[{"data":1,"prerenderedAt":1418},["ShallowReactive",2],{"article-\u002Farticles\u002Fshagai-ni-dasenai-joho-seisei-ai":3,"article-series-none":1417},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"category":11,"service":12,"author":13,"thumbnail":14,"tags":15,"faq":21,"body":37,"_type":1411,"_id":1412,"_source":1413,"_file":1414,"_stem":1415,"_extension":1416},"\u002Farticles\u002Fshagai-ni-dasenai-joho-seisei-ai","articles",false,"","社外に出せない情報で生成AIを使うには｜プロンプトの承認・自動マスキング・オンプレLLMの費用の考え方","機密性の高い情報を扱う会社が、生成AIを使うときの3つの選択肢（上長の事前承認・自動マスキング・社内で動くLLM）を、費用と手間の順に解説します。プロンプトから情報が出る経路、取り込む資料の利用規約、GPUの費用をP\u002FLで試算する考え方まで。","2026-11-19","AX・業務自動化","ai-security","K.S.ロジャース株式会社","\u002Fimg\u002Fcontents\u002Fnews\u002Fthumb160.jpg",[16,17,18,19,20],"生成AI","情報漏洩対策","RAG","オンプレミスLLM","AIガバナンス",[22,25,28,31,34],{"q":23,"a":24},"法人向けの生成AIサービスなら、入力した内容は学習に使われませんか？","主な法人向けサービスは、既定では入力と出力をモデルの学習に使わないと公式に示しています。ただし「学習に使わない」と「社外に送らない」は別の話です。不正利用の監視のために一定期間保存される場合や、処理される地域が設定で変わる場合があります。社内の決まりで社外に出せない情報は、学習の有無とは別に扱いを決めます。",{"q":26,"a":27},"プロンプトの自動マスキングだけで、機密情報は守れますか？","マスキングだけでは守りきれません。辞書にない新しい用語や表記ゆれは伏せ漏れになり、伏せすぎると AI が質問の意味を取れなくなります。また、何が会社固有の情報かを見つける仕組みそのものを社外の AI に頼ると、本末転倒になります。マスキングは、社内の決まりで「伏せれば出してよい」とされた情報に使い、伏せても出せない情報は別の手段で扱います。",{"q":29,"a":30},"オンプレミスのLLMには、どのくらい費用がかかりますか？","費用の中心は GPU を積んだマシンで、ほかに設置場所・電源・保守の手間、モデルや検索の仕組みの更新がかかります。金額は、使う人数・同時に使う量・求める精度で大きく変わるため、購入する場合と月額で借りる場合の両方を試算し、減らせる残業代や外注費など P\u002FL への効果と比べて、払えるかどうかを先に判断します。",{"q":32,"a":33},"社外の資料をRAGに取り込むときは、何に注意すればよいですか？","取り込む前に、資料の利用規約やライセンスを確かめます。購入した規格や技術資料の中には、AI の学習への利用や、複製・データベース化を禁じているものがあります。RAG への取り込みが禁止の対象に当たるか迷う場合は、発行元や社内の法務・知財の担当に確認し、取り込む資料の一覧と確認の結果を残しておきます。",{"q":35,"a":36},"プロンプトを上長が事前に承認する運用は、現実的ですか？","全部のプロンプトを承認にすると、承認待ちで使われなくなります。対象を「社外に出せない情報に触れる業務」に絞る、毎回ではなくプロンプトの型を一度承認する、それ以外は事後の確認にする、といった形で範囲を限ると現実的になります。",{"type":38,"children":39,"toc":1396},"root",[40,48,61,66,105,112,117,204,216,346,351,365,381,387,392,416,423,428,433,466,478,552,565,571,576,657,662,685,690,703,709,714,735,750,755,797,812,817,823,835,909,929,944,950,955,1054,1066,1085,1100,1105,1110,1126,1272,1277,1290,1296,1301,1350,1355,1360,1383],{"type":41,"tag":42,"props":43,"children":44},"element","p",{},[45],{"type":46,"value":47},"text","「生成AIを業務に使いたい。ただ、社内の決まりで技術資料をクラウドに出せない」――製造業・医療・自治体など、機密性の高い情報を扱う組織から、こうした相談をいただくことが増えています。",{"type":41,"tag":42,"props":49,"children":50},{},[51,53,59],{"type":46,"value":52},"よく話題になるのは「入力した内容が AI の学習に使われるか」です。しかし、機密を扱う現場で本当に問題になるのは、その手前です。社内の資料を検索させる仕組み（RAG）でも、調べものを任せる使い方でも、",{"type":41,"tag":54,"props":55,"children":56},"strong",{},[57],{"type":46,"value":58},"質問の文章（プロンプト）そのもの",{"type":46,"value":60},"に会社固有の用語や情報が入り、それがクラウドの AI に送られます。",{"type":41,"tag":42,"props":62,"children":63},{},[64],{"type":46,"value":65},"本記事では、社外に出せない情報で生成AIを使うための選択肢を費用と手間の順に3つに整理し、取り込む資料の利用規約、社内で動く AI（オンプレミスの LLM）の GPU の費用の考え方まで、当社の支援の現場で論点になったことを交えて解説します。",{"type":41,"tag":67,"props":68,"children":69},"blockquote",{},[70,78],{"type":41,"tag":42,"props":71,"children":72},{},[73],{"type":41,"tag":54,"props":74,"children":75},{},[76],{"type":46,"value":77},"この記事の要点",{"type":41,"tag":79,"props":80,"children":81},"ul",{},[82,88,93],{"type":41,"tag":83,"props":84,"children":85},"li",{},[86],{"type":46,"value":87},"情報が社外に出るのは、プロンプト・取り込む資料・出力と履歴の3か所です。「学習に使われない」は、社外に送らないこととは別の話です",{"type":41,"tag":83,"props":89,"children":90},{},[91],{"type":46,"value":92},"選択肢は、①送る前に人が確かめる（上長の事前承認）②固有の情報を自動で伏せる（自動マスキング）③社内で動く AI（オンプレミスの LLM）の3つで、後ろほど費用と手間が大きくなります",{"type":41,"tag":83,"props":94,"children":95},{},[96,98,103],{"type":46,"value":97},"オンプレにすると GPU が費用の中心になります。当社の支援先でも、その費用を P\u002FL で払えるかが重要な論点になりました。取り込みたい資料の利用規約で AI の学習への利用が禁じられていた例もあり、",{"type":41,"tag":54,"props":99,"children":100},{},[101],{"type":46,"value":102},"資料の権利と費用を最初に確かめる",{"type":46,"value":104},"ことが近道です",{"type":41,"tag":106,"props":107,"children":109},"h2",{"id":108},"プロンプトに書いた内容も社外に出る-入力資料出力のどこで情報が外に出るか",[110],{"type":46,"value":111},"「プロンプトに書いた内容」も社外に出る ―― 入力・資料・出力のどこで情報が外に出るか",{"type":41,"tag":42,"props":113,"children":114},{},[115],{"type":46,"value":116},"生成AIで情報が社外に出る経路は、大きく3つあります。",{"type":41,"tag":118,"props":119,"children":120},"table",{},[121,145],{"type":41,"tag":122,"props":123,"children":124},"thead",{},[125],{"type":41,"tag":126,"props":127,"children":128},"tr",{},[129,135,140],{"type":41,"tag":130,"props":131,"children":132},"th",{},[133],{"type":46,"value":134},"経路",{"type":41,"tag":130,"props":136,"children":137},{},[138],{"type":46,"value":139},"何が外に出るか",{"type":41,"tag":130,"props":141,"children":142},{},[143],{"type":46,"value":144},"見落としやすい点",{"type":41,"tag":146,"props":147,"children":148},"tbody",{},[149,168,186],{"type":41,"tag":126,"props":150,"children":151},{},[152,158,163],{"type":41,"tag":153,"props":154,"children":155},"td",{},[156],{"type":46,"value":157},"入力（プロンプト）",{"type":41,"tag":153,"props":159,"children":160},{},[161],{"type":46,"value":162},"質問の文章、貼り付けた本文",{"type":41,"tag":153,"props":164,"children":165},{},[166],{"type":46,"value":167},"製品の呼び名・社内の略語・顧客名など、会社固有の言葉が質問文に入る",{"type":41,"tag":126,"props":169,"children":170},{},[171,176,181],{"type":41,"tag":153,"props":172,"children":173},{},[174],{"type":46,"value":175},"取り込む資料",{"type":41,"tag":153,"props":177,"children":178},{},[179],{"type":46,"value":180},"RAG