NTTコムウェア's 30％ AI-run ops unit bans パワポ文書

The platform services team did not move to AI first and then fix its documents. According to Kurama, the team had long answered internal inquiries by asking people who already knew the answer, and it struggled with the state of its internal data before deploying AI agents. Its work spans AIOps, SRE, front-end and back-end development, and security operations, so the documents feeding those workflows were diverse and often not in AI-ready form. The team first ran RAG training on internal documents, but the results fell short of expectations.

Kurama points to three causes for that shortfall. One is that reporting formats themselves were not standardized, with files named with versions like 0.6 or 0.7 and copies scattered around, leaving AI unable to tell which data to reference. Another is that people could not frame the right questions. The パワポ文書 problem sat alongside these: slides read well to humans who filled in the gaps, but they carry no tags, so AI cannot follow them. Kurama also notes two broader shifts — falling human writing ability and advancing terminology technology — that appear to be driving demand for the text-only rule. He cites Amazon's practice of running decision meetings on Microsoft Word alone as a known reference point, because it forces participants to state the point, the issues, and who decides.

For readers, the stakes are less about the format rule itself and more about sequencing. Kurama argues that without a clear business scenario, AI adoption becomes a string of costly "token-wasting" experiments, and that companies should define purpose and outcome before deciding what AI should handle. Whether the team reaches its 70％ goal may depend less on model capability than on whether it can standardize its documents and terminology enough for AI to use them.