Skip to main content

Data Sources and Training Data

1. Human Oversight and Intervention Points​

The AI functions integrated into OmbuDesk – including AI summaries, translations, text simplifications, corrections, and the chatbot – serve exclusively as supporting assistance systems for case handlers. The AI system does not make any autonomous decisions about the outcome of a dispute resolution procedure. Final responsibility for case processing and outgoing communication always remains with the person handling the case.

To ensure safe and proper use of the software, all users must observe the following intervention points (human-in-the-loop) in their daily work:

  • Reviewing AI summaries: The AI summary provides a quick overview of the current status of a case file. However, before any decision-relevant steps are taken, the case handler must always critically evaluate the generated summary and, if necessary, check it for accuracy and completeness against the underlying original documents (e.g., application forms, emails from the parties).
  • Approval in communication: If AI features such as translation, text simplification, or the correction function are used when composing emails, the generated text must never be sent without review. Before clicking "Send," the user must fully read the email draft, assess its content, and make manual adjustments if necessary before the message is sent to the applicant or respondent.
  • Validating chatbot information: Answers from the OmbuDesk chatbot regarding case details or as a source of ideas for next steps are only recommendations. The case handler must professionally validate this information before incorporating it into the official case processing.
  • Ability to intervene at any time: All users retain full control over the software at every work step. Drafts or summaries generated by the AI can be ignored, manually overwritten, or deleted at any time.

2. Performance Limitations and Accuracy of the AI​

The OmbuDesk AI (based on a European-hosted model of ChatGPT) is a powerful language model for text processing, summarization, translation, and drafting assistance. Nevertheless, the system is subject to certain technical performance limitations that users must always take into account in their daily work:

  • Possible hallucinations and factual errors: As with all generative language models, the AI may occasionally produce plausible-sounding but factually incorrect or fabricated information (so-called "hallucinations"). The AI does not "understand" legal matters in the human sense but calculates the probability of word sequences.
  • Dependence on data quality: The accuracy of the AI outputs depends significantly on the quality, legibility, and completeness of the original inputs. Contradictory or highly complex inputs can lead to inaccurate AI results.
  • No legal or professional assessment: The chatbot and the summary functions under no circumstances replace professional or legal review. The AI is not able to make legally binding assessments.
  • Error-proneness with specific data: Particular care must be taken when adopting specific figures, deadlines, case numbers, or names. These data points must always be cross-checked by a human against the original documents.

3. Avoiding Bias and Discrimination​

Artificial intelligence learns from vast amounts of data and can therefore unintentionally reproduce human prejudices (bias). To guarantee fair and objective case processing, the following principles must be observed:

  • Checking for equal treatment: The results generated by the AI must always be checked to ensure they do not unduly disadvantage either party. No discrimination based on gender, ethnic origin, language, religion, age, or social origin may occur.
  • Balanced summaries: When using the AI summary, care must be taken to ensure that the arguments of both sides are represented objectively and equally.
  • Neutral inputs (prompts): When manually entering information, pay attention to using factual and neutral language.
  • Human responsibility: The AI has no ethical or moral authority. Sole responsibility always lies with the person handling the case.

4. Data Sources and Training Data​

For full transparency, clear guidelines apply regarding the models used and data processing:

  • AI models used (data sources): OmbuDesk uses a specially secured version of ChatGPT, hosted within Europe.
  • No training with customer data: All data entered, uploaded, or processed by the AI as part of case processing in OmbuDesk is explicitly not used for training purposes of the AI models employed.
  • Location of data processing: All data processing for the AI functions takes place exclusively on servers within the European Economic Area (EEA).