に入れる文書、アップロードしたファイル",{"type":41,"tag":153,"props":182,"children":183},{},[184],{"type":46,"value":185},"検索に使う「埋め込み」の作成や、資料の保管がクラウドで行われる",{"type":41,"tag":126,"props":187,"children":188},{},[189,194,199],{"type":41,"tag":153,"props":190,"children":191},{},[192],{"type":46,"value":193},"出力と履歴",{"type":41,"tag":153,"props":195,"children":196},{},[197],{"type":46,"value":198},"AI の回答、やりとりの履歴、ログ",{"type":41,"tag":153,"props":200,"children":201},{},[202],{"type":46,"value":203},"回答には質問と資料の内容が反映される。履歴やログがサービス側に保存される場合がある",{"type":41,"tag":42,"props":205,"children":206},{},[207,209,214],{"type":46,"value":208},"ここで押さえておきたいのは、",{"type":41,"tag":54,"props":210,"children":211},{},[212],{"type":46,"value":213},"「学習に使われない」と「社外に送らない」は別の話",{"type":46,"value":215},"だということです。主な法人向けサービスは、既定では入力と出力を学習に使わないと公式に示しています。",{"type":41,"tag":118,"props":217,"children":218},{},[219,235],{"type":41,"tag":122,"props":220,"children":221},{},[222],{"type":41,"tag":126,"props":223,"children":224},{},[225,230],{"type":41,"tag":130,"props":226,"children":227},{},[228],{"type":46,"value":229},"サービス",{"type":41,"tag":130,"props":231,"children":232},{},[233],{"type":46,"value":234},"公式に示されている条件（2026年10月確認）",{"type":41,"tag":146,"props":236,"children":237},{},[238,262,283,304,325],{"type":41,"tag":126,"props":239,"children":240},{},[241,246],{"type":41,"tag":153,"props":242,"children":243},{},[244],{"type":46,"value":245},"Microsoft 365 Copilot",{"type":41,"tag":153,"props":247,"children":248},{},[249,251,260],{"type":46,"value":250},"プロンプト・応答・Microsoft Graph 経由でアクセスしたデータは、基盤となる LLM のトレーニングに使われない（出典：",{"type":41,"tag":252,"props":253,"children":257},"a",{"href":254,"rel":255},"https:\u002F\u002Flearn.microsoft.com\u002Fja-jp\u002Fcopilot\u002Fplan\u002Fmicrosoft-365-copilot-privacy",[256],"nofollow",[258],{"type":46,"value":259},"Microsoft Learn「Microsoft Copilot のデータ、プライバシー、セキュリティ」",{"type":46,"value":261},"）",{"type":41,"tag":126,"props":263,"children":264},{},[265,270],{"type":41,"tag":153,"props":266,"children":267},{},[268],{"type":46,"value":269},"Azure で提供されるモデル（Azure OpenAI など）",{"type":41,"tag":153,"props":271,"children":272},{},[273,275,282],{"type":46,"value":274},"プロンプトと出力は OpenAI などモデルの提供元に渡らず、許可なく基盤モデルの学習に使われない。不正利用の監視で人の確認が必要になったときのために保存される場合があり、承認を受けた顧客はこの保存を止める申請ができる（出典：",{"type":41,"tag":252,"props":276,"children":279},{"href":277,"rel":278},"https:\u002F\u002Flearn.microsoft.com\u002Fja-jp\u002Fazure\u002Ffoundry\u002Fresponsible-ai\u002Fopenai\u002Fdata-privacy",[256],[280],{"type":46,"value":281},"Microsoft Learn「Microsoft Foundry における Azure が販売する Foundry モデルのデータ、プライバシー、セキュリティ」",{"type":46,"value":261},{"type":41,"tag":126,"props":284,"children":285},{},[286,291],{"type":41,"tag":153,"props":287,"children":288},{},[289],{"type":46,"value":290},"OpenAI API",{"type":41,"tag":153,"props":292,"children":293},{},[294,296,303],{"type":46,"value":295},"明示的に許可しない限り学習に使われない。不正利用の監視のログは既定で最大30日保存される（出典：",{"type":41,"tag":252,"props":297,"children":300},{"href":298,"rel":299},"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fyour-data",[256],[301],{"type":46,"value":302},"OpenAI「Data controls in the OpenAI platform」",{"type":46,"value":261},{"type":41,"tag":126,"props":305,"children":306},{},[307,312],{"type":41,"tag":153,"props":308,"children":309},{},[310],{"type":46,"value":311},"Anthropic の法人向け（API・Claude for Work）",{"type":41,"tag":153,"props":313,"children":314},{},[315,317,324],{"type":46,"value":316},"既定では入力と出力を学習に使わない。