EU AI Act Article 12 and the Interaction-Trace Gap: Logging Human–AI Workflows for High-Risk AI Compliance

Article 12 of the EU AI Act requires high-risk AI systems to support automatically generated logs that enable traceability, post-market monitoring, risk identification, and operational oversight. Current logging approaches typically focus on system-level events such as inputs, outputs, timestamps, tool calls, model versions, and errors. We argue that for high-risk human-AI workflows, such logs are conformance-insufficient. Many consequential failures arise not from isolated model outputs, but from breakdowns in role clarity, delegation, oversight exercise, grounding, and repair. We identify this as the interaction-trace gap. The paper proposes a Minimum Viable Interaction-Trace Logging Schema that records authority transitions, declared versus observed system role, oversight checkpoints, repair episodes, grounding signals, and final decision ownership. We compare the schema against existing agent-governance proposals and illustrate its value through a synthetic hiring workflow and the public Moffatt v. Air Canada case. The contribution is a technical interpretation of Article 12 for human-AI workflows and a concrete logging schema for future compliance, audit, and standardisation work.