手记14

最近总有人问我：AI 都这么强了，为什么公司里还没见它真把谁的活儿顶掉？

我一开始也以为是 AI 不够聪明——会出错，会瞎编，不懂实际情况。但想得越深，越觉得不对。问题可能反过来：不是 AI 进不了公司，是公司还没准备好被 AI 看懂。

举个例子。你随手丢给 AI 一个看似简单的问题，比如"这件事出了岔子，怎么办？"它大概只会给你几句正确的废话：先沟通、再协调、通知相关方。一听就是外行话。

不是它笨，是这个问题背后真正的难处，它一个都不知道：这事到底卡在哪个环节？牵动的是哪一方的利益？哪条线得优先保？还能拖多久？这话能不能落到白纸黑字上？跟上面汇报，怎么说才不给自己留下被追责的口子？

而这些东西，公司里根本没写在任何制度文件里。它们藏在邮件、会议、表格、和老员工的脑子里，靠的是多年攒下来的默契和分寸。所以 AI 不是不懂，是这些门道从来没被人讲清楚过——连公司自己都说不清。

更扎心的一层在这儿。你真把公司的资料敞开给 AI 去读，它会读到一团乱：过时的旧文件还躺在共享盘上，"我看看"和"我同意"混在一堆对话里，同一份东西好几个版本都在用。所谓 AI 的"幻觉"，很多时候不是它在胡思乱想，而是它忠实地照出了公司自己那本说不清、理不顺的糊涂账。 不是 AI 出了问题，是公司自己都没把自己搞明白。

于是真正的坎来了：信任。

哪怕 AI 讲得头头是道，决策的人也得问一句：我凭什么信你？靠它说得漂亮就信，是不保险的；真正能信的，是它每一个判断都能翻回去——根据哪封邮件、哪份记录、哪个数据。 这跟公司信一个人是一个道理：信你，不是因为你从不犯错，是因为有流程、有审批、有记录、有人担责。AI 想进核心，也得先进这套体系，而不是靠"它好像很聪明"。

那往后会怎样？我的判断是：AI 不会均匀地替掉所有人，它会把人悄悄分成两拨。

一拨是干活的人——整理、催进度、写汇总、做记录、做翻译。这些会被大量压掉。

另一拨是担责的人——那个最后拍板、签字、要为判断负责的人。这拨人反而更值钱。

因为 AI 能替你做事，替不了你负责。你越要求"出了事得有人扛"，AI 就越顶不掉那个签字的位置。岗位不是整片蒸发，是两头分化：手在干的活越来越自动，肩上扛的责越来越重。

所以对个人来说，出路不是跟 AI 比谁手快——那条路必输。真正的位置在中间：既懂实际情况、又懂 AI，知道关键信息在哪、也知道责任的边界在哪，能把那些只可意会的经验，翻译成 AI 能用、也能被追溯的规则。 这种"业务和 AI 之间的翻译者"，短期内谁都替不了。

一句话收尾：

AI 落地，买工具是最容易的一步。难的是，公司愿不愿意先把自己那本糊涂账，理成一份 AI 能看懂、也扛得起追责的明白账。AI 真正改变公司的那天，不是因为它更聪明了，而是公司终于准备好，被看清楚了。


English

People keep asking me: AI is so capable now — why hasn't it actually taken over anyone's job inside companies yet?

At first I assumed AI just wasn't smart enough — it makes mistakes, it makes things up, it doesn't grasp the real situation. But the deeper I thought about it, the more it seemed backwards. The problem may run the other way: it's not that AI can't get into the company — it's that the company isn't ready to be understood by AI.

Take an example. You toss AI a question that sounds simple — "this went wrong, what do I do?" — and it gives you a few correct but useless lines: communicate first, coordinate, notify the relevant parties. Instantly you can tell it's an outsider talking.

It's not stupid. It just doesn't know any of the real difficulty behind the question: where exactly is this stuck? Whose interests does it touch? Which thread has to be protected first? How long can it wait? Can this be put in writing at all? And when reporting upward, how do you phrase it so you don't leave yourself open to blame?

None of this is written in any policy document. It lives in emails, meetings, spreadsheets, and the heads of long-time staff — built on years of unspoken understanding and judgment. So AI isn't clueless; it's that these unwritten rules were never spelled out — not even by the company to itself.