評価ボタンなどでフィードバックを送った場合は例外（出典：",{"type":41,"tag":252,"props":318,"children":321},{"href":319,"rel":320},"https:\u002F\u002Fprivacy.claude.com\u002Fen\u002Farticles\u002F7996868-is-my-data-used-for-model-training",[256],[322],{"type":46,"value":323},"Anthropic「Is my data used for model training?」",{"type":46,"value":261},{"type":41,"tag":126,"props":326,"children":327},{},[328,333],{"type":41,"tag":153,"props":329,"children":330},{},[331],{"type":46,"value":332},"Google Workspace の Gemini",{"type":41,"tag":153,"props":334,"children":335},{},[336,338,345],{"type":46,"value":337},"許可なくドメインの外で生成AIモデルの学習に使われることはない（出典：",{"type":41,"tag":252,"props":339,"children":342},{"href":340,"rel":341},"https:\u002F\u002Fknowledge.workspace.google.com\u002Fadmin\u002Fgemini\u002Fgenerative-ai-in-google-workspace-privacy-hub?hl=ja",[256],[343],{"type":46,"value":344},"Google「Google Workspace の生成 AI に関するプライバシー ハブ」",{"type":46,"value":261},{"type":41,"tag":42,"props":347,"children":348},{},[349],{"type":46,"value":350},"一方で、表のとおり「一定期間保存される」「設定によって処理される地域が変わる」といった条件もあります。たとえば Azure では、「グローバル」と付くデプロイの種類を選ぶと、モデルが置かれた任意の地域でプロンプトが処理される場合があります（出典は表と同じ Microsoft Learn のページ）。社内の決まりや取引先との契約で「社外に出さない」と決めている情報は、学習の有無とは別に扱いを決める必要があります。",{"type":41,"tag":42,"props":352,"children":353},{},[354,356,363],{"type":46,"value":355},"個人情報を含む場合は、法律の面からの確認も要ります。個人情報保護委員会は、生成AIに個人データを含むプロンプトを入力し、それが応答の出力以外の目的で扱われる場合は法律違反になる可能性があるとして、提供事業者が機械学習に利用しないことなどを十分に確認するよう注意を促しています（出典：",{"type":41,"tag":252,"props":357,"children":360},{"href":358,"rel":359},"https:\u002F\u002Fwww.ppc.go.jp\u002Fnews\u002Fcareful_information\u002F230602_AI_utilize_alert\u002F",[256],[361],{"type":46,"value":362},"個人情報保護委員会「生成AIサービスの利用に関する注意喚起等について」",{"type":46,"value":364},"、2023年6月2日公表、2026年10月確認）。",{"type":41,"tag":67,"props":366,"children":367},{},[368,376],{"type":41,"tag":42,"props":369,"children":370},{},[371],{"type":41,"tag":54,"props":372,"children":373},{},[374],{"type":46,"value":375},"K.S.Rogers の現場から",{"type":41,"tag":42,"props":377,"children":378},{},[379],{"type":46,"value":380},"機密性の高い技術情報を扱う製造業の会社様では、ディープリサーチのように調べものを任せる使い方も、RAG を組んで社内の資料を検索させる使い方も、入力したプロンプトをそのままクラウドの AI に送るのは難しい、という結論になりました。質問の文章に会社固有の用語や情報が入り、それがプロンプトを通じて外に出てしまう可能性があるためです。そこで、送る前に上長がプロンプトを確かめる案と、会社固有の情報を自動で見つけて伏せる案の2つを検討しました。",{"type":41,"tag":106,"props":382,"children":384},{"id":383},"選択肢は3つ費用と手間の順",[385],{"type":46,"value":386},"選択肢は3つ（費用と手間の順）",{"type":41,"tag":42,"props":388,"children":389},{},[390],{"type":46,"value":391},"社外に出せない情報を扱いながら生成AIを使う方法は、次の3つに分けられます。上から順に、費用と手間が大きくなります。",{"type":41,"tag":393,"props":394,"children":395},"picture",{},[396,398,406,407],{"type":46,"value":397},"\n  ",{"type":41,"tag":399,"props":400,"children":405},"source",{"media":401,"srcSet":402,"width":403,"height":404},"(max-width: 768px)","\u002Fimg\u002Fcontents\u002Farticles\u002Fshagai-ni-dasenai-joho-seisei-ai\u002Fsentakushi-3-sp.svg",420,620,[],{"type":46,"value":397},{"type":41,"tag":408,"props":409,"children":415},"img",{"src":410,"alt":411,"width":412,"height":413,"loading":414},"\u002Fimg\u002Fcontents\u002Farticles\u002Fshagai-ni-dasenai-joho-seisei-ai\u002Fsentakushi-3.svg","社外に出せない情報で生成AIを使う3つの選択肢：①上長の事前承認は費用が小さいが承認待ちで使われなくなりやすい、②自動マスキングは伏せ漏れと伏せすぎが起きる、③オンプレミスのLLMは情報を社外に出さないがGPUの費用が大きい",800,380,"lazy",[],{"type":41,"tag":417,"props":418,"children":420},"h3",{"id":419},"送る前に人が確かめる上長の事前承認",[421],{"type":46,"value":422},"① 送る前に人が確かめる（上長の事前承認）",{"type":41,"tag":42,"props":424,"children":425},{},[426],{"type":46,"value":427},"利用者がプロンプトを書いたら、上長が中身を確かめ、問題がなければ送る、という運用です。仕組みを作る費用はほとんどかからず、明日からでも始められます。",{"type":41,"tag":42,"props":429,"children":430},{},[431],{"type":46,"value":432},"ただし、運用の負担は軽くありません。",{"type":41,"tag":79,"props":434,"children":435},{},[436,446,456],{"type":41,"tag":83,"props":437,"children":438},{},[439,444],{"type":41,"tag":54,"props":440,"children":441},{},[442],{"type":46,"value":443},"承認待ちで使われなくなる",{"type":46,"value":445}," ―― 生成AIの良さは、思いついたときにすぐ試せることです。1回ごとに承認を待つと、利用者は使うのをやめてしまいます",{"type":41,"tag":83,"props":447,"children":448},{},[449,454],{"type":41,"tag":54,"props":450,"children":451},{},[452],{"type":46,"value":453},"上長も判断に迷う",{"type":46,"value":455}," ―― 何が「会社固有の情報」に当たるかの基準がないと、上長ごとに判断がばらつきます",{"type":41,"tag":83,"props":457,"children":458},{},[459,464],{"type":41,"tag":54,"props":460,"children":461},{},[462],{"type":46,"value":463},"量が増えると形だけになる",{"type":46,"value":465}," ―― 承認の件数が増えると、中身を読まずに通す状態になりがちです",{"type":41,"tag":42,"props":467,"children":468},{},[469,471,476],{"type":46,"value":470},"そのため、事前承認は",{"type":41,"tag":54,"props":472,"children":473},{},[474],{"type":46,"value":475},"対象を絞って使う",{"type":46,"value":477},"のが現実的です。",{"type":41,"tag":118,"props":479,"children":480},{},[481,497],{"type":41,"tag":122,"props":482,"children":483},{},[484],{"type":41,"tag":126,"props":485,"children":486},{},[487,492],{"type":41,"tag":130,"props":488,"children":489},{},[490],{"type":46,"value":491},"考え方",{"type":41,"tag":130,"props":493,"children":494},{},[495],{"type":46,"value":496},"具体例",{"type":41,"tag":146,"props":498,"children":499},{},[500,513,526,539],{"type":41,"tag":126,"props":501,"children":502},{},[503,508],