UK listed firms' AI risk mentions hit 41.2% in 2025 reports but substantive content stays at 4.3%

A two-stage LLM pipeline classified passages from 2020-2025 filings after validation against 474 human annotations. Recall remained high while label agreement stayed moderate. Mentions of AI adoption climbed in parallel from 13.8% to 45.2%, with vendor references concentrated on Microsoft and a handful of other providers. Critical National Infrastructure sectors and AIM-listed firms lagged Main Market peers in risk disclosure; Energy and Data Infrastructure showed the widest gaps. Harm-specific language appeared in just seven documents across the full corpus. Most risk statements remained non-substantive, listing generic exposure without mitigation details or scenario quantification. This pattern indicates disclosure volume tracks regulatory attention more than operational change. The low substantiveness rate limits the data's value for societal resilience models that require actionable signals. Annual reports processed at this scale can surface sector-level trends but currently under-represent concrete intervention points in executable digital systems. Future pipelines will need finer-grained extraction of control measures to close the gap between stated risk and documented response.