Many large enterprises now face shadow AI problems: employees quietly using ChatGPT, Claude, or similar tools to get through master-data-heavy and judgement-intensive work, because the sanctioned tools and processes haven't caught up with their work demands. The usual response is a stricter policy. This talk argues that's the wrong approach.

We will go through the findings of a mixed-methods Master's thesis (which combines qualitative interviews with a quantitative investigation into a larger pool of professionals), conducted at a Dutch multinational trading and distribution firm running a proprietary, federated ERP. The research tested which organisational preconditions actually determine the success of GenAI integration in ERP-embedded, knowledge-intensive processes, like product master data enrichment, and whether they lead to overall improved process performance.

I translate the findings, together with the latest 2025-2026 research on agentic AI governance, enterprise BPM adoption, and the EU AI Act's shifting timeline, into a tiered framework for practitioners which any data or IT leader can apply to sequence AI readiness: data foundations first, governance second, confidence-gated automation as third.