{"type":41,"tag":153,"props":504,"children":505},{},[506],{"type":46,"value":507},"業務で絞る",{"type":41,"tag":153,"props":509,"children":510},{},[511],{"type":46,"value":512},"社外に出せない資料に触れる業務（設計・開発・顧客の技術相談など）だけを承認の対象にする",{"type":41,"tag":126,"props":514,"children":515},{},[516,521],{"type":41,"tag":153,"props":517,"children":518},{},[519],{"type":46,"value":520},"毎回ではなく「型」を承認する",{"type":41,"tag":153,"props":522,"children":523},{},[524],{"type":46,"value":525},"よく使うプロンプトの型（ひな形）を一度承認し、型の範囲で使うときは承認を省く",{"type":41,"tag":126,"props":527,"children":528},{},[529,534],{"type":41,"tag":153,"props":530,"children":531},{},[532],{"type":46,"value":533},"それ以外は事後の確認にする",{"type":41,"tag":153,"props":535,"children":536},{},[537],{"type":46,"value":538},"一般的な調べものや文章の手直しは、利用の記録を後から抜き取りで確かめる",{"type":41,"tag":126,"props":540,"children":541},{},[542,547],{"type":41,"tag":153,"props":543,"children":544},{},[545],{"type":46,"value":546},"判断の基準を書き出す",{"type":41,"tag":153,"props":548,"children":549},{},[550],{"type":46,"value":551},"入れてはいけない言葉（製品の開発名、取引先名など）の一覧を作り、上長と利用者の両方が見られるようにする",{"type":41,"tag":42,"props":553,"children":554},{},[555,557,563],{"type":46,"value":556},"最後の「入れてはいけない言葉の一覧」は、次の自動マスキングの辞書にもそのまま使えます。入力してよい情報の区分の決め方は、",{"type":41,"tag":252,"props":558,"children":560},{"href":559},"\u002Farticles\u002Fseisei-ai-shanai-rule-sample\u002F",[561],{"type":46,"value":562},"生成AIの社内ルールの作り方",{"type":46,"value":564},"で詳しく扱っています。",{"type":41,"tag":417,"props":566,"children":568},{"id":567},"固有の情報を自動で伏せる自動マスキング",[569],{"type":46,"value":570},"② 固有の情報を自動で伏せる（自動マスキング）",{"type":41,"tag":42,"props":572,"children":573},{},[574],{"type":46,"value":575},"プロンプトを送る前に、会社固有の情報を仕組みで見つけて伏せ字や仮の名前に置き換え、AI の回答が返ってきたら元に戻す方法です。仕組みは大きく3つの部品でできています。",{"type":41,"tag":118,"props":577,"children":578},{},[579,600],{"type":41,"tag":122,"props":580,"children":581},{},[582],{"type":41,"tag":126,"props":583,"children":584},{},[585,590,595],{"type":41,"tag":130,"props":586,"children":587},{},[588],{"type":46,"value":589},"部品",{"type":41,"tag":130,"props":591,"children":592},{},[593],{"type":46,"value":594},"やること",{"type":41,"tag":130,"props":596,"children":597},{},[598],{"type":46,"value":599},"限界",{"type":41,"tag":146,"props":601,"children":602},{},[603,621,639],{"type":41,"tag":126,"props":604,"children":605},{},[606,611,616],{"type":41,"tag":153,"props":607,"children":608},{},[609],{"type":46,"value":610},"辞書・パターンで伏せる",{"type":41,"tag":153,"props":612,"children":613},{},[614],{"type":46,"value":615},"社内用語の辞書、品番や図面番号の書式、メールアドレスや電話番号の形に当てはまる文字を置き換える",{"type":41,"tag":153,"props":617,"children":618},{},[619],{"type":46,"value":620},"辞書にない新しい用語や表記ゆれ（略称・英語表記など）を見逃す",{"type":41,"tag":126,"props":622,"children":623},{},[624,629,634],{"type":41,"tag":153,"props":625,"children":626},{},[627],{"type":46,"value":628},"AI で固有名を見つける",{"type":41,"tag":153,"props":630,"children":631},{},[632],{"type":46,"value":633},"文章から人名・社名・製品名らしい言葉を見つけて置き換える",{"type":41,"tag":153,"props":635,"children":636},{},[637],{"type":46,"value":638},"見つける AI 自体を社外のサービスに頼ると本末転倒。社内で動かす必要がある。精度は100%にならない",{"type":41,"tag":126,"props":640,"children":641},{},[642,647,652],{"type":41,"tag":153,"props":643,"children":644},{},[645],{"type":46,"value":646},"伏せたあとに戻す",{"type":41,"tag":153,"props":648,"children":649},{},[650],{"type":46,"value":651},"「元の言葉と仮の名前」の対応表を手元だけに持ち、回答の中の仮の名前を元に戻す",{"type":41,"tag":153,"props":653,"children":654},{},[655],{"type":46,"value":656},"AI が仮の名前を言い換えたり省略したりすると、正しく戻せない",{"type":41,"tag":42,"props":658,"children":659},{},[660],{"type":46,"value":661},"どの方法でも、次の2つの失敗が起きます。",{"type":41,"tag":79,"props":663,"children":664},{},[665,675],{"type":41,"tag":83,"props":666,"children":667},{},[668,673],{"type":41,"tag":54,"props":669,"children":670},{},[671],{"type":46,"value":672},"伏せ漏れ",{"type":46,"value":674}," ―― 伏せるべき言葉が残る。単語は伏せても、文脈や数値の組み合わせから中身が推測できることもあります",{"type":41,"tag":83,"props":676,"children":677},{},[678,683],{"type":41,"tag":54,"props":679,"children":680},{},[681],{"type":46,"value":682},"伏せすぎ",{"type":46,"value":684}," ―― 一般的な技術用語まで伏せてしまい、AI が質問の意味を取れず回答の質が落ちる。質問したいこと自体が会社固有の技術である場合は、伏せると質問が成り立ちません",{"type":41,"tag":42,"props":686,"children":687},{},[688],{"type":46,"value":689},"つまり、自動マスキングは「伏せれば社外に出してよい」と社内で決めた情報に使う方法で、技術の中身そのものが機密の場合には向きません。辞書の更新を誰が担うかも最初に決めておきます。",{"type":41,"tag":42,"props":691,"children":692},{},[693,695,701],{"type":46,"value":694},"開発の現場では、AI ツールに読ませないファイルを設定で決める方法もあります。