Book a Demo

Regulation (EU) 2024/1689

EU AI Act transparency & governance

Last updated

Which AI models Dehurdle uses, what data they receive and what we keep, how we classify Dehurdle under the EU AI Act, and how to ask for a human review of a scorecard.

View the human review processView works council readiness

  • Formative training purpose

    Dehurdle is an educational simulator for skill rehearsal. When used as intended, we assess it as outside the EU AI Act's high-risk employment categories.

  • No automated HR actions

    Dehurdle scores are developmental metrics. Our Terms of Service prohibit using them as the sole or primary basis for disciplinary actions, terminations, or pay decisions (see GDPR Article 22).

  • Human review on request

    Any employee or user can dispute an AI scorecard and request a review by a member of the Dehurdle team.

Classification under EU Regulation 2024/1689 (EU AI Act)

The European Union Artificial Intelligence Act establishes transparency and governance requirements based on risk tiering. Here is how we assess Dehurdle against them:

  • 1. Exclusion from high-risk employment AI (Annex III, Point 4)

    Formative L&D

    Annex III(4) of the EU AI Act classifies as “high-risk” AI systems intended for recruitment or selection, for decisions on promotion, termination or task allocation, or for monitoring and evaluating the performance and behavior of workers. Dehurdle is not designed or offered for these purposes, and our Terms of Service prohibit using it as the sole or primary basis for such decisions. Dehurdle is a formative practice tool where employees rehearse complex communication scenarios.

  • 2. Transparency obligations (Article 50 EU AI Act)

    Full disclosure

    Simulations are presented as conversations with an AI persona (Gemini Live). Every scorecard states that scores, pacing analysis, and rationale are AI-generated and links to a human review request.

  • 3. Prohibition on punitive use

    Binding terms

    Our Terms of Service prohibit using Dehurdle scorecards and competency reports as the sole or primary basis for high-stakes employment decisions, including hiring, terminations, promotions, or performance-based disciplinary actions.

  • 4. No emotion recognition (Article 5(1)(f) EU AI Act)

    Not used

    The EU AI Act prohibits AI systems that infer the emotions of people at work or in education, except for medical or safety reasons. Dehurdle does not infer emotions or psychological states, and it does not label anyone as, for example, anxious, stressed or warm. Its metrics describe observable delivery only: pace in words per minute, pauses and filler words, measured from the timing of your words. Volume, pitch and voice steadiness aren't measured.

    The camera is optional and off by default. If your organization allows it and you turn it on, analysis runs in your browser and no video or images leave your device. Only these numbers are stored with the scorecard: eye-contact ratio, seconds of direct gaze, head movement, number of rapid movements, nods, smiles, a facial expressiveness score (how much the face moves) and a composure score (based on head steadiness and rapid movements).

AI models and processing pipeline

Dehurdle uses Google Cloud (Vertex AI and Speech-to-Text) and no other AI provider. For each model: what it receives, what Dehurdle stores, and for how long. Each organization's retention period is a number of days (default 365) that Dehurdle sets at the organization's request. Persona images are also generated with Google models and use no personal data.

  • Gemini Live

    gemini-live-2.5-flash-native-audio

    Live practice conversations

    Google Cloud · us-central1 (Iowa, USA)

    Plays the AI persona in web and WhatsApp practice calls. It hears what you say and answers in the persona's voice in real time.

    Input data:
    Your live audio during the session, and the scenario the persona plays.
    What's stored:
    No audio. While the session runs, Dehurdle's server holds the audio in memory only; it is never written to disk and is gone when the session ends. The transcript is saved after automatic redaction of email addresses, phone numbers, payment card numbers and some national ID numbers (names and addresses are not removed).
    Retention:
    Audio is not stored after the session ends. Transcripts, including the spoken words kept for voice analysis, are deleted after the organization's retention period (default 365 days).

  • Google Cloud Speech-to-Text

    Transcription

    Google Cloud · Global endpoint

    Turns the AI persona's spoken replies, and anything you say to the coach by voice, into text.

    Input data:
    The persona's audio, and your voice input to the coach.
    What's stored:
    Text only: the persona's lines in the session transcript and your message in the coach chat. The audio is not stored.
    Retention:
    Transcripts and coach chats follow the organization's retention period (default 365 days) and are deleted when the organization is deleted.

  • Gemini 3.5 Flash

    gemini-3.5-flash

    Scoring and feedback

    Google Cloud · asia-south1 (Mumbai, India)

    Reads the finished conversation, scores the five competencies chosen for the scenario and writes feedback.

    Input data:
    The session transcript and the scenario, including the five competencies it scores.
    What's stored:
    Scores and written feedback, saved with the scorecard.
    Retention:
    Written feedback is deleted after the organization's retention period. Scores are kept for progress tracking.

  • Gemini 2.5 Flash

    gemini-2.5-flash

    Coach, speech check and WhatsApp messages

    Google Cloud · asia-south1 (Mumbai, India)

    Powers the in-app coach, analyzes recordings uploaded to the speech check, and understands WhatsApp messages and voice notes.

    Input data:
    Your coach messages, speech-check recordings (up to 25 MB), and WhatsApp messages and voice notes.
    What's stored:
    Coach chats, with email addresses, phone numbers, payment card numbers and some national ID numbers filtered out from October 10, 2026 (earlier messages as written). The speech-check result (transcript, scores and tips), not the recording. No WhatsApp message content or voice notes.
    Retention:
    Coach chats: the organization's retention period (default 365 days). Speech-check results: open to anyone with the result link, can be deleted from the result page, and are deleted automatically after 30 days. WhatsApp conversation state: up to 30 days, keyed by a hashed phone number.

No model training on your data: Dehurdle does not train or fine-tune any AI model on customer data. Google Cloud's terms for Vertex AI do not allow Google to use customer data to train its models.

Human review & appeal protocol

Every employee and participating professional can request a human review of any AI scorecard generated on Dehurdle.

  1. Initiate request

    Select “Request Human Review” at the bottom of any scorecard in the platform, or email reach@dehurdle.com with your Interaction ID.

  2. Human assessment

    A member of the Dehurdle team reviews the conversation transcript, scenario requirements, and competency rationale.

  3. Written response

    You receive a written response by email that either confirms the assessment or explains any correction.

Submit a review request directly

Are you an employee seeking a human review of a scorecard from a completed simulation?

Email a review request

Enterprise AI governance inquiries

For procurement questionnaires or questions about how Dehurdle uses AI, contact our Data Protection and Grievance Officer, Vishal Verma.

Contact reach@dehurdle.com