There are at least four users in the same buyer

The person trying an AI tool is rarely one simple persona. In the same company, the curious builder wants speed, the operator wants fewer mistakes, the manager wants predictable output, and the reviewer wants evidence. A tool that only delights the first person may still fail the adoption path.

This is why "the demo was impressive" and "the team adopted it" are different statements. The demo answers whether the tool can produce something. Adoption answers whether other people can inspect, correct, and depend on that something.

The trust ladder

I think about AI tooling as a ladder rather than a category. At the bottom, the tool suggests. Then it drafts. Then it edits under supervision. Then it acts inside a narrow boundary. Each rung needs a different proof.

• Suggestion needs relevance: did it surface something I would have missed?
• Drafting needs taste: can I keep enough of it that starting from blank would be worse?
• Editing needs traceability: can I see what changed and why?
• Action needs containment: can the tool fail without causing a mess?

Handoff is the hidden product surface

The most neglected part of AI tooling is the handoff back to the human. A tool can be accurate and still unusable if the result lands as a blob. People need a way to skim confidence, inspect sources, compare alternatives, undo bad steps, and explain the output to someone else.

A concrete example: a resume evaluator that says "improve impact" is not enough. A useful one shows which line triggered the feedback, what assumption it made, what a stronger rewrite would change, and where the user should ignore the advice. The product is not the score. The product is the handoff.

Where small tools can still win

Large platforms will keep expanding horizontally. Small tools can still win by being opinionated about a specific handoff. They can make one review step clearer, one messy input safer, one decision easier to audit. That is not less ambitious. It is a different kind of ambition: owning the moment where trust usually breaks.

同一個買方裡其實有好幾種使用者

試用 AI 工具的人很少是一個單純 persona。同一家公司裡，好奇的 builder 想要速度，operator 想要少出錯，manager 想要穩定輸出，reviewer 想要證據。工具如果只讓第一種人覺得驚艷，仍然可能走不到真正採用。

所以「demo 很厲害」和「團隊採用」是兩句不同的話。demo 回答的是工具能不能產生東西；採用回答的是別人能不能檢查、修正、依賴那個東西。

信任是一段階梯

我比較喜歡把 AI 工具想成一段信任階梯，而不是一個分類。最底層是建議，接著是草稿，再來是在監督下修改，最後才是在窄邊界內行動。每一階需要的證據都不同。

• 建議需要相關性：它有沒有指出我會漏掉的東西？
• 草稿需要品味：它有沒有讓我比從空白開始更快？
• 修改需要可追蹤：我看不看得出改了什麼、為什麼改？
• 行動需要可控制：它失敗時會不會把事情搞爛？

交接才是隱形產品面

AI 工具最常被忽略的地方，是結果怎麼交還給人。工具可以很準，但如果輸出只是一坨東西，仍然很難用。人需要快速看信心、檢查來源、比較替代方案、撤回錯誤步驟，並且能把結果解釋給下一個人。

具體一點說，履歷工具只說「impact 不夠」是不夠的。更有用的是指出哪一行觸發回饋、它假設了什麼、改寫會改變哪裡，以及什麼時候應該忽略建議。產品不是分數，產品是交接。

小工具還能贏在哪裡

大平台會繼續橫向擴張。小工具仍然能贏，但不是靠假裝自己也是平台，而是對一個具體交接場景有立場：讓一個 review 步驟更清楚，讓一份髒輸入更安全，讓一個決策更容易追查。這不是比較沒野心，而是另一種野心：抓住信任最容易斷掉的那一刻。