Abstract

Large language models (LLMs) can generate fluent organizational text, but fluency does not establish who is responsible for a proposal, instruction, authorization, or commitment. In conversational workflows, a short message such as “OK” may signal understanding, politeness, or approval, and the organizational standing of such utterances can become ambiguous when LLMs are involved in drafting or mediating exchanges. This paper proposes a conceptual accountability wrapper for LLM-mediated organizational work. The wrapper introduces a small set of interaction-layer markers—acting role, declared interactional purpose, discourse-act label, explicit acceptance or refusal, and a compact audit export—at conversational moments where decisions or obligations may arise. These markers do not attempt to represent the full complexity of human intentions or commitments. Rather, they record limited but governance-relevant traces that can help organizations reconstruct how a decision unfolded and who accepted responsibility for it. Drawing on Language-Action Theory and socio-technical research in information systems, the paper presents the wrapper as a bounded design-science proposal illustrated through organizational scenarios. The contribution is not a complete accountability architecture and does not replace organizational governance mechanisms such as role structures, authority allocation, or compliance procedures. Instead, it offers a lightweight interaction-layer artifact that may improve traceability in LLM-mediated workflows. The paper concludes by outlining organizational preconditions, ethical risks, and an evaluation path for prototype development and pilot studies.

Download

Share

COinS