「AIエージェントに任せたら、言わなくても先回りして全部やってくれた」。そんな体験談を見かけませんか?
先回りしてくれるなら助かる。でも実際に任せてみると頼んでいないことまで進んでいて驚くことがあります。「OK」と返しただけなのにどこまでのOKだったのかが自分でも分からなくなる。
このままだと頼んでいない変更や送信が積み重なり、あとから元に戻す手間のほうが大きくなります。でも承認の書き方を少し変えるだけでAIは頼んだ範囲で止まるようになります。
そこでこの記事では私が運用で決めている承認の書き方を3つお伝えします。読み終わる頃には次の「OK」から書き方を変えられます。
「OK」は人によって広さが違う
「この文面でOK」と言ったとき私が承認したのは文面です。でもAIは「送ってよい」まで含めて読むことがあります。文面のOKと送信のOKは別のものです。
これはAIだけの話ではありません。部下に「それでいいよ」と言ったら思ったより先まで進んでいた経験はないでしょうか。
買い物を頼むときに似ています。「牛乳を買ってきて」と頼んだのに気を利かせて卵とパンも買ってくる。親切ですが頼んだのは牛乳だけです。
私は運用のルールに「承認の範囲を広げて解釈しない」を入れました。これはAIが止まるべき条件として置いています。
書き方1:「何を・どこまで」を1行で書く
承認するときは「OK」だけで返しません。「文面はOK。送信はまだ」のように何をどこまで認めたかを1行で書きます。
10秒ほどの手間です。でもこの1行があるとAIは次の段階の手前で止まります。
書き方2:承認の言葉をそのまま記録させる
私はAIに私の承認を要約させず原文のまま引用して記録させています。どの会話での承認かも一緒に残します。
要約すると意味が少しずつ広がります。「文面はOK」が記録の中で「承認済み」になり、いつのまにか「送信も承認済み」として扱われる。原文が残っていればそれを防げます。
書き方3:範囲の外に出そうなら聞かせる
AIには「承認された範囲を超えそうなときは、進めずに確認して」と頼んでおきます。
聞かれる回数は少し増えます。でも勝手に進んだ作業を戻すよりずっと短い時間で済みます。
ちょっと考えてみてください。最後にAIへ「OK」と返したときあなたは何をどこまで認めたつもりでしたか?
アクションのヒント
- 次にAIへ承認を返すときは「OK」の後ろに1行足してください。「ここまでOK。その先はまだ」。
- これだけで頼んでいない作業を減らせます。
ポイント
- 承認は「OK」だけで返さず何をどこまで認めたかを1行で書く
- 承認の言葉は要約させず原文のまま記録させる
- キーワード:「ここまでOK。その先はまだ」
どこで確認を挟むか、仕事ごとに先に決めます
承認の1行は次の「OK」から書けます。ただ毎回その場で範囲を考えていると、確認そのものが負担になります。アシッドハウスは今の仕事を一緒に見て、AIに任せきりにする所と必ず人が確認する所を先に決めます。
お問い合わせ内容に、確認や差し戻しが多い仕事を一つ書いてください。通常1〜2営業日以内に返信します。
"I handed it to an AI agent and it got ahead of me and did everything without being told." Have you seen stories like that?
It would help if it got ahead of you. But when you actually delegate, you're sometimes surprised to find it has gone ahead with things you never asked for. You only replied "OK," and now even you can't tell how far that OK reached.
If this continues, changes and sends you didn't ask for pile up, and undoing them later takes more effort. But by changing slightly how you write approvals, AI will stop within the scope you asked for.
So in this article I'll share three ways of writing approvals that I use in my own operations. By the end you can change how you write your very next "OK."
How wide an "OK" is depends on who reads it
When I say "this wording is OK," what I approved is the wording. But AI sometimes reads it as including "you may send it." An OK for the wording and an OK to send are different things.
This isn't only about AI. Have you ever told a staff member "that's fine" and found they had gone further than you expected?
It's like asking someone to shop for you. You ask them to "buy milk," and they thoughtfully buy eggs and bread too. It's kind, but milk was all you asked for.
I put "do not interpret the scope of an approval broadly" into my operating rules. I set it as a condition under which AI must stop.
Way 1: Write "what, and how far" in one line
When I approve, I don't reply with just "OK." I write in one line what I approved and how far, such as "The wording is OK. Don't send yet."
It takes about ten seconds. But with that one line, AI stops before the next stage.
Way 2: Have it record your approval word for word
I have AI record my approvals by quoting them exactly, not summarizing them. It also keeps which conversation the approval was given in.
Summaries widen the meaning little by little. "The wording is OK" becomes "approved" in the record, and before you know it, it's treated as "sending approved too." Keeping the original wording prevents that.
Way 3: Have it ask when it's about to go beyond the scope
Tell the AI in advance: "When you're about to go beyond what was approved, don't proceed. Check with me."
You'll be asked a little more often. But it takes far less time than undoing work that went ahead on its own.
Think about it for a moment. The last time you replied "OK" to AI, what did you mean to approve, and how far?
Action tips
- The next time you give AI an approval, add one line after "OK." "OK up to here. Not beyond that yet."
- That alone cuts down on work you didn't ask for and makes checking after delegating easier.
Key points
- Don't approve with just "OK." Write in one line what you approved and how far
- Have your approvals recorded word for word, not summarized
- Key phrase: "OK up to here. Not beyond that yet"
We decide in advance, task by task, where a check goes in
You can write the one-line approval starting with your next "OK." But if you work out the scope on the spot every time, the checking itself becomes a burden. ACIDHOUSE looks at your current work with you and decides in advance where AI can be left to it and where a person must always check.
In your message, describe one task with a lot of checking or sending back. We usually reply within 1–2 business days.