設定で防げることと防げないことの実測は、",{"type":41,"tag":252,"props":696,"children":698},{"href":697},"\u002Farticles\u002Fclaude-code-deny-settei\u002F",[699],{"type":46,"value":700},"Claude Codeの秘密情報対策",{"type":46,"value":702},"にまとめています。",{"type":41,"tag":417,"props":704,"children":706},{"id":705},"社内で動く-aiオンプレミスの-llm",[707],{"type":46,"value":708},"③ 社内で動く AI（オンプレミスの LLM）",{"type":41,"tag":42,"props":710,"children":711},{},[712],{"type":46,"value":713},"自社の設備の中で LLM と検索の仕組みを動かし、情報を社外に送らない方法です。情報の扱いでは最も安心できますが、費用と手間は最も大きくなります。",{"type":41,"tag":42,"props":715,"children":716},{},[717,719,724,726,733],{"type":46,"value":718},"オンプレを検討する前に、",{"type":41,"tag":54,"props":720,"children":721},{},[722],{"type":46,"value":723},"自社が契約しているクラウドの中で閉じる構成",{"type":46,"value":725},"で足りないかを確かめます。たとえば Azure では、AI のサービスをインターネットから切り離し、自社の仮想ネットワークからプライベート エンドポイント経由でのみ使えるように構成できます（出典：",{"type":41,"tag":252,"props":727,"children":730},{"href":728,"rel":729},"https:\u002F\u002Flearn.microsoft.com\u002Fja-jp\u002Fazure\u002Fai-services\u002Fcognitive-services-virtual-networks",[256],[731],{"type":46,"value":732},"Microsoft Learn「Foundry Tools の仮想ネットワークを構成する」",{"type":46,"value":734},"、2026年10月確認）。情報システムの担当の方には、社内のネットワークとどこが分かれていて、どこを通るかを1枚の図にして説明すると、話が早く進みます。",{"type":41,"tag":67,"props":736,"children":737},{},[738,745],{"type":41,"tag":42,"props":739,"children":740},{},[741],{"type":41,"tag":54,"props":742,"children":743},{},[744],{"type":46,"value":375},{"type":41,"tag":42,"props":746,"children":747},{},[748],{"type":46,"value":749},"図面を社外のクラウドに出しにくいという製造業の会社様では、すでにお使いの Microsoft の環境の中で完結する構成にすることで、検討が前に進みました。一方で、その環境に当社がゲストとして招待されることは認められませんでした。そのため、Power Automate で作ったフローも、当社から直接つないで設定することはできません。当社の環境で作ったフローをエクスポートし、お客様のパソコンからインポートする形になりました。お客様の環境の中で閉じる構成では、作ったものを移す手間が課題になりやすいため、移し方まで最初に決めておくことをおすすめしています。",{"type":41,"tag":42,"props":751,"children":752},{},[753],{"type":46,"value":754},"それでもクラウドが使えないのは、次のような場合です。",{"type":41,"tag":79,"props":756,"children":757},{},[758,768,778],{"type":41,"tag":83,"props":759,"children":760},{},[761,766],{"type":41,"tag":54,"props":762,"children":763},{},[764],{"type":46,"value":765},"契約や業界の決まりで、クラウドでの処理そのものが認められない",{"type":46,"value":767}," ―― 取引先との契約や、扱う技術情報の決まりによっては、学習の有無に関係なくクラウドに置けません",{"type":41,"tag":83,"props":769,"children":770},{},[771,776],{"type":41,"tag":54,"props":772,"children":773},{},[774],{"type":46,"value":775},"医療機関の患者の情報",{"type":46,"value":777}," ―― 患者の情報を外に出せないため、オンプレで動く AI が前提になるという相談がありました",{"type":41,"tag":83,"props":779,"children":780},{},[781,786,788,795],{"type":41,"tag":54,"props":782,"children":783},{},[784],{"type":46,"value":785},"自治体",{"type":46,"value":787}," ―― 自治体のネットワークは、マイナンバー利用事務系・LGWAN 接続系・インターネット接続系に分けて運用されています（出典：",{"type":41,"tag":252,"props":789,"children":792},{"href":790,"rel":791},"https:\u002F\u002Fwww.soumu.go.jp\u002Fmain_content\u002F001000932.pdf",[256],[793],{"type":46,"value":794},"総務省「地方公共団体における情報セキュリティポリシーに関するガイドライン（令和7年3月版）」",{"type":46,"value":796},"、2026年10月確認）。業務で扱う情報によっては LGWAN の中で動かす必要があり、庁内で動く LLM が求められることが多くなります",{"type":41,"tag":67,"props":798,"children":799},{},[800,807],{"type":41,"tag":42,"props":801,"children":802},{},[803],{"type":41,"tag":54,"props":804,"children":805},{},[806],{"type":46,"value":375},{"type":41,"tag":42,"props":808,"children":809},{},[810],{"type":46,"value":811},"自治体からのご相談では、LGWAN の中に構築する必要があるため、ローカルで動く LLM を求められるケースが多くあります。ただ、LGWAN の環境に GPU が載っているわけではないので、GPU マシンを購入していただく話になりがちです。200〜300万円ほどかかることもあり、投資対効果の判断が難しくなりやすい傾向があります。一方で、Microsoft や AWS と提携している自治体もあり、別の自治体とお話しする中では「Azure なら」「AWS なら」使ってよい、というケースもありました。最初に「どういう環境なら使えますか」と伺うのも、一つの手だと考えています。",{"type":41,"tag":42,"props":813,"children":814},{},[815],{"type":46,"value":816},"オンプレにする場合は、GPU を積んだマシン、公開されているモデルの中からの選定（モデルごとに商用利用の条件が違うため、ライセンスも確かめます）、社内の資料を検索する仕組み、動かし続けるための保守が必要です。最新のクラウドの AI と比べて回答の質が下がることもあるため、対象の業務で足りるかを小さく試してから広げます。",{"type":41,"tag":106,"props":818,"children":820},{"id":819},"取り込む資料の利用規約を先に確かめる",[821],{"type":46,"value":822},"取り込む資料の「利用規約」を先に確かめる",{"type":41,"tag":42,"props":824,"children":825},{},[826,828,833],{"type":46,"value":827},"RAG を組むときに見落とされやすいのが、",{"type":41,"tag":54,"props":829,"children":830},{},[831],{"type":46,"value":832},"取り込む資料の権利",{"type":46,"value":834},"です。