XARA Privacy Policy

This privacy policy explains which personal data afca ag processes when operating the AI conversation trainer Xara. Xara consists of the web client XaraWeb, the Xara app for Meta Quest and the administration tool XaraAdmin.

Personal data means any information relating to an identified or identifiable person. This policy is designed to meet the requirements of the Swiss Federal Act on Data Protection (FADP) and the EU General Data Protection Regulation (GDPR). Whether and to what extent these laws apply depends on the individual case. Additional information under the GDPR is provided in Section 11.

Xara is used in education and professional training and is intended for people aged 16 and over. Institutions must comply with the applicable legal requirements when their learners use the service, including any required notices, permissions and consents. This does not affect afca's own legal obligations.

afca ag, Industriestrasse 35, 3052 Zollikofen, Switzerland. For data protection requests: datenschutz@afca.ch (contact: Markus Brönnimann).

Xara is normally used by educational institutions and companies (the "institution") for their learners. The institution decides who trains, which scenarios are used and how long conversations are retained. It is the data controller for this processing; afca processes this data on its behalf and according to its instructions. afca is the controller for the operation of the platform itself (access management, security, technical logs).

Sign-in data is not linked to conversation content.

Your institution may require you to enter a user identifier before a conversation, for example a number or a short code. This identifier is stored with the conversation record so that your institution can attribute your exercises to you. afca does not know what the identifier means.

Your institution can determine in XaraAdmin which staff members may view its conversation records and restrict this access to individual persons. Authorised staff also see the feedback and the user identifier.

Conversations are stored pseudonymised. afca does not hold the mapping between the user identifier assigned by your institution and your identity. Conversation content may nevertheless contain personal information if you include it yourself. Do not mention real names or details about real people in training conversations.

Operation generates automatic backend logs: time, function called, session ID, error messages and an anonymised IP address. These backend logs contain no conversation content. Logs of XaraWeb and XaraAdmin may contain your sign-in identifier. If an optional user identifier is used, it may appear in backend logs together with the session ID.

The Quest app stores your access code, the selected language and your settings locally on the headset. With the "Upload log" function you can send the app's most recent technical log entries to afca for troubleshooting. These entries contain the session ID, device model and app version and may contain text recognised during the current conversation as well as the access code you entered. They are transmitted only when you trigger the upload.

The Quest app does not access camera images. The passthrough view of your surroundings is rendered by the headset's operating system and is not visible to the app. The app does not store or transmit any hand, eye, face or body tracking data and collects no location, no advertising ID and no data from your Meta account. No analytics, advertising or social media services are embedded.

The app requires access to the microphone (voice input) and to the internet. No other permissions are requested.

Microsoft Azure AI Speech converts your speech to text and the avatar's replies to speech. Processing takes place in Microsoft data centres in Switzerland or in the EU. Microsoft stores neither audio nor text. In live conversation mode, Azure AI Speech is not used; speech recognition and speech synthesis are handled by the language model (Section 5.3).

For the avatar's replies and the feedback we use the Microsoft Azure OpenAI Service (including GPT-4o and Realtime models). The language model receives your institution's scenario instructions, the conversation so far and your current utterance as text. It does not receive sign-in data, your user identifier or, except in live conversation mode (Section 5.3), your voice recording.

All models used by default are deployed with the deployment type "Data Zone Standard" in Microsoft's EU data zone. Processing takes place within this data zone, which follows the Microsoft Azure EU Data Boundary and comprises the EU member states as well as EFTA states such as Switzerland and Norway.

Microsoft does not use your inputs to train AI models and does not store them, with one exception: Microsoft automatically screens inputs for abuse and may retain content flagged as potentially abusive for up to 30 days for review. This content is stored in the geography of the respective Azure resource (Switzerland or EU). For resources in the European Economic Area, the review is carried out by authorised Microsoft employees located in the European Economic Area.

Institutions can use a live conversation mode for individual scenarios in XaraWeb, in which you speak with the avatar without waiting times, as in a real conversation, and can also interrupt it. The mode is switched off by default. It is used only if it has been enabled for your institution and switched on in XaraAdmin for the scenario concerned. It is not available in the Xara app for Meta Quest.

In live mode, your voice is not converted to text by Azure AI Speech but streamed directly to a Realtime language model of the Microsoft Azure OpenAI Service. The model understands your utterance from the audio and replies in spoken language. It hears not only what you say but also how you say it (for example tone of voice, speaking rate, pauses) and can respond to this in the conversation. It also picks up noises and voices from your surroundings; where possible, train without other people talking in the background.

The audio data passes through an afca server in Switzerland (Microsoft Azure, Switzerland North region), which only forwards it and does not store it. For the conversation record, a Microsoft transcription model (also Azure OpenAI Service) converts your utterances to text. As in standard mode, only the text is stored (Section 3.3); afca stores neither your voice recording nor the avatar's speech output.

The language model and the transcription model are deployed in the EU data zone by default; Section 5.2 applies (exception: Section 5.4). In live mode, the abuse monitoring described there may also cover your voice recording.

Your voice is not analysed to recognise or identify you, and no voice profiles are created. In live mode, too, AI feedback is based on the conversation record in text form (Section 5.5).

Institutions may select additional models for individual scenarios: models offered through Microsoft Azure with the deployment type "Global Standard" (for example gpt-audio) or Google Gemini. These options are disabled by default and can only be used if the institution explicitly lifts the setting "AI models Switzerland/EU only" and assigns a corresponding model to a scenario. These models also receive conversation text, but not your voice recording. The exception is a Microsoft Realtime model with Global Standard in a scenario with live conversation mode: it also receives your voice recording. The institution selects the models in XaraAdmin; this privacy policy is also linked there.

