Agentic Coding評測：32%企業棄買軟體改自建，獲利卻沒漲 | Agentic Coding Review: 32% Skip Software, Profit Flat

By Kit 小克 | AI Tool Observer | 2026-09-03

🇹🇼 Agentic Coding評測：32%企業棄買軟體改自建，獲利卻沒漲

Agentic Coding（用 AI 代理寫代碼）正在改寫企業「買軟體」還是「自己做」的老問題。麥肯錫最新《State of AI 2026》調查訪問 1,719 位企業領袖，發現有 32% 的組織因為 Agentic Coding 工具的出現，直接放棄了原本要買的軟體，改成自己動手建置。聽起來像是軟體業的地震，但數字背後藏著一個更誠實的故事：獲利根本沒有等比例成長。

32% 企業放棄外購，但高手和後段班差很多

調查把企業分成「高績效者」（AI 貢獻至少 5% EBIT 的組織，僅佔 6%）和其他人。高績效者裡有將近 50% 選擇自建，一般企業只有 31%。產業別的差距也很大：

科技業：41% 選擇自建，領先所有產業
醫療、專業服務、能源：介於 36%–39%
保險（19%）、公部門（17%）：明顯落後，法規和舊系統包袱重

同時，年營收超過 10 億美元的大企業裡，已經在一個以上部門「規模化」導入 AI Agent 的比例，從去年的 27% 跳到 40%。

為什麼獲利沒漲？成本只是換了個地方藏

這是這份報告最誠實的部分：即使 Agentic Coding 用得這麼兇，整體「AI 貢獻 EBIT」的企業比例還是卡在 37% 沒動，高績效者也只有 6%。原因很簡單——省下的授權費，換成了：

維運成本：軟體生命週期裡維護就吃掉 60%–90% 的總成本，寫代碼只是小頭
資安風險：Veracode 2026 資安報告指出，AI 生成代碼的安全通過率平均只有 56%，未特別要求資安時，約 44% 的產出會帶著已知漏洞
舊系統反效果：METR 的對照實驗發現，資深工程師在成熟舊代碼庫上用 AI 工具，速度反而慢了 19%；Google DORA 也指出 AI 在全新專案能提速 35%–40%，但碰到複雜舊代碼只剩不到 10%

換句話說，企業不是真的省錢，是把「買軟體的錢」換成「養工程師、輪值 on-call、補資安漏洞」的錢，帳面上好看，實際上是轉移支付。

小克怎麼看：分場景，別一刀切

Agentic Coding 在「從零開始的新專案」上確實好用，這點數據不騙人。但如果你是想拿它去改一個跑了十年的舊系統，或是完全不做資安審查就上線，那 32% 放棄買軟體的企業裡，恐怕有不少人正在踩 Gartner 預測「2027 年底前 40% 的 Agentic AI 專案會被砍掉」的坑。自建前先問自己：這是新專案還是舊系統？資安審查排進流程了嗎？維運的人手夠嗎？想清楚這三題，再決定要不要跳過採購那張 PO。

好不好用，試了才知道。

🇺🇸 Agentic Coding Review: 32% Skip Software, Profit Flat

Agentic coding - using AI agents to write software - is rewriting the classic build-vs-buy decision. McKinsey's latest State of AI 2026 survey polled 1,719 business leaders and found that 32% of organizations skipped a planned software purchase because agentic coding tools made it feasible to build the thing themselves instead. That sounds like an earthquake for the software industry - until you look at the profit numbers, which tell a much more honest story.

32% Skip the Purchase, But the Gap Between Leaders and Laggards Is Huge

McKinsey splits respondents into "high performers" (organizations attributing at least 5% of EBIT to AI - just 6% of the sample) and everyone else. Nearly 50% of high performers chose to build in-house, versus 31% of the rest. The industry split is just as telling:

Technology: 41% build over buy, leading every sector
Healthcare, professional services, energy: 36%-39%
Insurance (19%) and public sector (17%): far behind, weighed down by regulation and legacy systems

Meanwhile, large enterprises above $1B in revenue scaling agents in one or more functions jumped from 27% to 40% year over year.

Why Didn't Profit Move? Costs Just Got Relocated

This is the most honest part of the report: despite the surge in agentic coding, the share of organizations reporting any real EBIT impact from AI stayed flat at 37%, with only 6% qualifying as high performers. The reason is simple - license fees saved just got shifted into:

Maintenance costs: 60%-90% of a software system's lifetime cost is maintenance, not the initial code
Security risk: Veracode's 2026 report found AI-generated code has an average security pass rate of only 56%, and about 44% of outputs ship with known vulnerabilities when security isn't explicitly requested
Legacy code backfires: METR's controlled trial found experienced developers were actually 19% slower using AI tools on mature codebases; Google's DORA research shows agentic coding gains of 35%-40% on greenfield tasks but under 10% on complex legacy code

In other words, companies aren't really saving money - they're converting "software license spend" into "headcount, on-call rotations, and security remediation" spend. It looks better on the balance sheet, but it's a transfer, not a saving.

Kit's Take: This Is a Scenario Question, Not a Blanket Yes

Agentic coding genuinely earns its keep on greenfield projects - the data backs that up. But if you're pointing it at a decade-old legacy system, or skipping security review entirely before shipping, you may be one of the 32% heading straight for what Gartner predicts: 40% of agentic AI projects canceled by end of 2027. Before you skip that purchase order, ask three questions: is this greenfield or legacy? Is security review built into the pipeline? Do you actually have the headcount to maintain what you build? Answer those honestly before deciding whether to build or buy.

You won't know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

Claude Code 實測：AI 幫你寫程式到底行不行？ | Claude Code Review: Can AI Really Code for You?

3月 26, 2026

閱讀完整內容

搜尋此網誌

AI小貼士