When I first started integrating AI into automation workflows, I was actually very enthusiastic. My intuition was simple: if a system could become a bit smarter, wouldn't that reduce the burden on people? Because of that, whenever I looked at a process, my first reaction was often:

Can we add some AI here?

Can this step be automated a little more?

After several real-world experiences, I slowly realized that the problem was not that AI itself was not good enough. The real issue was that, in many small and medium-sized businesses, the system itself was not ready to support AI.

In many SMBs, what I see is not a complete and continuous flow of information, but fragmented pieces of data that do not connect well. Some critical information does not exist in any system at all, but lives in someone's head. Other information is scattered across different software tools, different Excel files, or even paper documents.

As a result, people become the busiest part of the system. They copy data from system A to system B, move numbers from one Excel file to another, convert paper into digital form, and then manually check what "looks wrong." This often creates a strange situation: individual steps become faster because of tools, but the overall workload actually increases.

Sometimes, the problem is not AI at all. The problem is that the process itself is still unstable.

When a process is unstable and changes frequently, automation already carries risk. Once a process is handed to a system, its mistakes get fixed into place as well. AI makes this even more obvious. AI is very good at producing results quickly under existing rules, but it does not know which exceptions are actually signals that the process itself is broken. When exceptions are treated as normal inputs, AI will simply make wrong decisions faster — and in a way that looks more "reasonable."

This is why I often see that when AI is introduced too early, before a process is truly understood and stabilized, later corrections become much harder. Many times, choosing to move a bit slower at the beginning is actually a way to create room for faster progress later.

This idea of "slowing down" often conflicts with the desire of many SMBs to do things "all at once." In small companies, "one-shot solutions" sound very attractive. They seem to save time and money, and they feel advanced. But in practice, doing everything at once often leads to systems that are too complex to understand. Over time, no one dares to touch them, and no one wants to take responsibility.

When a system breaks, the safest option for most people is to go back to familiar Excel files or manual processes. That is why, instead of building a system that "can do everything," I care more about whether a system can be rolled back, adjusted, and re-understood. To me, keeping a way back is not being conservative — it is risk management.

So when do I think AI is actually appropriate? Usually only after certain conditions appear. The process is no longer being overturned every day. People start to trust the system instead of working around it. Decision boundaries can be clearly explained, instead of being patched after problems happen. Before these foundations exist, AI easily becomes an extra burden rather than real help.

That is why I do not start with AI. Not because the technology is not capable, but because the system, the process, and the people are not ready to absorb it. For me, what really matters is not how fast AI is introduced, but whether the system can be understood, trusted, and sustained over time. Sometimes, building a solid foundation first is already more than enough.

剛開始把 AI 放進自動化流程時，我其實非常積極。那時候的直覺很簡單 -- 如果系統能多一點智能，人的負擔是不是就能少一點？所以我幾乎每看到一個流程，就會下意識想著：這裡能不能加點 AI？那裡是不是也可以再自動一點？

直到幾次實際經驗之後，我才慢慢意識到，問題其實不在於 AI 做得不夠好，而是在很多中小型企業裡，系統本身還承受不起 AI。

在不少中小型企業中，我看到的不是一個完整、連續的資訊流，而是一段一段接不起來的資料。有些重要資訊不存在任何系統裡，而是放在某個人的腦中；有些則散落在不同軟體、不同 Excel，甚至紙本文件之間。

於是人反而成了系統裡最忙的一環：把資料從 A 系統抄到 B 系統，從一個 Excel 補到另一個 Excel，從紙本轉成數位，再人工檢查哪裡看起來怪怪的。也因此，常常會出現一種很弔詭的狀況 -- 工具讓單一步驟變快了，但整體工作量卻變更多。

有時候，問題真的不在 AI，而是在流程本身還不穩定。

在流程還不穩定、每天都在變動的狀態下，自動化本身就帶著風險。因為一旦把流程交給系統，錯誤也會一起被固定下來。AI 特別容易放大這件事，它很擅長在既有規則下快速產生結果，卻不會分辨哪些例外其實是在提醒流程本身出了問題。當例外被視為正常輸入，AI 只會更快、也更「合理」地做出錯誤判斷。

這也是為什麼我常常看到，在流程還沒被真正理解與穩定之前，越早導入 AI，後面的修正反而越困難。很多時候，看起來慢一點的選擇，其實是在幫系統留出之後能加速的空間。

這個「慢一點」的選擇，常常和中小型企業追求的「一次到位」相衝突。在中小型企業裡，「一次到位」是一個很有吸引力的想法，看起來可以省錢、省時間，也顯得很先進。但實務上，一次到位往往意味著系統變得複雜、難以理解，久了之後，沒人敢動，也沒人敢負責。

當系統一出錯，人們最安全的選擇，往往就是回到熟悉的 Excel 或人工流程。所以比起打造一個「什麼都能做」的系統，我更在意的是：這個系統能不能被退回、被修正、被重新理解。對我來說，保留退路從來不是保守，而是一種風險管理。

那什麼時候，我才會覺得 AI 是合適的？通常要等到某些狀態出現，我才會覺得「現在是合適的時機」-- 流程不再每天被推翻，人開始信任系統而不是繞過它，決策的邊界能被說清楚，而不是事後補救。在這些基礎出現之前，AI 很容易變成額外的負擔，而不是幫助。

所以，我不會一開始就導入 AI。不是因為技術做不到，而是因為系統、流程與人，還沒準備好承受它。對我來說，真正重要的不是「多快導入 AI」，而是這個系統能不能被理解、被信任，並且長期活下來。有些時候，先把基礎做好，就已經很足夠了。