AI利用料で企業の3社に1社が予算超過、適切なモデル選択で費用抑制へ

業務ぎょうむで人工知能じんこうちのう（AI）を使つかう企業きぎょうが増ふえるなか、想定そうていを超こえる利用料りようりょうに頭あたまを抱かかえる現場げんばが広ひろがっています。調査対象ちょうさたいしょうとなった企業きぎょうの3社しゃに1社しゃが、すでに年間ねんかんのトークン料りょう予算よさんを超過ちょうかしていました。コンサルティング大手おおてのアクセンチュアが世界せかい17カ国こくの経営層けいえいそう750人にんを対象たいしょうに行おこなった調査ちょうさによると、今後こんご2年間ねんかんでトークン消費量しょうひりょうは78％増ふえる見通みとおしで、企業きぎょうの負担ふたんはさらに重おもくなる懸念けねんがあります。

As more companies use artificial intelligence (AI) in their business operations, workplaces struggling with usage fees that exceed expectations are becoming widespread. One in three surveyed companies had already exceeded its annual token budget. According to a survey of 750 executives across 17 countries conducted by consulting giant Accenture, token consumption is projected to increase by 78% over the next two years, raising concerns that the burden on companies will grow even heavier.

Vるなか= amid / as (describes an ongoing background circumstance under which another event unfolds)頭を抱える= to be deeply troubled / at one's wit's end (an idiom describing someone struggling with a difficult, headache-inducing problem)

実じつは、費用ひようの急増きゅうぞうを招まねいている大おおきな要因よういんは「過剰かじょうな性能せいのうのモデル」の利用りようです。アクセンチュアが米労働省べいろうどうしょうの職業情報しょくぎょうじょうほうデータベースをもとに9368件けんの業務ぎょうむタスクを分析ぶんせきしたところ、最先端さいせんたんの「フロンティアモデル」の能力のうりょくを真しんに必要ひつようとするタスクは全体ぜんたいの約やく1割わりにとどまりました。それにもかかわらず、実際じっさいにはAIへの依頼いらいの54％が必要以上ひつよういじょうの高性能こうせいのうモデルに送おくられており、簡単かんたんな作業さぎょうにまで高価こうかな処理能力しょりのうりょくが使つかわれている実態じったいが浮うき彫ぼりになっています。

In fact, a major factor driving the rapid increase in expenses is the use of 'overpowered models.' When Accenture analyzed 9,368 work tasks based on the U.S. Department of Labor's occupational information database, tasks genuinely requiring the capabilities of cutting-edge 'frontier models' accounted for only about 10% of the total. Nevertheless, 54% of actual requests to AI are sent to unnecessarily high-performance models, bringing to light a reality in which expensive processing power is being used even for simple tasks.

それにもかかわらず= nevertheless / in spite of that (a conjunction indicating that what follows contradicts what would be expected from the preceding fact)浮き彫りになる= to become clearly apparent / to be highlighted (an expression describing hidden problems or actual conditions becoming visibly evident)

こうした無駄むだを抑おさえる手立てだてとして注目ちゅうもくされているのが、作業さぎょうの内容ないように応おうじて最適さいてきなAIモデルへ自動じどうで振ふり分わける「モデルルーティング」です。経営層けいえいそうの多おおくがコスト削減さくげんに最もっとも重要じゅうようだと考かんがえているものの、実際じっさいに導入どうにゅうしている企業きぎょうは30％にとどまります。アクセンチュアの試算しさんによると、仕事しごとに合あわせたモデルの使つかい分わけや、部門ぶもんごとの使用責任しようせきにんの明確化めいかくかなどを徹底てっていすれば、今後こんご2年間ねんかんで見込みこまれるトークン費用ひようの増加ぞうかを最大さいだい44％抑おさえられる可能性かのうせいがあるといいます。

Gaining attention as a means to curb such waste is 'model routing,' which automatically directs tasks to the optimal AI model depending on the nature of the work. Although many executives consider this most important for cost reduction, only 30% of companies have actually implemented it. According to Accenture's estimates, thoroughly implementing measures such as selecting different models tailored to specific tasks and clarifying usage responsibility for each department could potentially reduce the anticipated growth in token costs over the next two years by up to 44%.

Nに応じて= depending on N / in response to N (indicates adapting or changing actions according to the nature, condition, or requirements of N)〜にとどまる= to remain at / to stay limited to (indicates that a quantity, level, or extent does not advance beyond a certain limit)

AIの利用料りようりょうを管理かんりすることは、単たんなる費用ひようの節約せつやくにとどまりません。今回こんかいの調査ちょうさでは、トークン料りょうの支出ししゅつが定量的ていりょうてきな事業成果じぎょうせいかに結むすびついていると証明しょうめいできている企業きぎょうは2割強わりきょうにすぎず、投資とうしに見合みあう効果こうかが得えられているか不透明ふとうめいな状況じょうきょうが続つづいています。高性能こうせいのうなモデルばかりを求もとめるのではなく、仕事しごとの難むずかしさに応おうじてモデルを選えらび、成果せいかにつなげる経営けいえいの仕組しくみを作つくれるかどうかが、企業きぎょうのAI活用かつようの成否せいひを分わけそうです。

Managing AI usage fees goes beyond mere cost savings. In this survey, just over 20% of companies were able to prove that their token spending is tied to quantitative business outcomes, and the situation remains uncertain as to whether results commensurate with the investment are being achieved. Rather than solely chasing high-performance models, whether companies can build management mechanisms that select models according to task difficulty and connect their use to results looks set to determine the success or failure of corporate AI adoption.

Nにすぎない= nothing more than N / merely N (emphasizes that something does not exceed a small number, level, or degree)Nに見合う= to correspond to N / to be commensurate with N (indicates being well-balanced with cost, effort, or value)

AI利用料で企業の3社に1社が予算超過、適切なモデル選択で費用抑制へ

Learning Notes

Grammar Patterns

Vocabulary

Language Insights

Practice

Pattern	Meaning
Vるなか	amid / as (describes an ongoing background circumstance under which another event unfolds)
頭を抱える	to be deeply troubled / at one's wit's end (an idiom describing someone struggling with a difficult, headache-inducing problem)
それにもかかわらず	nevertheless / in spite of that (a conjunction indicating that what follows contradicts what would be expected from the preceding fact)
浮き彫りになる	to become clearly apparent / to be highlighted (an expression describing hidden problems or actual conditions becoming visibly evident)
Nに応じて	depending on N / in response to N (indicates adapting or changing actions according to the nature, condition, or requirements of N)
〜にとどまる	to remain at / to stay limited to (indicates that a quantity, level, or extent does not advance beyond a certain limit)
Nにすぎない	nothing more than N / merely N (emphasizes that something does not exceed a small number, level, or degree)
Nに見合う	to correspond to N / to be commensurate with N (indicates being well-balanced with cost, effort, or value)

Word	Reading	Level	Meaning
超過	ちょうか	N2	excess；surplus；exceeding (a limit, estimate, etc.)；going over
過剰	かじょう	N2	excess；surplus；superabundance；overabundance
削減	さくげん	N1	cut；reduction；curtailment
懸念	けねん		worry；fear；anxiety；concern
試算	しさん		trial calculation；preliminary calculation；estimation
定量的	ていりょうてき		quantitative
成否	せいひ		success or failure；outcome；result