Case study
Trust by Design: from principles to AI guidelines
A framework for responsible AI adoption in public institutions, developed through MA research into public sector AI systems and tested with government teams.
The situation
Public sector AI systems in Germany are often built with real rigour around accuracy, security and compliance. Yet technical quality alone doesn't make people trust a system, use it well or know what to do when it gets something wrong. The technical layer can be solid while the human layer is missing. My research question was how to design the layer that makes technical excellence trustworthy to the people who depend on it.
Audit
What I found
I studied three publicly documented government AI systems: Assistent.iQ from the Bundesdruckerei KI-Kompetenz-Center, the federal data analytics platform PLAIN, and the AI quality benchmarking initiative MOVE. Using critical journey mapping across helpdesk, identity and authorisation, and cross-system credentials, plus a gap analysis of literacy, explainability, recovery and accountability, I found the same kind of gap in each. Assistent.iQ was designed to generate accurate guidance, but its public documentation showed no way for people to understand or contest an output. PLAIN enabled secure, sovereign analysis, but without onboarding and usability context, adoption across teams struggled. MOVE benchmarked model quality against EU AI Act criteria, but measured technical performance only, never how people experienced the system.
Pathway
The pathway we designed
I developed a Trust Layer: a distinct layer between AI infrastructure and the people who use it, whether citizens, patients or frontline staff. It turns four pillars into design requirements that any team can check before a system goes live. Explainability: can people understand what the system did and why? Override: can they question, stop or correct it? Recoverability: can they get back on track when it goes wrong? Accountability: is it clear who is responsible?
Prototype
What we made tangible
The framework is published openly at trust-by-design.org. I also built TrustAudit (trustaudit.tools), an interactive self-assessment that lets organisations find their own explainability, recovery and control gaps, and TrustBridge (trustbridge.design), which shows how the Trust Layer applies across regulated sectors. All three were built with v0, Cursor and Vercel.
Test
What testing showed
I presented the framework to more than six communities within German government and ran advisory sessions with process management leaders and cross-functional teams in regulated government institutions. Operational teams recognised the gap quickly once it was named, and connected the principles to their own systems without prompting. These are qualitative responses from those sessions, not measured outcomes.
Scale
What scaled, changed or stopped
The framework became the basis of an independent practice offering trust audits, framework implementation and EU AI Act readiness.
What the organisation can do now
Teams have a shared vocabulary for a problem they had felt but couldn't name, four questions they can check before go-live, and a free self-assessment to find their own gaps.
My role
Sole researcher and author. The work formed my MA thesis "Trust by Design", graduated with distinction (9.01/10). I designed and built the three tools and delivered the advisory sessions.