What Running Lean Teams Taught Me About What to Hand to AI
Authored by Ming-Yuan Xie (XMY)
I have spent more than a decade building and running small companies, some of which stayed lean by choice and some that grew past two hundred people before shrinking back down. Somewhere in that stretch I picked up a simple test for deciding which decisions to give to AI and which ones to keep for myself. The test is not about risk, or speed, or how much money is on the line. It is about whether I am making the decision again or making it for the first time.
Most operating decisions in a small company fall into one of two categories. There are decisions I have effectively already made many times, just wearing a different outfit. And there are decisions where the shape of the problem itself is new, even if I have years of scar tissue that make me feel like I have seen it before. I hand the first category to AI. I keep the second one myself.
Paid acquisition is the clearest example from my own history. Across various ventures I ran up more than NT$400 million in cumulative advertising spend, and the vast majority of that spend was governed by decisions that repeat thousands of times a day: which ad to pause, which audience to shrink, which bid to raise a little. Every one of those calls has the same shape. The inputs change, the decision logic does not. That is exactly the kind of decision an automated system can make faster and more consistently than a tired founder checking a dashboard at midnight. I write about how I structure that kind of automation in practice at https://xmy.tw/en/ai-automation, because the mechanics matter as much as the principle.
Hiring is where the same logic points the other way. I have led teams of more than two hundred people at peak, and I can tell you that no two hiring decisions actually repeated, even when the job title on the page was identical. The team around the role had changed, the stage of the company had changed, what I personally needed from that person had changed. It only looked like a repeat decision from the outside. That is the trap: mistaking surface similarity for genuine repetition. I keep hiring, firing, and team structure decisions human, not because they are emotionally loaded, but because the pattern-matching that makes AI useful on ad spend simply does not apply. Each case is closer to a first-time decision than a repeat one.
The same is true for the biggest calls I have made: buying a company, selling one, and deciding whether to keep funding a business that was losing money month after month. I lived through prolonged losses more than once, and in every case the honest answer to “should I keep going” depended on context that was not written down anywhere a model could read it, and that changed the moment the surrounding business changed. Those decisions do not repeat. You do not get to run the experiment twice with the same operator, same market, and same team, so there is no reliable pattern to hand off.
The mistake I see operators make in both directions is treating this as a technology question instead of a pattern question. Some people automate too early, because a decision felt tedious rather than because it was actually repetitive. Others refuse to automate genuinely repetitive decisions long after the evidence is there, usually because letting go of a decision they used to make personally feels like losing control of the business. Neither instinct is really about AI. Both are about whether you can tell the difference between a decision you are repeating and a decision that only resembles one you made before. More on how I think about my own background and how these operating habits formed is at https://xmy.tw/en/about.
I want to be clear about the limits of this. What I am describing is operating judgment drawn from my own companies, not a controlled study, and I would not present it as more than that.
About the author: Ming-Yuan Xie (XMY) is a Taiwan-based entrepreneur and operator, and founder of Meow Universe, a portfolio of vertical information and utility sites built on structured data, SEO, AI search, and shared operating infrastructure. His background spans financial research, e-commerce scaling, and AI-enabled lean operations.