社内で作った資料なら問題になりにくいのですが、外から入手した資料には、それぞれ利用の条件があります。",{"type":41,"tag":118,"props":836,"children":837},{},[838,854],{"type":41,"tag":122,"props":839,"children":840},{},[841],{"type":41,"tag":126,"props":842,"children":843},{},[844,849],{"type":41,"tag":130,"props":845,"children":846},{},[847],{"type":46,"value":848},"資料の種類",{"type":41,"tag":130,"props":850,"children":851},{},[852],{"type":46,"value":853},"確かめること",{"type":41,"tag":146,"props":855,"children":856},{},[857,870,883,896],{"type":41,"tag":126,"props":858,"children":859},{},[860,865],{"type":41,"tag":153,"props":861,"children":862},{},[863],{"type":46,"value":864},"購入した規格・技術資料・有料のデータベース",{"type":41,"tag":153,"props":866,"children":867},{},[868],{"type":46,"value":869},"利用規約やライセンスで、AI の学習への利用、複製、データベース化、社外での処理が禁じられていないか",{"type":41,"tag":126,"props":871,"children":872},{},[873,878],{"type":41,"tag":153,"props":874,"children":875},{},[876],{"type":46,"value":877},"取引先から受け取った資料",{"type":41,"tag":153,"props":879,"children":880},{},[881],{"type":46,"value":882},"秘密保持契約の範囲（目的外の利用の禁止、社外の処理の可否、返却・廃棄の義務）",{"type":41,"tag":126,"props":884,"children":885},{},[886,891],{"type":41,"tag":153,"props":887,"children":888},{},[889],{"type":46,"value":890},"Web で公開されている資料",{"type":41,"tag":153,"props":892,"children":893},{},[894],{"type":46,"value":895},"サイトの利用規約、転載・機械的な収集の扱い",{"type":41,"tag":126,"props":897,"children":898},{},[899,904],{"type":41,"tag":153,"props":900,"children":901},{},[902],{"type":46,"value":903},"社内で作った資料",{"type":41,"tag":153,"props":905,"children":906},{},[907],{"type":46,"value":908},"社外秘などの区分と、見られる人の範囲（検索の仕組みで、見せてはいけない人に出ないか）",{"type":41,"tag":42,"props":910,"children":911},{},[912,914,919,921,927],{"type":46,"value":913},"RAG への取り込みが規約の「学習」に当たるかどうかは、文言によって解釈が分かれることがあります。迷う場合は自社だけで判断せず、発行元や社内の法務・知財の担当に確かめ、",{"type":41,"tag":54,"props":915,"children":916},{},[917],{"type":46,"value":918},"取り込む資料の一覧と、確かめた結果",{"type":46,"value":920},"を残しておきます。開発会社に渡す資料の扱いは、",{"type":41,"tag":252,"props":922,"children":924},{"href":923},"\u002Farticles\u002Fkaihatsu-gaichu-ai-riyou-rule\u002F",[925],{"type":46,"value":926},"システム開発を外注するときのAI利用ルール",{"type":46,"value":928},"も参考になります。",{"type":41,"tag":67,"props":930,"children":931},{},[932,939],{"type":41,"tag":42,"props":933,"children":934},{},[935],{"type":41,"tag":54,"props":936,"children":937},{},[938],{"type":46,"value":375},{"type":41,"tag":42,"props":940,"children":941},{},[942],{"type":46,"value":943},"機密性の高い技術情報を扱う製造業の会社様では、その分野ならではの規格の資料を RAG に取り込みたいという要望がありました。ところが、資料の利用規約を確かめると「AI の学習に使ってはいけない」とされているものがありました。何を集めて取り込むかによって作業の量も変わるため、取り込む対象の権利の条件を最初にお客様と合意しておくと、見積もりがぶれにくくなります。",{"type":41,"tag":106,"props":945,"children":947},{"id":946},"オンプレにすると-gpu-が費用の中心になる-pl-で払えるかを先に試算する",[948],{"type":46,"value":949},"オンプレにすると GPU が費用の中心になる ―― P\u002FL で払えるかを先に試算する",{"type":41,"tag":42,"props":951,"children":952},{},[953],{"type":46,"value":954},"オンプレの LLM を選ぶと、費用の中心は GPU を積んだマシンになります。費用の項目は次のとおりです。",{"type":41,"tag":118,"props":956,"children":957},{},[958,979],{"type":41,"tag":122,"props":959,"children":960},{},[961],{"type":41,"tag":126,"props":962,"children":963},{},[964,969,974],{"type":41,"tag":130,"props":965,"children":966},{},[967],{"type":46,"value":968},"項目",{"type":41,"tag":130,"props":970,"children":971},{},[972],{"type":46,"value":973},"中身",{"type":41,"tag":130,"props":975,"children":976},{},[977],{"type":46,"value":978},"試算のときの注意",{"type":41,"tag":146,"props":980,"children":981},{},[982,1000,1018,1036],{"type":41,"tag":126,"props":983,"children":984},{},[985,990,995],{"type":41,"tag":153,"props":986,"children":987},{},[988],{"type":46,"value":989},"GPU を積んだマシン",{"type":41,"tag":153,"props":991,"children":992},{},[993],{"type":46,"value":994},"購入、または月額で借りる",{"type":41,"tag":153,"props":996,"children":997},{},[998],{"type":46,"value":999},"使う人数と、同時に使う量で必要な台数が変わる",{"type":41,"tag":126,"props":1001,"children":1002},{},[1003,1008,1013],{"type":41,"tag":153,"props":1004,"children":1005},{},[1006],{"type":46,"value":1007},"設置と電源",{"type":41,"tag":153,"props":1009,"children":1010},{},[1011],{"type":46,"value":1012},"置き場所、電源、冷却",{"type":41,"tag":153,"props":1014,"children":1015},{},[1016],{"type":46,"value":1017},"社内のサーバー室に入るか、電源の容量が足りるかを先に確かめる",{"type":41,"tag":126,"props":1019,"children":1020},{},[1021,1026,1031],{"type":41,"tag":153,"props":1022,"children":1023},{},[1024],{"type":46,"value":1025},"検索の仕組みの構築",{"type":41,"tag":153,"props":1027,"children":1028},{},[1029],{"type":46,"value":1030},"資料の取り込み、検索、利用画面",{"type":41,"tag":153,"props":1032,"children":1033},{},[1034],{"type":46,"value":1035},"取り込む資料の量と、利用規約の確認の手間で変わる",{"type":41,"tag":126,"props":1037,"children":1038},{},[1039,1044,1049],{"type":41,"tag":153,"props":1040,"children":1041},{},[1042],{"type":46,"value":1043},"保守と更新",{"type":41,"tag":153,"props":1045,"children":1046},{},[1047],{"type":46,"value":1048},"モデルの入れ替え、障害の対応、利用者の問い合わせ",{"type":41,"tag":153,"props":1050,"children":1051},{},[1052],{"type":46,"value":1053},"社内で担う人がいるか。いなければ外部に任せる費用がかかる",{"type":41,"tag":42,"props":1055,"children":1056},{},[1057,1059,1064],{"type":46,"value":1058},"金額は、使う人数・同時に使う量・求める回答の質で大きく変わります。