Here's the part that stings. If you actually open up your files to AI, what it reads is a mess: outdated documents still sitting on the shared drive, "let me check" and "I agree" tangled together in the same threads, several live versions of the same file. What we call AI "hallucination" is often not the AI inventing things — it's the AI faithfully reflecting the company's own muddled, untidy books. The problem isn't the AI. It's that the company never sorted itself out.

Which brings the real hurdle: trust.

Even when AI argues persuasively, the decision-maker still has to ask: why should I believe you? Believing it because it sounds polished is unsafe. What you can actually trust is when every judgment traces back — to which email, which record, which data point. It's the same as trusting a person: you trust them not because they never err, but because there's process, approval, records, and someone accountable. For AI to reach the core, it has to enter that same system — not coast on "it seems smart."

So what comes next? My read: AI won't replace everyone evenly. It will quietly split people into two groups.

One group does the work — sorting, chasing, summarizing, recording, translating. This gets heavily compressed.

The other group carries responsibility — the one who finally decides, signs, and answers for the judgment. This group becomes more valuable.

Because AI can do the work for you; it can't be accountable for you. The more you insist that someone must answer when things go wrong, the less AI can occupy that signing seat. Jobs don't evaporate in one sheet — they polarize: the hands-on work grows ever more automated, the responsibility on your shoulders grows ever heavier.

So for an individual, the way out isn't racing AI on speed — that race is lost in advance. The real position is in between: understanding the real situation and understanding AI; knowing where the key information lives and where the lines of responsibility fall; able to translate that intuitive, unspoken experience into rules AI can use and that can be traced. This "translator between business and AI" can't be replaced anytime soon.

One line to close on:

With enterprise AI, buying the tool is the easy step. The hard part is whether a company is willing to first turn its own muddled books into clear ones — legible to AI, and able to bear accountability. The day AI truly changes a company won't be the day it gets smarter. It'll be the day the company is finally ready to be seen clearly.


日本語要約

「AIはこれほど優秀なのに、なぜ社内でまだ誰かの仕事を本当に奪っていないのか」とよく聞かれる。

最初はAIの賢さ不足だと思っていた。だが考えるほど逆だと感じる。AIが会社に入れないのではなく、会社の側がAIに理解される準備をしていないのだ。

簡単そうな問いをAIに投げても、返ってくるのは「まず連絡を、次に調整を、関係者に通知を」といった正しいが中身のない言葉だ。問いの背後にある本当の難しさ——どこで詰まっているのか、誰の利害に触れるのか、どれを優先するのか、文書に残してよいのか、上にどう報告すれば責任を問われずに済むのか——をAIは何も知らない。

これらは制度文書のどこにも書かれていない。 メール、会議、表計算、ベテランの頭の中に、長年の暗黙の呼吸として隠れている。AIが的外れなのは、この暗黙知が一度も言語化されてこなかったからだ。会社自身ですら説明できていない。

さらに痛いのは、実際に資料をAIに開くと、それが映し出すのは混乱そのものだという点だ。古い版が共有フォルダに残り、「確認します」と「同意します」が混ざり、同じ資料の複数版が併存する。AIの「幻覚」の多くは、勝手な空想ではなく、会社自身の整理されていない知識の現状を忠実に映したものにすぎない。

だから本当の壁は「信頼」だ。AIがいくら筋の通った話をしても、決める人は問う——なぜ信じられるのか。表現の巧みさで信じるのは危うい。信じられるのは、すべての判断が根拠（どのメール、どの記録、どのデータ）まで遡れるときだけだ。

今後、AIは全員を均等に代替するのではなく、人を二つに分ける。作業する人（整理・催促・集計・記録・翻訳）は大きく圧縮され、責任を負う人（最後に署名し判断に責任を持つ人）はむしろ価値が高まる。AIは作業を代替できても、責任は代替できないからだ。

個人の活路は、AIと速さを競うことではない。現場とAIの両方を理解し、重要情報の所在と責任の境界を知り、暗黙の経験をAIが使える・追跡できる形のルールに翻訳できる人——「業務とAIの翻訳者」になることだ。

結論： AI導入でツールを買うのは最も簡単な一歩にすぎない。難しいのは、会社が自らの曖昧な実態を、AIにも理解でき責任にも耐える明確な形へ整理する覚悟があるかどうかだ。AIが会社を本当に変える日は、AIがさらに賢くなる日ではなく、会社がようやく「見られる」準備を整えた日である。