With Global Standard, Microsoft may process data in any Azure region worldwide where the model is available; content retained for abuse monitoring stays in the geography of the Azure resource (EU). Gemini processing takes place in the USA; transient storage or caching may also occur in other countries where Google or its agents maintain facilities.

Microsoft and Google do not use the transmitted content to train or improve their AI models. Google retains content for abuse monitoring for a limited period that it does not specify.

AI feedback supports your reflection on the training conversation. It is generated automatically from the conversation content in text form, not from your voice recording even in live conversation mode, and is not a grade. afca does not use it to assess individuals, makes no automated decisions with legal effect and does not carry out profiling. As part of the conversation record, the feedback is visible in XaraAdmin to the staff your institution has authorised (Section 3.3). Your institution decides whether and how it uses conversation records.

We disclose personal data only to the following processors, to the extent necessary for operation. No data is passed on for advertising or for third parties' own purposes.

The scenario optimiser in XaraAdmin processes only scenario texts written by trainers, not conversations from XaraWeb or the Quest app. Anthropic does not use this content to train its models and normally deletes it within 30 days. If Anthropic's automated safety systems flag content as violating its usage policy, Anthropic may retain it for up to 2 years. Please do not include personal or confidential information in scenario texts either.

If you access Xara through a hep Verlag account, hep Verlag AG, Bern, verifies your sign-in and licence under its own responsibility and passes us the data listed in Section 3.1.

Under Swiss law, EU and EEA countries are recognised as providing an adequate level of data protection. For transfers to the USA and other countries by Google and Anthropic and for any other international transfers by Microsoft, we rely on the contractual safeguards named above. For transfers to third countries from the perspective of the GDPR, see Section 11.3.

When the customer relationship ends, all of the institution's data is exported on request and then deleted.

The Quest app and XaraWeb use no cookies for analytics or advertising. XaraWeb stores technically necessary settings and your access token locally in the browser. When you sign in through Microsoft Entra ID (XaraWeb, XaraAdmin), Microsoft sets cookies for authentication and single sign-on; Microsoft's privacy statement applies (privacy.microsoft.com). Google Analytics, social media plugins and trackers are not embedded.

We take appropriate technical and organisational measures: encryption in transit (TLS) and at rest, separation of data per institution, role-based access limited to staff who need it for operation, key management in Azure Key Vault, separate test and production environments, monitoring and logging. The infrastructure is covered by Microsoft Azure's certifications (including ISO/IEC 27001, ISO/IEC 27701, SOC 2).

Within the scope of the applicable data protection law you have the right to access, rectify and delete your data and to object to its processing. Please contact datenschutz@afca.ch. Please do not send passwords or access tokens. Where the user identifier assigned by your institution is needed to locate the relevant records, we coordinate the request with the institution.

afca handles correction and deletion requests without undue delay. We delete conversation and feedback data, synthesised avatar audio, usage logs and exports within 7 days, and technical logs, including uploaded app logs, within 30 days at the latest. Deleted records may remain in rolling backups for up to 7 days.

You may also lodge a complaint with the competent data protection authority, in Switzerland the Federal Data Protection and Information Commissioner (www.edoeb.admin.ch). For people in the EU and the EEA, Section 11.4 also applies.

This section applies to the extent that the GDPR applies to the processing of your data, in particular if you use Xara through an institution in the EU or the EEA. It supplements the other sections. The German version of this policy uses the Swiss legal terms "Personendaten", "Bearbeitung" and "Auftragsbearbeiter"; they correspond to the GDPR terms personal data, processing and processor.

For your conversation data, your institution is the controller within the meaning of Art. 4(7) GDPR; afca processes this data as a processor (Art. 28 GDPR). afca itself is the controller for the operation of the platform (Section 2).

Your institution determines the legal basis for the conversation data and the optional user identifier (Sections 3.2 and 3.3) and informs you about it.

Where afca itself is the controller, we rely on the following legal bases:

You are not obliged to provide us with data. However, without the information required for access you cannot use Xara.

From the perspective of the GDPR, Switzerland is a third country. The European Commission has adopted an adequacy decision for Switzerland; the transfer of your data to afca is based on it (Art. 45 GDPR). With the default settings, your conversation data is processed only in Switzerland and in EU and EEA states (Sections 5 and 6).

Processing in the USA or in other third countries takes place only in the cases described in Sections 5.4 and 6. It is based on an adequacy decision (Art. 45 GDPR) or on the European Commission's Standard Contractual Clauses named in Section 6 (Art. 46(2)(c) GDPR). You can obtain a copy of these safeguards on request from datenschutz@afca.ch.

In addition to the rights under Section 10, you have the right to restriction of processing (Art. 18 GDPR) and to data portability (Art. 20 GDPR). You may object, on grounds relating to your particular situation, to processing that is based on a legitimate interest (Art. 21 GDPR). Where processing is based on your consent, you may withdraw it at any time with effect for the future.

Your institution is the controller for your conversation data. You can exercise your rights with the institution or with us; we forward your request to the institution and support it in responding.

You have the right to lodge a complaint with a data protection supervisory authority, in particular in the Member State where you live or work or where the alleged infringement took place (Art. 77 GDPR).

No automated decision-making, including profiling, within the meaning of Art. 22 GDPR takes place (Section 5.5).

We update this policy when our data processing changes. The version published on afca.ch applies. The German version is binding.