ローカルLLMでISMS適合状況評価を支援する ― ヒアリングから報告書まで5日間の「人とLLMの分担」A practitioner report on delegating ISMS conformity assessments to a local LLM…

A practitioner has published a detailed account on Qiita describing how a local large language model, running entirely on a business PC, can assist with assessing conformity to an Information Security Management System (ISMS). The central appeal is straightforward but consequential: because the model runs locally, a client's confidential material never leaves the machine at any point during the evaluation, which addresses one of the most persistent objections to using generative AI in security and compliance work.

The reported workflow spans five days and is framed explicitly as a division of labor between the human assessor and the LLM. According to the report, once processing is kicked off in the evening, the model works overnight, and by the next morning pass or fail judgments for the full set of ISMS requirements are complete, along with draft improvement proposals. Throughout that period, the author emphasizes, the customer's confidential information does not travel outside the business PC. The scope described covers the entire span of an engagement, from initial hearings or interviews through to the finished report, rather than a single narrow task.

For readers unfamiliar with the domain, ISMS conformity assessment typically refers to checking an organization's controls and documentation against the requirements of the ISO/IEC 27001 standard and its associated control catalogue. This is inherently document-heavy work: assessors read policies, procedures, records, and interview notes, then decide whether each requirement is satisfied and, where gaps exist, recommend corrective action. That combination of structured requirements, large volumes of text, and repetitive judgment makes it a plausible candidate for LLM assistance, while the sensitivity of the underlying evidence makes data confidentiality a hard constraint.

The reason the local-versus-cloud distinction matters so much here is regulatory and contractual as well as technical. Sending client security documentation to an external API can conflict with confidentiality agreements, data residency expectations, or the very ISMS controls being assessed. Running the model on the same PC where the work is performed sidesteps that exposure, which is why the report frames "how far you can delegate to a local LLM" as the question it set out to measure. The piece reads as an empirical record of testing that boundary rather than a vendor claim about a finished product.

The broader ecosystem context helps explain why this kind of experiment is now feasible. Tools such as Ollama, LM Studio, and llama.cpp have made it considerably easier to run open-weight models locally, and model families like Llama, Qwen, Gemma, and Mistral have narrowed the capability gap with hosted services for many text tasks. Quantization techniques allow sizeable models to run on ordinary business hardware, though performance and output quality remain constrained by available memory and compute. The overnight batch approach described in the report is consistent with those constraints: local inference can be slower than cloud endpoints, so scheduling long runs outside working hours is a pragmatic way to absorb that cost.

It is worth being cautious about how much the LLM is actually being trusted to decide. The report positions the model as producing pass/fail judgments and draft improvement proposals, but the framing as a human–LLM division of labor implies that a qualified assessor still reviews, corrects, and takes responsibility for the conclusions. Generative models are known to produce confident but incorrect statements, and conformity judgments carry real audit and certification consequences, so treating the model's output as a first draft rather than a final verdict appears to be the intended posture. The five-day figure likely reflects the full cycle including human hearings and review, not autonomous model runtime alone.

As a single practitioner account, the write-up should be read as one data point rather than a benchmark. It does not, based on the available excerpt, specify the exact model, hardware, or accuracy measurements, and results are likely to vary with the model chosen, the prompt design, and the maturity of the organization being assessed. Even so, it reflects a growing interest in applying locally hosted LLMs to regulated, confidential workflows where cloud services are difficult to use, and it offers a concrete template for how such tasks might be structured, with the human retaining judgment and the model handling volume.