Trustworthy GenAI
Designing an enterprise explainable GenAI framework to calibrate employee trust in GenAI decisions
—focused on internal GenAI chatbots for work, with a reusable library of guidelines and UX patterns.
RoleEmployerTimelineCollaboration
UX Designer
BMW Group7 Months
Ai Stakeholders
UX Community
Key & End Users
The explainable GenAI framework
A glimpse of best practices applied to a real use case, following the framework guidelines.
Problem
Within BMW, over 1,800 GenAI solutions have been deployed for work. However, user testing shows that employees generally lack trust and still stick to traditional tools. Around 27% of users abandon the AI tools after a single use.
Additionally, the absence of unified, reusable standards has led to redundant development across departments, reducing efficiency and compromising experience consistency.
Approach
In building this framework from 0 to 1, I focused on three main goals.
Calibrating Trust
Avoiding both overtrust and mistrust by calibrating user trust aligns with AI’s actual capabilities.
User-centered Explanations
Focusing on clear, user-helpful explanations over technical details.
Increasing user control
Providing balanced control to users as AI becomes more automated.
Based on theoretical research, I categorized the factors influencing trust in AI into four main categories, forming the foundation of this framework.
The GenAI chatbot, trained on internal company documents, has become a key application and technology. For this research, I selected a mature use case—an GenAI chatbot that retrieves information from production manuals to support troubleshooting. Maintenance technicians can ask questions directly, saving time on searches.
To ensure alignment between business and user needs, I collaborated with stakeholders and key users through multiple iterations, ultimately designing a high-fidelity prototype for testing.
We then conducted three iterative rounds of usability testing with 8 users and extracted key UX patterns from best practices.
Usability Testing Results Mapped Using Affinity Diagramming
Highlights
Design Principle
Humanization
Guideline Applied
Explain AI trustworthiness
Pattern Used
Adaptive Trust Indicator
Why
Avoiding overconfidence – Users tend to trust the AI after several correct answers. With the document together, it provides double verification and encourages critical thinking in an understanding way.
The AI confidence level is represented in a humanized form.
Expressing varying confidence levelsInteraction
Progressive Disclosure –
A color-coded assessment enables quick evaluation, with more details on hover and optional links for deeper insights, ensuring transparency without overload.
❌ Don’t
Misleading Numbers - Assuming users understand probability, a 95% trust rating might make them think the answer is correct. However, it only reflects the AI’s confidence, not its actual accuracy.Design mistakes from the past
Impact
The results were delivered to BMW’s internal AI platform and UX platform for further operational follow-up.
Fachportal RedesignCareCalGenAI Maintenance AssistantMusclePal