そのため、",{"type":41,"tag":54,"props":1060,"children":1061},{},[1062],{"type":46,"value":1063},"購入する場合と月額で借りる場合の両方",{"type":46,"value":1065},"を試算し、使い始めの規模と、広げたときの規模の2通りで比べます。",{"type":41,"tag":42,"props":1067,"children":1068},{},[1069,1071,1076,1078,1084],{"type":46,"value":1070},"効果の側は、減らせる残業代や外注費、増員せずに済む人件費など、",{"type":41,"tag":54,"props":1072,"children":1073},{},[1074],{"type":46,"value":1075},"P\u002FL（損益）にどれだけ効くか",{"type":46,"value":1077},"で数え、毎月の費用を払えるかを判断します。届かない場合は、対象の業務や利用者を絞る、オンプレは「クラウドに出せない資料」の検索だけにする、といった組み合わせを考えます。費用対効果の数え方は、",{"type":41,"tag":252,"props":1079,"children":1081},{"href":1080},"\u002Farticles\u002Fai-donyu-shien-hiyou-erabikata\u002F",[1082],{"type":46,"value":1083},"AI導入の費用相場",{"type":46,"value":564},{"type":41,"tag":67,"props":1086,"children":1087},{},[1088,1095],{"type":41,"tag":42,"props":1089,"children":1090},{},[1091],{"type":41,"tag":54,"props":1092,"children":1093},{},[1094],{"type":46,"value":375},{"type":41,"tag":42,"props":1096,"children":1097},{},[1098],{"type":46,"value":1099},"先の製造業の会社様では、プロンプトをクラウドに送れず、取り込みたい資料にも利用の制限があったことから、社内で動く LLM を作る方向で検討が進みました。そこで最大の論点になったのが、GPU マシンの費用です。費用が大きいため、技術的にできるかどうかより先に、その費用を P\u002FL で払えるかどうかが判断の分かれ目になりました。オンプレを検討するときは、構成を決める前に費用と効果の試算をそろえることをおすすめしています。",{"type":41,"tag":106,"props":1101,"children":1103},{"id":1102},"どれを選ぶかの判断の軸",[1104],{"type":46,"value":1102},{"type":41,"tag":42,"props":1106,"children":1107},{},[1108],{"type":46,"value":1109},"どの方法が合うかは、情報の区分と、使う量と、払える費用で決まります。",{"type":41,"tag":393,"props":1111,"children":1112},{},[1113,1114,1119,1120],{"type":46,"value":397},{"type":41,"tag":399,"props":1115,"children":1118},{"media":401,"srcSet":1116,"width":403,"height":1117},"\u002Fimg\u002Fcontents\u002Farticles\u002Fshagai-ni-dasenai-joho-seisei-ai\u002Fhandan-flow-sp.svg",688,[],{"type":46,"value":397},{"type":41,"tag":408,"props":1121,"children":1125},{"src":1122,"alt":1123,"width":412,"height":1124,"loading":414},"\u002Fimg\u002Fcontents\u002Farticles\u002Fshagai-ni-dasenai-joho-seisei-ai\u002Fhandan-flow.svg","選び方の流れ：契約や決まりでクラウドでの処理が認められない情報なら、GPUの費用をP\u002FLで試算してオンプレを検討する。会社契約のクラウドで閉じれば扱える情報なら、閉じた構成と権限の整備で使う。伏せれば出せる情報なら、自動マスキングと事前承認を組み合わせる。伏せても出せない情報は、オンプレと同じく費用を試算して判断する",432,[],{"type":41,"tag":118,"props":1127,"children":1128},{},[1129,1155],{"type":41,"tag":122,"props":1130,"children":1131},{},[1132],{"type":41,"tag":126,"props":1133,"children":1134},{},[1135,1140,1145,1150],{"type":41,"tag":130,"props":1136,"children":1137},{},[1138],{"type":46,"value":1139},"判断の軸",{"type":41,"tag":130,"props":1141,"children":1142},{},[1143],{"type":46,"value":1144},"① 事前承認",{"type":41,"tag":130,"props":1146,"children":1147},{},[1148],{"type":46,"value":1149},"② 自動マスキング",{"type":41,"tag":130,"props":1151,"children":1152},{},[1153],{"type":46,"value":1154},"③ オンプレの LLM",{"type":41,"tag":146,"props":1156,"children":1157},{},[1158,1181,1204,1227,1250],{"type":41,"tag":126,"props":1159,"children":1160},{},[1161,1166,1171,1176],{"type":41,"tag":153,"props":1162,"children":1163},{},[1164],{"type":46,"value":1165},"扱える情報",{"type":41,"tag":153,"props":1167,"children":1168},{},[1169],{"type":46,"value":1170},"人が確かめて出してよいもの",{"type":41,"tag":153,"props":1172,"children":1173},{},[1174],{"type":46,"value":1175},"伏せれば出してよいもの",{"type":41,"tag":153,"props":1177,"children":1178},{},[1179],{"type":46,"value":1180},"クラウドに出せないもの",{"type":41,"tag":126,"props":1182,"children":1183},{},[1184,1189,1194,1199],{"type":41,"tag":153,"props":1185,"children":1186},{},[1187],{"type":46,"value":1188},"向く使い方",{"type":41,"tag":153,"props":1190,"children":1191},{},[1192],{"type":46,"value":1193},"対象を絞った一部の業務",{"type":41,"tag":153,"props":1195,"children":1196},{},[1197],{"type":46,"value":1198},"毎日多くの人が使う業務",{"type":41,"tag":153,"props":1200,"children":1201},{},[1202],{"type":46,"value":1203},"特定の資料の検索や、特定の部署の業務",{"type":41,"tag":126,"props":1205,"children":1206},{},[1207,1212,1217,1222],{"type":41,"tag":153,"props":1208,"children":1209},{},[1210],{"type":46,"value":1211},"費用",{"type":41,"tag":153,"props":1213,"children":1214},{},[1215],{"type":46,"value":1216},"小さい（人の手間が中心）",{"type":41,"tag":153,"props":1218,"children":1219},{},[1220],{"type":46,"value":1221},"中くらい（仕組みづくりと辞書の更新）",{"type":41,"tag":153,"props":1223,"children":1224},{},[1225],{"type":46,"value":1226},"大きい（GPU と保守）",{"type":41,"tag":126,"props":1228,"children":1229},{},[1230,1235,1240,1245],{"type":41,"tag":153,"props":1231,"children":1232},{},[1233],{"type":46,"value":1234},"始めるまでの期間",{"type":41,"tag":153,"props":1236,"children":1237},{},[1238],{"type":46,"value":1239},"すぐ",{"type":41,"tag":153,"props":1241,"children":1242},{},[1243],{"type":46,"value":1244},"数週間〜",{"type":41,"tag":153,"props":1246,"children":1247},{},[1248],{"type":46,"value":1249},"試算・調達・構築が必要",{"type":41,"tag":126,"props":1251,"children":1252},{},[1253,1258,1262,1267],{"type":41,"tag":153,"props":1254,"children":1255},{},[1256],{"type":46,"value":1257},"主な弱点",{"type":41,"tag":153,"props":1259,"children":1260},{},[1261],{"type":46,"value":443},{"type":41,"tag":153,"props":1263,"children":1264},{},[1265],{"type":46,"value":1266},"伏せ漏れと伏せすぎ",{"type":41,"tag":153,"props":1268,"children":1269},{},[1270],{"type":46,"value":1271},"費用と、回答の質",{"type":41,"tag":42,"props":1273,"children":1274},{},[1275],{"type":46,"value":1276},"実際には、1つに決めるより組み合わせることが多くなります。