System Card: Celonis AI Annotation Builder

April 29, 2025

System Name: Celonis AI Annotation Builder

System Version: GA version, updated regularly

1. System Overview

Description: The AI Annotation Builder is a capability within the Celonis Platform. It provides a no-code environment allowing users to interpret and structure data by generating AI-powered annotations.
AI Integration: It utilizes a Generative AI model (LLM) to process selected data based on natural language instructions (prompts) provided by the user. The AI interprets the prompt's logic and criteria to classify, categorize, or enrich data records with meaningful labels (annotations) and provides reasoning for its decisions.
User Interaction: Users select data within the Celonis Process Intelligence Graph, write natural language prompts defining how the data should be interpreted or classified, configure the desired output format (e.g., categories), run the annotation process, and review the generated annotations along with the optional reasoning provided for each.

Model Name & Version: Utilizes a Generative AI (Large Language Model) capability integrated within the Celonis Platform.
- By default it uses these models. Other models (e.g. BYOM) can be configured by administrators
Impact of Model Output: The AI model's primary output is "AI Annotations" – structured labels, classifications (e.g., priority level, category, root cause), or recommendations – applied to individual data objects. It also provides optional explicit reasoning for each annotation. These annotations enrich the organization's data model within Celonis, making previously unstructured or complex data more valuable for analysis, reporting, and potentially triggering downstream automations (though it does not build the automations itself).

Data Inputs:
- User-selected data from the Celonis Process Intelligence Graph (e.g., free-text fields like emails, service tickets, descriptions).
- User-authored natural language prompts containing instructions, criteria, definitions, and potentially examples.
- User configuration settings defining the structure and possible values of the output annotations.
Data Processing: The integrated Generative AI model analyzes the selected input data for each record according to the logic and criteria defined in the user's prompt. It applies this logic to determine the appropriate annotation.
Data Outputs:
- AI Annotations: Structured data labels appended to the relevant records in the Celonis data model (e.g., a 'Priority' column added with values like 'High', 'Medium', 'Low').
- (optional) Reasoning: Text explaining the specific factors and prompt logic that led to each annotation for each individual record.

Level of Automation: Human-in-the-loop / Semi-automated. The user defines the task, provides the intelligence via the prompt, initiates the process, and validates the output. The AI executes the classification/annotation task based on these explicit instructions.
Human Intervention Points:
- Data Selection: User decides which data to annotate.
- Prompt Engineering: User writes, tests, and refines the natural language instructions that dictate the AI's logic. This is the primary control point.
- Configuration: User defines the output structure (e.g., categories).
- Validation & Testing: Users are expected to run tests, review the generated annotations and the accompanying reasoning for accuracy, relevance, and potential biases.
- Refinement: User modifies prompts or configurations based on validation results.
- Application: User decides how to leverage the enriched data (e.g., for analysis, dashboarding, or as input for separate automation workflows).
Monitoring & Evaluation: Primarily occurs during the setup and testing phase through user review of annotations and reasoning. Ongoing quality checks might involve periodic re-validation or monitoring the impact of annotations on downstream processes. The quality heavily relies on the clarity and accuracy of the user's prompt and the quality of the source data.

Data Security: Leverages the security framework of the Celonis Platform, including:
- Data encryption at rest and in transit.
- Role-based access controls, ensuring users only process data they are authorized to access within Celonis.
- Processing adheres to Celonis' data processing agreements and security standards.
System Reliability: Dependent on the reliability and uptime of the underlying Celonis cloud platform.

Fairness & Non-Discrimination: Bias can be introduced primarily through the user-defined prompt (if instructions contain biased logic) or the underlying source data being annotated. Careful crafting and review of prompts are essential to ensure they define fair criteria. The reasoning output can aid in auditing annotations for unintended consequences or biases originating from the prompt logic or data.
Transparency & Explainability: The system offers significant transparency:
- The core logic is user-defined via the natural language prompt.
- The system provides explicit reasoning for each annotation generated, linking it back to the input data and prompt criteria.
- This System Card describes the functionality. While the internal workings of the underlying GenAI model are complex, the application of the AI is made transparent through the user-controlled prompt and the reasoning output.