System Card: Celonis Process Copilot
April 29, 2025
System Name: Celonis Process Copilots
System Version: GA version, updated regularly
1. System Overview
Description: Celonis Process Copilots are integrated features within the Celonis Platform. They function as Generative AI-powered chatbots designed to provide quick access to process insights and data through a conversational interface.
AI Integration: The core functionality of Celonis Process Copilots relies on a Generative AI model (here: Large Language Model). The AI interprets input from users and systems to retrieve relevant information based on a pre-configured "knowledge model" derived from the organization's process data or additional connected data sources within Celonis. It allows users to query data sources, retrieve knowledge, request visualizations, and identify areas for process improvement.
User Interaction: Users interact with the Celonis Process Copilots via an conversational chat interface. They type questions in natural language (e.g., "What was supplier X's late payment rate last quarter?") and receive answers, tables, or charts directly within the chat window.
2. AI Model Dependency
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: Celonis Process Copilots are agentic systems that are orchestrated by Large Language Models. Under the hood, Process Copilots are agents in which the AI takes an input from a user or system, creates an execution plan given a set of predefined tools, decides to execute one of the provided tools, repeats the process until an answer was provided, and then generates the response. The AI model output determines the execution plan, orchestrates the tool execution, and response given the contextual data and knowledge from Celonis. The quality and relevance of the output are highly dependent on the scope and accuracy of the configured knowledge.
3. Data Flow
Data Inputs:
Requests (natural language questions or structured input) are entered into the chat interface or are provided through an API.
The pre-configured knowledge and context: This defines the specific subset of the organization's process data (e.g., specific KPIs, records, attributes, modeled relationships) or other connected sources that Celonis Process Copilots are authorized to access within the Celonis platform.
Instructions and selection/configuration of tools provided by the user configuration.
Data Processing:
Inputs to Celonis Process Copilots are interpreted by the Generative AI model.
The AI queries the structured information within the configured knowledge and data sources.
Instructions and tools may expand the knowledge and context of the Generative AI model to ground requests and responses.
Data Outputs:
Natural language answers to user questions or structured responses.
Generated tables and charts visualizing the requested data.
Outputs are displayed within the Celonis Process Copilot chat interface or through an API response.
4. Human Oversight & Control
Level of Automation: Users or systems initiate interactions with Celonis Process Copilots by asking questions in the chat interface or call an API. The AI generates responses based on its configuration, data sources, and knowledge. Users interpret and decide how to use the provided information. Other systems (also outside of the Celonis platform) may consume the generated response from Celonis Process Copilots and respective system owners decide on the use of the provided response.
Human Intervention Points:
Configuration: Analysts or Center of Excellence (CoE) teams define and curate the knowledge and data sources, for example, selecting which KPIs, records, and attributes the copilot can access. This is a critical control point determining the scope and reliability of the copilot.
Prompting: Users control the questions asked.
Interpretation: Users evaluate the relevance and accuracy of the AI's responses.
Action: Users decide whether and how to act upon the insights provided by the copilot.
Refinement: Users can rephrase or ask follow-up questions if the initial response is unsatisfactory.
Monitoring & Evaluation:
Performance is implicitly monitored through usage patterns and user interactions.
User feedback mechanisms within the interface may be available.
User questions and generated responses may be available for investigation.
The accuracy and usefulness of the copilot are heavily dependent on the quality and relevance of the underlying data sources and provided knowledge within the Celonis platform. Ongoing evaluation of the configured knowledge base by the deploying organization is recommended.
5. Safety & Security
Data Security: Relies on the security infrastructure of the Celonis platform. This typically includes:
Data encryption at rest and in transit.
Role-based access controls inherent to the Celonis platform, ensuring users can only query data they are permitted to see via the copilot, as defined within the knowledge and data source configuration.
Adherence to Celonis' data processing agreements and security standards.
System Reliability: Dependent on the reliability and uptime of the underlying Celonis platform, which typically involves redundancy and failover mechanisms.
6. Ethical Considerations
Fairness & Non-Discrimination: Potential biases could arise from the underlying business process data or the selection choices made during the knowledge and data source configuration. It is crucial that the data fed into the Celonis platform and the specific data points selected for the copilot's knowledge are representative and carefully reviewed to avoid reflecting or perpetuating undesirable systemic biases. The AI itself primarily retrieves configured information; bias mitigation focuses on data source and configuration.
Transparency & Explainability: Transparency is provided through:
This System Card.
The defined scope: The copilot only operates on data explicitly made available through the knowledge and data source configuration.
The Process Copilot also shows which tools were called with which arguments and in which order, allowing users to understand how the Copilot generated the final answer.
While deep explainability of the LLM's internal reasoning for phrasing might be limited (common to GenAI), the source of the factual information can be traced back to the configured knowledge elements within the Celonis platform.