Skip to main content

System Card: Celonis Integrated AI Assistants

April 29, 2025

System Name: Celonis Integrated AI Assistants (including Dashboard Assistant, Component Configuration Assistant, ETL Assistant)

System Version: These are integrated features within the Celonis platform, potentially in different release stages (e.g., preview, development, as indicated in source documents). Refer to specific Celonis release notes for current status.

1. System Overview

  • Description: Celonis Integrated AI Assistants are a suite of AI-powered tools embedded within various Celonis interfaces (like Studio Views and the SQL Data Integration interface). They are designed to augment user workflows by providing contextual help, automating certain tasks, and simplifying complex operations. This card covers assistants focused on dashboard summarization, component configuration aid, and SQL query generation/editing.

  • AI Integration: Each assistant utilizes AI models (often Large Language Models - LLMs) tailored to its specific task. The AI interprets user context, natural language inputs, or existing configurations/code to generate relevant outputs such as summaries, configuration suggestions, or SQL queries.

  • User Interaction: Interaction varies by assistant:

    • Dashboard Summarizer Assistant: Users viewing a Celonis Studio View may see an automatically generated summary of the dashboard's content and purpose, aiding quick understanding. Interaction is primarily consuming the AI-generated summary.

    • Component Configuration Assistant: While configuring components within Studio Views, users receive AI-powered assistance. This aims to help less proficient users by potentially suggesting configurations or interpreting natural language requests related to component setup.

    • ETL Assistant for Extractions: Currently available for JDBC extractions, within the SQL editor interface for data extraction, users can interact with this AI assistant. They can provide natural language commands to generate SQL queries, or receive AI suggestions for correcting errors in existing queries.

2. AI Model Dependency

  • Model Name & Version: Utilizes various AI models, including LLMs, integrated within the Celonis platform.

    • TBD

  • Impact of Model Output: The AI outputs directly assist users in their tasks:

    • Dashboard Summarizer Assistant: Generates summaries to accelerate dashboard onboarding and comprehension for business users.

    • Component Configuration Assistant: Provides configuration aid, aiming to make View component setup easier and faster, especially for newer users.

    • ETL Assistant for Extractions: Generates or corrects SQL extraction queries based on natural language, aiming to increase productivity, lower the technical barrier for data extraction, and improve query quality through validation.

3. Data Flow

  • Dashboard Assistant:

    • Inputs: Data samples, metadata and structure of the currently viewed Celonis Studio View/Dashboard.

    • Processing: AI analyzes the dashboard's components and layout to generate a concise description.

    • Outputs: A natural language summary displayed within the View interface.

  • Component Configuration Assistant:

    • Inputs: User's current context within the Studio View editor, selected component, data samples, and potentially user's natural language instructions regarding configuration.

    • Processing: AI interprets the context and user intent to suggest relevant component configurations.

    • Outputs: Suggestions or actions within the component configuration panel in Studio.

  • ETL Assistant for Extractions:

    • Inputs: User's natural language request for a SQL query, or an existing SQL query for analysis/correction. Context regarding the relevant database schema tables.

    • Processing: AI interprets the natural language or analyzes the existing SQL to generate/correct the query. A crucial step involves validating the generated/modified SQL using both AI checks and predefined rules for safety and correctness.

    • Outputs: A proposed SQL query within the editor. If validation fails or errors are detected, outputs include suggestions for correction.

4. Human Oversight & Control

  • Level of Automation: Primarily Human-in-the-loop / Augmentation. The assistants support the user rather than acting autonomously.

  • Human Intervention Points:

    • Users initiate the context (viewing a dashboard, editing a component, opening the SQL editor).

    • Users provide explicit input for the Component Configuration and SQL Assistants (e.g., natural language prompts).

    • Users review all outputs: dashboard summaries for accuracy, configuration suggestions before applying them, and SQL queries before execution.

    • Users make the final decision: accepting configurations, running generated SQL queries, acting based on dashboard summaries.

    • For the ETL Assistant for Extractions, the user reviews AI correction suggestions and manually applies fixes. The validation step provides a safety check, but user verification before running any query is critical.

  • Monitoring & Evaluation: Primarily through user experience – whether the assistance is helpful, accurate, and efficient. For the ETL Assistant for Extractions, success is also measured by the validity and effectiveness of the generated queries and the usefulness of correction suggestions. Feedback mechanisms within the Celonis platform may be used.

5. Safety & Security

  • Data Security: Relies on the security infrastructure of the Celonis platform, including data encryption and Role-Based Access Controls. Assistants operate strictly within the data permissions scope of the logged-in user.

  • System Reliability: Dependent on the reliability and uptime of the underlying Celonis platform. For the SQL Assistant, the integrated AI and rule-based query validation serves as an explicit safety measure to prevent execution of potentially harmful or syntactically incorrect queries generated by the AI.

6. Ethical Considerations

  • Fairness & Non-Discrimination: Risks are generally lower than in systems making decisions about people but require attention. Inaccurate dashboard summaries or SQL queries could lead to flawed business understanding or actions. Biased outputs are possible if the AI's training data contained biases or if user prompts inadvertently introduce them, but the primary focus is on functional correctness and safety. User review is key mitigation.

  • Transparency & Explainability:

    • Dashboard Summarizer / Component Assistants: Transparency comes from knowing the output is AI-generated. Explainability of why specific phrasing or configuration was chosen might be limited (common LLM challenge).

    • ETL Assistant for Extractions: Offers higher transparency. Users provide prompts and see the generated SQL. The validation step is transparent. Crucially, if queries are invalid or need correction, the AI provides suggestions, offering a degree of explainability for identified issues.