General-purpose large language models are increasingly used by employees to interpret standards, troubleshoot systems, and draft or refine engineering artefacts. This routine assistance creates bidirectional flows: proprietary documentation is occasionally externalised as prompts, uploads, screenshots, or connector-mediated retrieval, while model outputs are pasted back into internal tickets, runbooks, and repositories. This integrative review synthesises heterogeneous evidence (peer-reviewed research, provider and regulator materials, and structured incident reporting) to map employee-driven leakage mechanisms along the documentation lifecycle and to derive a governance approach that is auditable under policy drift and multi-vendor toolchains. We identify a recurrent set of boundary-crossing transition points such as copy-paste, upload/OCR, connector invocation, and paste-back, where risk concentrates and where observability is often weakest. Across these pathways, four cross-cutting risk dimensions recur: confidentiality and competitive exposure, compliance and cross-border transfer, model-side effects (including extraction, spillover, and contamination risks), and incentive-driven governance gaps that sustain shadow workflows. Building on the mechanism map, we propose a proportionate "minimal guardrail stack" and an organisational evaluation framework combining qualitative risk scoring, rule-based escalation, and simple, trackable metrics (e.g., consolidation onto sanctioned channels, blocking effectiveness, inspection false positives, policy-drift lag, and time to a compliant alternative). The paper does not assert prevalence. Instead, it aims to make assumptions explicit and support cautious, workflow-compatible adoption decisions.
Latinović et al. (Thu,) studied this question.