AI
Your AI Output Is Bad? I Review AI Work Every Day. It's Usually One of Three Diseases.
The draft is done. The AI ran smoothly. But when you read it again, something is off. Not wrong exactly. Just lifeless.
The most common complaint I hear is that AI output feels flat, and people cannot figure out why. The prompt was long. The tool was expensive. The model is the newest one. The result still feels dead.
I review AI output every day. Content for my 8 channels, reports, research, decks. I do it the same way I used to review my team's work at Tokopedia and Flip. After hundreds of reviews, the pattern comes down to three diseases. The model is not dumb.
Disease one: you accept the first draft
This is the most common one. Someone writes a prompt, the AI produces a result, they ship it, done.
Think back to when you were the junior. Your boss never accepted your first draft, right? So why do you accept the AI's first draft?
Back when I reviewed junior PMs, my routine was always the same: read once to catch the intent, read twice to find the weak spots, then give specific notes. Now I do exactly the same to AI. The difference is that AI does not sulk when I mark it up.
AI work is like the work of a new hire. Smart, fast, but it does not know your standards yet. Standards never transfer through a single prompt. They transfer through consistent review.
A first draft is raw material. The final touch shows up in the second round, the third, sometimes the fourth.
Disease two: your feedback is global, not specific
If you only say "not good enough, fix it", the AI will fix things based on its own guess. That guess usually goes the wrong direction.
Same as giving feedback to a person. "The deck lacks punch" changes nothing. But "slide three, the numbers do not support the claim, either change the numbers or soften the claim" is what actually lifts the result.
I have an old habit from my PM years: feedback should stick to the part that has the problem. Do not waste time commenting on everything. Now, with every AI output I review, I mark which part is weak and why. This paragraph is flat because there is no example. This sentence goes because it repeats the previous one.
AI picks up patterns from the specific. Show it your standard in one section and it can carry that standard to the rest. If you only say "not good enough", it guesses. And it keeps making the same guess.
Disease three: you never show an example of what you want
This one costs the most.
Most people describe the result they want in words. But the fastest way to set a standard is to show an example. Do not say "write like a professional speaks". Paste one piece of writing that actually sounds like that, then tell the AI to follow the pattern.
I realized this while reviewing my own content output. The results I called lifeless were not lifeless because the AI could not write. They were lifeless because I never showed it writing that I consider good.
The moment I started including examples of final outputs I like, plus notes on why I like them, quality jumped. Now every type of work has its own reference example.
AI is your team, not a machine
Out of all those reviews, one shift changed how I work: treat AI like your team.
How can you be so careful giving feedback to a junior, but with AI you just say "make it"? Both need to know your standards. Both need correction when they are wrong. The difference: AI does not get tired, does not get offended, and can be corrected ten times without complaint.
Pushing work quality to the next level used to be a leadership skill. Now the same skill applies to AI output. Making companies efficient is an old habit of mine, long before AI was trendy: $4M+ (USD) per year in savings, from a mix of initiatives I drove as a product leader, including process improvements, cost optimization, and sharper product decisions. Automation was only one of them. That kind of leverage used to require a position, an engineering team, and expensive systems. Now AI is the sharpest tool for the same habit, and it can be trained into everyone on the team.
The proof is in my current work. The 8-channel content distribution system I built and run alone, the Applied-AI Certification, and AI Circle all came from the same loop: draft, review, specific correction, show an example, repeat.
No magic. Just disciplined review.
Where to start
Try it tomorrow. Take one AI output you normally produce, then review it like you would review a person's work: mark the weak parts, write down why they are weak, then have it fix them one by one. Do not say "fix everything".
If you want to practice this pattern with me and others working on the same thing, join AI Circle. Every month we break down real workflows, not theory.
If you lead a team or own a company and want your whole team working this way, I also train companies directly. Check the details on the corporate page.