or this step is skipped. Then, 1 and, optionally, the result of b are inserted into the prompt c , and the LLM request is performed d (cf. [ 8, 10 ] for more information about prompt design, parameters, and examples). Finally, the output of d is transformed back into a CPEE-Tree 4 in e and passed back to the user interface for further usage/refinement. Possible Subsequent Steps. After having obtained the final model the following steps can be taken: • Retrieve items for documentation and communication to other participants of the process from models created in the scenarios described above (i.e., semantic/conceptual layer). The tool ofers two ways to do this: (i) obtain a visual representation of the model (“save svg graph” in “Instance” tab shown in Fig. 4) and (ii) transform the model to text (again, using LLM) using the “Save as Text” button (again, in “Instance” tab shown in Fig. 4). An example of (ii) is shown in Fig. 5a A .

• Make the process model ready for execution by adding information needed for execution (data objects, functionality and in-/output data objects for tasks, ...) directly to models created using the tool which only include control-flow but are already described in the form of a CPEE-Tree (which is a 1taken from https://zenodo.org/records/7783492 2https://mermaid.js.org/, accessed 2025-06-26

User Interface of AutoBPMN.AI

Current Model 3 / 4ProvNideeSwhoeMDeotaidlsel a2 After "Provide Shoe Details" add a new task saving shoe details into database.

1 User Input 4

2 Selected LLM Translate d to CPEE-Tree 4 e

Create Mermaid Using LLM in 2 d

LLM Service

Add 1 , b * to Prompt c

Verify Email

a3 ProvideSxhcoluesDievteails a2 VerifyEEmmaialil Exists a3 exclusive

Email Does Not Exist EmailExists EmailDPoreosvNidoteExEismtailAddress a6 ProvideSEamvaeilAEdmdraesilsto Da6tabase a7 SaveEmailtoDatabase a7

Send Confirmation Email a5 SendConfirmationEmail a5 1 2 3 Translate 3 to Mermaid b *

Identify Scenario based on 3 a Current Model (i.e., output of b ), Mermaid representation of New Model (i.e., output of c ), and New Model 4 (i.e., output from d ) in CPEE-Tree representation (see Fig. 4). A shortened log is shown in

3. Maturity

in CPEE in, e.g., [ 1 ].

Scientific maturity: The underlying concepts have been published in several outlets, i.e., text2model in [ 5 ], model2text in [ 11 ], conversational process modeling in [ 8, 6 ], and process automation and execution CPEE maturity: The underlying process cloud execution engine CPEE is an open source engine; the 3 engine and connected components have been downloaded 1.677.929 times so far . CPEE has been developed into orchestration framework centurio.work which implements and operatively runs 16 diferent processes at 7 manufacturing companies [ 12 ].

Maturity of Conversational Process Modeling: The concepts of text2model transformations have 4 been developed in the BPAIS project funded by SAP Signavio; text2process transformation has become 5 part of the modeling capabilities of the AI-assisted process modeler of SAP Signavio . Moreover, we have conducted two user studies: the first one was conducted to evaluate text2model concepts with 40 participants with diferent experience levels in graphical modeling, e.g., UML, ER, BPMN, and 3 4 https://rubygems.org/profiles/eTM, accessed 2025-06-25 https://www.cs.cit.tum.de/bpm/projects/bpais/, accessed 2025-06-25 https://news.sap.com/2025/03/sap-signavio-launches-ai-process-modeler-text-to-process/, accessed 2025-06-25 programming skills [ 13 ]. The second user study assesses the concepts of conversational process redesign with 64 participants. Finally, we are currently conducing a lab course at the Technical University of Munich on conversational process modeling and automation.

4. Videos and Tool

The tool is available at autobpmn.ai, where a new empty process instance can be created. For each instance the process model can be created/redesigned as described in Sect. 2. The user interface is realized by creating a new spin6 of a CPEE cockpit7 to create and redesign models using Conversational Modeling. The LLM Service which supports the Conversational Modeling outlined in blue in Fig. 4 is also available on github8. A video demonstrating the functionality of the tool is available at autobpmn.ai.

Declaration on Generative AI

The author(s) have not employed any Generative AI tools.

References

[1] S. Rinderle-Ma , J. Mangler , D. Ritter , Fundamentals of Information Systems Interoperability , Springer, 2024 .

[2] S. Schüler , S. Alpers , State of the art: Automatic generation of business process models , in: BPM 2023 International Workshops , Utrecht, The Netherlands, September 11-15 , 2023 , Revised Selected Papers, volume 492 of Lecture Notes in Business Information Processing , Springer, 2023 , pp. 161 - 173 .

[3] H. Kourani , A. Berti , D. Schuster , W. M. van der Aalst, Promoai: Process modeling with generative AI , in: Proceedings of the Thirty-ThirdInternational Joint Conference on Artificial Intelligence, IJCAI-2024, International Joint Conferences on Artificial Intelligence Organization , 2024 .

[4] J. Köpke , A. Safan , Introducing the bpmn-chatbot for eficient llm-based process modeling , in: BPM Dempos , 2024 , pp. 86 - 90 . URL: https://ceur-ws. org/ Vol- 3758 /paper-15.pdf.

[5] N. Klievtsova , J. Benzin , T. Kampik , J. Mangler , S. Rinderle-Ma, Conversational process modelling: State of the art, applications, and implications in practice , in: Business Process Management Forum , 2023 , pp. 319 - 336 . doi: 10 .1007/978-3- 031 -41623-1\_ 19 .

[6] N. Klievtsova , T. Kampik , J. Mangler , S. Rinderle-Ma, Conversational process model redesign , arXiv preprint arXiv:2505.05453 ( 2025 ).

[7] J. Claes , I. Vanderfeesten , J. Pinggera , H. A. Reijers , B. Weber , G. Poels , A visual analysis of the process of process modeling, Inf . Syst. E Bus . Manag. 13 ( 2015 ) 147 - 190 .

[8] N. Klievtsova , T. Kampik , J. Mangler , S. Rinderle-Ma, Conversationally actionable process model creation , in: Cooperative Information Systems , 2024 , pp. 39 - 55 .

[9] B. Weber , M. Reichert , S. Rinderle-Ma, Change patterns and change support features - enhancing lfexibility in process-aware information systems , Data Knowl. Eng . 66 ( 2008 ) 438 - 466 .

[10] N. Klievtsova , J. Mangler , T. Kampik , J. Benzin , S. Rinderle-Ma, How can generative AI empower domain experts in creating process models? , in: 19th International Conference on Wirtschaftsinformatik (WI 2024 ), Würzburg, Germany, September 16-19 , 2024 , Proceedings, AISeL, 2024 , p. 66 .

[11] N. Klievtsova , J. Mangler , T. Kampik , S. Rinderle-Ma, Utilizing process models in the requirements engineering process through model2text transformation , in: RE, 2024 , pp. 205 - 217 .

[12] S. Rinderle-Ma , J. Mangler, Process automation and process mining in manufacturing , in: Business Process Management, 2021 , pp. 3 - 14 .

[13] N. Klievtsova , J. Benzin , J. Mangler , T. Kampik , S. Rinderle-Ma, Process modeler vs. chatbot: Is generative AI taking over process modeling? , in: Process Mining Workshops , 2024 , pp. 637 - 649 .