Everyone Has AI, But Why Doesn't Everyone Work Better?
A story from a worker who used AI to help check chat quality, before discovering that the problem wasn't solely with the AI, but with unclear work standards.
AI Didn’t Make Me Better at My Job. It Just Showed Me That This Office Never Had a System to Begin With.
It all started with the sound of packing tape, ripping right next to my ear.
I work in chat quality assurance at a small company that sells both thermal printers and children’s clothing.
The two products have almost nothing to do with each other, yet somehow they live together in the same warehouse.
Our office is about the size of a one-bedroom condo’s bedroom.
My desk sits in a corner, facing a window. It isn’t a pretty view.
It looks out onto the wall of the neighboring building, an arm’s length away, along with somebody’s air-conditioner compressor, which hums all day long.
Behind me is the stock and packing area. Boxes are stacked almost as high as the back of my chair, and every time someone pulls off a strip of bubble wrap,
I flinch as if someone had set off a firecracker without warning.
My job is to open the chats between our admins and customers and check how well the admins responded.
Did they greet the customer properly?
Did they use polite language?
Did they close the sale?
Sounds like a well-organized job, doesn’t it?
The only catch is that our scoring criteria live in a file called “New_Criteria_final_use_this_one2,” and its contents don’t match the file “New_Criteria_final” that my boss posted in the LINE group two months ago.
The night before the monthly summary was due, I had more than three hundred chats left to review. My coffee had gone cold long ago. So I did what most people in my situation would do: I opened a new tab and asked an AI:
“Please review these chats and score the admins.”
A few minutes later, I had a beautifully formatted score table, complete with strengths, weaknesses, and individual recommendations.
I felt like I’d won the lottery. I even started daydreaming that I might finally get to leave work on time like everyone else.
Until I went back and read the report a second time.
Lesson One: The Admin Who Never Said That
One line in the report said an admin had told a customer, “Size 100 fits up to four years old.” The AI had deducted points because that information didn’t match the company’s size chart.
So I went back and opened the original chat to check.
The admin had never said that. Not a single line in the chat contained that sentence. And the size chart the AI had compared it against wasn’t our company’s chart at all. It belonged to some stranger on the internet I’d never heard of.
Just think about it. If I had submitted that report to my boss, that admin could have been called in for a talk, or given a warning, for something they never did, backed by “evidence” that didn’t exist.
That would have turned into a completely different movie from the one I meant to show.
I burst out laughing so loudly that Phi Nui, who was packing boxes behind me, turned around to look. Then, for the first time, I started thinking seriously: what I’d received from the AI wasn’t “finished work.” It was just a first draft in a very nice outfit.
And that difference is what I want to talk about.
Lesson Two: Something That Looks Finished Is More Dangerous Than Something That Isn’t
A rough draft rarely fools anyone, because everyone can see right away that it still needs work.
But a neatly organized score table, written in polished language and presented with confidence, makes us feel the work has already been checked.
It seems to say, “All done. You can use this. Go to sleep.”
That is exactly the sentence I most wanted to hear at ten o’clock at night.
AI can always make its writing sound credible, whether the content is right or wrong, because the confidence it displays is no guarantee of accuracy.
It’s like an airport announcement, which sounds equally trustworthy whether it’s telling you your flight is on time or that your boarding gate has changed.
The cruelest irony is that my profession is verifying what people actually said, and yet I almost believed a report I hadn’t verified a single line of.
It reminded me of a watch repairman who wears a slow watch to work every day without ever noticing that something is wrong with his own.
Lesson Three: The AI’s Questions Exposed the Office, Not Me
After that incident, I started using AI differently.
Instead of ordering it to review the chats and do all the scoring for me, I tried asking:
“If I’m scoring chats about clothing sizes, what should I take into consideration?”
The AI answered, and then asked me back:
“Does the company have a standard size chart? And should admins base their recommendations on the customer’s age or height?”
I sat still for quite a while.
Because the answer was: no.
Each admin answers differently. Some go by age, some by height, and another asks for the child’s weight. And the admin who answered size questions most accurately was the one who was willing to walk over and check the tag on the actual garment in the warehouse before replying.
Yet under my evaluation criteria, that person was marked down on response time for taking longer than everyone else.
I had been doing chat quality assurance for a year, and I had never once asked whether the company had a clear standard for what a “good response” should be.
All along, I had simply been scoring by gut feeling, based on which answers seemed to understand the customer. And honestly, feeling that you understand and actually understanding are quite different things.
The AI wasn’t the cause of the problems in my work. It merely asked a few ordinary questions, and that made me see that the foundation of work in this office had never been level.
We had been walking on that bumpy ground for so long that we got used to it, and never thought to look down and see what was wrong in the first place.
Lesson Four: The Best Person at Checking AI’s Work May Be Someone Who Has Never Used AI
One day, I took an answer about thermal printers, which I’d had the AI help draft, to show Phi Nui in the packing department.
Phi Nui has unboxed who knows how many hundreds of printers, and in his entire life he has never opened a chatbot even once.
After reading just two lines, he said:
“This is wrong. A sticker roll this size can’t go in this model. The core is too big.”
The AI’s answer ran three paragraphs, with headings, explanations, and cautionary notes. It looked so professional there was almost nothing to criticize.
But it was wrong from start to finish.
Phi Nui caught the mistake in a few seconds, without opening a manual or looking anything up, because he knows what the actual products look like and which pieces of equipment work together and which don’t.
Since then, I’ve shifted my trust away from what appears in the chat window and toward the people sitting behind my chair.
People who’ve never been called experts, who hold no grand titles, and who may never have touched any new technology, yet know exactly what the real products are like.
These people are the most accurate fact-checking tool in the company.
And I had been sitting just two meters away from him the whole time, without ever thinking of walking over to ask.
That incident made me understand that AI is like a megaphone. If the person holding the microphone really knows their stuff, it amplifies their ability and carries them further.
But if the person holding the microphone doesn’t know what they’re talking about, it just makes their nonsense louder.
The Ending I Hadn’t Planned
The following month, I started working in a new way.
First, I sat down and wrote the scoring criteria on a single sheet of paper, to define clearly what we were checking.
Then I walked over to talk with Phi Nui and a few of the admins, asking which issues customers tend to ask about over and over, and what kind of answer should count as correct.
Only after gathering all of that did I give the criteria to the AI, to check whether any items contradicted each other, whether any conditions were missing, and which parts could be made clearer.
As for the chats that needed reviewing and scoring, I still opened and read every case myself, just as before.
The result: that day, I finished work at four in the afternoon.
The time when normal humans get off work.
I sat still for a while, not knowing what to do with the rest of the day. Eventually I looked up and gazed out the window.
The neighboring building’s wall was still a wall, and the air-conditioner compressor was still humming as always. But for the first time, I noticed a water stain on the wall, shaped a bit like a map of some country.
I sat there looking at it for about five minutes.
The extra time didn’t come from AI doing all my work for me. It came from seeing more clearly what my real job was, and what was merely chaos I had mistaken for work.
Sometimes new technology doesn’t arrive to make anyone faster. It arrives as another mirror, showing us what we’ve actually been wasting our time on.
For me, the reflection in that mirror was an office that had never clearly defined what a “good response” means, but which had always had someone in charge of checking the quality of those responses.
And that someone was me.
Not once had I questioned the standard I used to judge others, or asked what exactly we were measuring when we talked about “quality.”
As for that coffee cup, I still haven’t washed it.
It’s still sitting beside my computer, a memorial to the night I almost wrote a warning letter to an admin over a sentence he never said, not even once.