たとえば、一般的な業務は会社契約のクラウドの AI で、会社固有の用語は伏せてから送り、クラウドに出せない資料の検索だけを社内で動かす、という形です。",{"type":41,"tag":42,"props":1278,"children":1279},{},[1280,1282,1288],{"type":46,"value":1281},"また、Microsoft 365 Copilot をすでに契約していても「機密の情報は入れられない」という理由で使われていない会社では、情報の区分と権限を整える（ガバナンス）だけでなく、その区分の中で使える業務の自動化をセットで用意すると、使われ方が変わります。Copilot の権限やラベルの設定は、",{"type":41,"tag":252,"props":1283,"children":1285},{"href":1284},"\u002Farticles\u002Fcopilot-security-settei-chusho\u002F",[1286],{"type":46,"value":1287},"Microsoft 365 Copilotのセキュリティ設定",{"type":46,"value":1289},"で整理しています。",{"type":41,"tag":106,"props":1291,"children":1293},{"id":1292},"ksrogers-の支援",[1294],{"type":46,"value":1295},"K.S.Rogers の支援",{"type":41,"tag":42,"props":1297,"children":1298},{},[1299],{"type":46,"value":1300},"K.S.Rogers は、機密性の高い情報を扱う会社の生成AIの活用を、情報の区分の整理から構成の選定、費用の試算まで一緒に進めています。",{"type":41,"tag":79,"props":1302,"children":1303},{},[1304,1314,1332],{"type":41,"tag":83,"props":1305,"children":1306},{},[1307,1312],{"type":41,"tag":54,"props":1308,"children":1309},{},[1310],{"type":46,"value":1311},"選択肢の整理と費用の試算",{"type":46,"value":1313}," ―― どの情報がクラウドに出せないか、どの業務に使うかを伺い、事前承認・自動マスキング・閉じたクラウド・オンプレのどれが合うかを、P\u002FL への効果とあわせて整理します",{"type":41,"tag":83,"props":1315,"children":1316},{},[1317,1322,1324,1330],{"type":41,"tag":54,"props":1318,"children":1319},{},[1320],{"type":46,"value":1321},"業務の自動化",{"type":46,"value":1323}," ―― AI業務自動化の導入支援サービス",{"type":41,"tag":252,"props":1325,"children":1327},{"href":1326},"\u002Fbusiness\u002Fopezero\u002F",[1328],{"type":46,"value":1329},"OpeZero",{"type":46,"value":1331},"では、最初にシステム環境を調べたうえで、構築から運用までを担います。料金は運用報酬型（初期構築費は原則0円・月額固定）です",{"type":41,"tag":83,"props":1333,"children":1334},{},[1335,1340,1342,1348],{"type":41,"tag":54,"props":1336,"children":1337},{},[1338],{"type":46,"value":1339},"社内で動く AI や検索の仕組みの開発",{"type":46,"value":1341}," ―― オンプレの LLM や RAG の構築は、",{"type":41,"tag":252,"props":1343,"children":1345},{"href":1344},"\u002Fbusiness\u002Fss\u002F",[1346],{"type":46,"value":1347},"システム開発",{"type":46,"value":1349},"として、範囲とデータを確認したうえで個別にお見積りします",{"type":41,"tag":42,"props":1351,"children":1352},{},[1353],{"type":46,"value":1354},"ご相談は全国からお受けしています（よく伺うのは関西・関東・名古屋）。情報の扱いのように社内の事情を詳しく伺う必要があるテーマは、面着でのお打ち合わせもおすすめしています。",{"type":41,"tag":106,"props":1356,"children":1358},{"id":1357},"まとめ",[1359],{"type":46,"value":1357},{"type":41,"tag":79,"props":1361,"children":1362},{},[1363,1368,1373,1378],{"type":41,"tag":83,"props":1364,"children":1365},{},[1366],{"type":46,"value":1367},"生成AIで情報が社外に出る経路は、プロンプト・取り込む資料・出力と履歴の3つです。「学習に使われない」と「社外に送らない」は分けて考えます",{"type":41,"tag":83,"props":1369,"children":1370},{},[1371],{"type":46,"value":1372},"選択肢は、①上長の事前承認 ②自動マスキング ③オンプレミスの LLM の3つです。事前承認は対象を絞り、マスキングは伏せ漏れと伏せすぎを前提に「伏せれば出せる情報」に使います",{"type":41,"tag":83,"props":1374,"children":1375},{},[1376],{"type":46,"value":1377},"RAG に取り込む資料は、利用規約やライセンスで AI での利用が禁じられていないかを先に確かめます",{"type":41,"tag":83,"props":1379,"children":1380},{},[1381],{"type":46,"value":1382},"オンプレにすると GPU が費用の中心になります。購入と月額の両方を試算し、P\u002FL で払えるかを構成を決める前に判断します",{"type":41,"tag":42,"props":1384,"children":1385},{},[1386,1388,1394],{"type":46,"value":1387},"「社外に出せない情報があり、生成AIをどこまで使えるか分からない」という段階から、",{"type":41,"tag":252,"props":1389,"children":1391},{"href":1390},"\u002Fcontact\u002F",[1392],{"type":46,"value":1393},"お問い合わせ",{"type":46,"value":1395},"よりご相談ください。",{"title":7,"searchDepth":1397,"depth":1397,"links":1398},2,[1399,1400,1406,1407,1408,1409,1410],{"id":108,"depth":1397,"text":111},{"id":383,"depth":1397,"text":386,"children":1401},[1402,1404,1405],{"id":419,"depth":1403,"text":422},3,{"id":567,"depth":1403,"text":570},{"id":705,"depth":1403,"text":708},{"id":819,"depth":1397,"text":822},{"id":946,"depth":1397,"text":949},{"id":1102,"depth":1397,"text":1102},{"id":1292,"depth":1397,"text":1295},{"id":1357,"depth":1397,"text":1357},"markdown","content:articles:shagai-ni-dasenai-joho-seisei-ai.md","content","articles\u002Fshagai-ni-dasenai-joho-seisei-ai.md","articles\u002Fshagai-ni-dasenai-joho-seisei-ai","md",[],1795014881147]