INNO 139.045 Argentic AI Coach
BFH studies how an agentic AI coach designed to build confidence in young men can be evaluated for safety, bias, and psychological quality, and how its challenges can be adapted to individual users.
Factsheet
- Schools involved School of Engineering and Computer Science
- Institute(s) Institute for Data Applications and Security (IDAS)
- Funding organisation Innosuisse
- Duration (planned) 01.07.2026 - 23.12.2026
- Head of project Prof. Dr. Souhir Ben Souissi
- Project staff Veronica Caselli
- Partner Konstruct Labs SA
Situation
Many young men seek opportunities for personal growth and greater confidence, yet the support currently available does not always meet their needs. Traditional approaches such as therapy and professional coaching can face barriers related to stigma, cost, and accessibility, while online self-development content is often passive and provides limited support for translating advice into concrete behavioural change. At the same time, some online communities targeting men, including the so-called “manosphere,” may promote extreme or potentially harmful advice. Existing AI companions, wellness applications, and digital coaching platforms provide alternative forms of support, but they lack mechanisms for systematically adapting the level of challenge to the user and for ensuring that AI-generated interactions follow appropriate psychological and safety principles. Particularly there is a need for approaches that combine personalised, action-oriented guidance with systematic evaluation of response quality, appropriateness, bias, and safety. Konstruct Labs SA aims to address this gap by developing a science-backed AI Coach-Agent for men aged 16-34. The envisioned solution should provide psychologically aligned, and action-oriented guidance, adapted to different user profiles and contexts. It should incorporate the Comfort-Growth Paradox to dynamically select and adjust the intensity of challenges, while integrating safety guardrails to minimise harmful, biased, or unsuitable responses
Course of action
The Generative AI Lab at BFH will conduct a comprehensive concept and feasibility study to address critical questions related to the development and evaluation of the English-language Coach-Agent. The study will explore how state-of-the-art NLP and conversational AI evaluation methods can be used to systematically assess the agent’s responses in terms of quality, appropriateness, and alignment with established psychological frameworks. Additionally, it will investigate methods for detecting potentially problematic responses, assessing whether the agent’s language is appropriate for the target group, and identifying potential differences in response quality across user profiles characterised by emotional state, cultural background, or socioeconomic signals. The study will also examine how the behavioural specificity of responses can be automatically assessed to distinguish action-oriented guidance from passive or purely reflective responses. A particular focus will be placed on translating the Comfort-Growth Paradox into a decision engine capable of selecting appropriate challenges and adapting their intensity to the user and context. Finally, suitable guardrail architectures will be investigated to ensure that simulations of difficult conversations remain challenging and useful while minimising the risk of harmful or uncontrolled responses.
