Skip to main content
1. Where exactly is AI applied in the Honestly software?
  • Topics – The AI assistant groups similar open-text responses into topics and rates the sentiment of each response.
  • Supporting passages in Topics – A separate pass highlights the passage in the response text that a topic assignment rests on, so you can see why a response was assigned to a topic.
  • Data masking – A language model reads incoming open-text responses and marks the passages containing personal details – first and last names, email addresses, phone numbers, postal addresses, and unique IDs or customer numbers – so that they can be replaced.
  • Initiatives – AI provides recommendations for initiatives based on dashboard data.
  • Bubble Chart on the dashboard – All open-text responses are summarised and their sentiment analysed. The results are clearly displayed in the chart.
  • Competency Matrix in 360° Feedback – Under AI Analysis of this Data, the AI evaluates the matrix results and turns them into an interpretation, a reflection question, and a development step.
  • Qualitative Hints in 360° Feedback – If Include open text answers is switched on, a second pass evaluates the free-text answers from the 360° feedback and derives recurring themes from them.
  • Company context – On request, the AI researches your company from its Web domain and pre-fills the wizard. You then review and edit every field yourself.
  • Translation – With Translate, open-text responses in the Topics section, in Text charts on a dashboard and in the topic responses of the Bubble Chart and Topic Trend Chart, as well as messages in the Chat, can be translated on click. They are translated by the Azure models described in question 2. Separately, when you build a multilingual survey, a 360° Feedback run, or the Code Participation login screen, Suggest translation pre-fills the other languages. These translation suggestions run through DeepL and Google Translate, not Azure (see question 2).
  • Speech-to-text – On a mobile device, spoken responses in the survey are transcribed into text live, while the participant is speaking.
2. Which AI is specifically used? We use Microsoft’s Azure AI Foundry, specifically generative models, which are operated in the Azure Cloud in Europe. The models are fully isolated from the public OpenAI service. Processing is carried out exclusively by Microsoft in their EU data centres. Here is the official privacy policy of Azure OpenAI Service. One function does not use the generative Foundry models: Speech-to-text. Spoken responses on a mobile device are transcribed by a separate Microsoft service, the Azure speech-recognition service, which likewise runs in Microsoft’s EU data centres. For the translation suggestions in surveys, 360° Feedback, and Code Participation, we use two dedicated translation services rather than Azure: DeepL for the languages DeepL supports, and Google Translate for every other language.
One exception: the company context research. For this one function, the model accesses publicly available websites through a web search. Together with the translation suggestions, it is one of the two data paths that lead beyond Microsoft’s EU data centres. What goes into it is solely the Web domain you enter and the desired output language – no employee data and no survey responses. The function is not active by default, has to be enabled for your account, and requires consent to Microsoft as a subprocessor.
3. Is the AI trained with our data? No. Microsoft does not use the data to improve OpenAI models or its own products. Honestly also does not use your data for model training. The models used are stateless and non-learning, unless fine-tuning is explicitly initiated – which we do not do. 4. How transparent is the way the AI works? The Topics AI uses language models to identify frequent terms, concepts, and patterns in open-text responses and automatically assign them to topics. The logic is based on probabilities (semantic similarity). 5. Is the use of AI mandatory? No. The functions are optional. Companies can disable them – using Honestly is fully possible without AI. 6. What data is processed and where? The following data is processed:
  • Open-text responses from surveys – for Topics, supporting passages, the Bubble Chart, data masking, and – when a translation is requested – for the translation. They may potentially contain personal information.
  • Free-text answers from 360° Feedback – for AI Analysis of this Data on the Competency Matrix, provided Include open text answers is switched on.
  • Messages from the Chat – when a translation is requested.
  • Survey content sent for translation suggestions – survey questions and answer options, 360° Feedback texts, and Code Participation texts, when you request a translation. This is content written by admins; employee responses and other personal data are not sent.
  • Your company’s web domain – for the company context research. No employee data and no survey responses go into it.
  • Audio of spoken responses – for Speech-to-text, when the add-on is enabled. The audio is processed solely for live transcription and is not stored.
The data is encrypted and processed on Microsoft servers in the EU. There are two exceptions: the web search used for the company context research (see question 2), and the translation suggestions, which are handled by the external services DeepL and Google Translate. Speech-to-text is an add-on you book separately. It requires consent to Microsoft as a subprocessor, and the audio is used only for the live transcription and is never stored. 7. How does Honestly address the risk of personal data being included in free-text responses? We recommend using free-text fields intentionally and with specific questions to minimise the likelihood of personal data being entered. If desired, the Topics AI can be applied only to certain questions or disabled entirely. 8. Is a Data Protection Impact Assessment (DPIA) necessary? Depending on your company’s assessment, a DPIA may be required – for example, when conducting extensive analysis of employee feedback. Upon request, we provide information to support this process, such as function descriptions, data flows, and sub-processor information. 9. How can Honestly support with legal reviews (e.g. with the works council or data protection team)? We can provide: Function description (German/English) List of all subprocessors (incl. Microsoft) Technical description of the Topics AI Privacy policy Support with DPIA Data processing agreement & TOMs FAQs for employees Deactivation option 10. What co-determination rights does the works council have? According to §87 BetrVG, co-determination is required if technical systems are used, or could be used, to monitor behaviour or performance. Even if no active monitoring takes place, the mere possibility may be sufficient. Therefore, we recommend involving the works council as early as possible and keeping them continuously informed and engaged throughout the entire project. This creates transparency, fosters a trusting collaboration, and ensures the success of the project. 11. What risks exist when using LLMs (Large Language Models) in an HR context? Interpretation errors in semantic analyses Lack of explainability in decisions Entry of personal data by employees Perception as a monitoring tool Challenges in GDPR-compliant implementation We address these risks through: Transparency Clearly defined data flows Data processing agreements with subprocessors Freedom of choice (enable/disable function) Training and help articles