Picking up where I left off.
The boss goes off to learn AI, but cannot make it work back at the office because he has no real use case. He has been away from the front line too long to say what he needs.
The boss says, “So it’s my fault?”
What about someone else? If the employee who works on the front line every day goes to learn, will that finally cut costs and improve efficiency?
Not necessarily.
There is another problem.
He knows what AI should do, but cannot feed it what it needs
Front-line staff have all the real-world situations in their hands. The boss cannot compete with them there.
Which customer’s payment terms can be relaxed a little; which inquiries come from competitors sounding us out; when the written acceptance standard says A but the goods are actually handled as B; why last year’s order really fell through, when the reason given at the review meeting was not the real one.
He can spot all of that at a glance.
The problem is, that glance has never been written down.
It is not in the ERP, the CRM or the policy manual. It lives in his head under the name experience. For years, nobody thought that was a problem—until someone wanted AI to take over.
AI can only take on a task if it has access to all of this. Without it, it can only do the work whose rules are already written down.
So the employee comes back from training and runs into a strange wall: he knows what AI should do, but cannot explain the rules.
![[raw/images/459616433c1f84033033f97b133326b4_MD5.png]]
I used to be that hard-to-replace employee
I have been a translator for twenty years. Everything that makes my work valuable is this kind of knowledge: which term cannot be translated that way in this industry; which sentence a client objected to last time; whether a reader will lose interest by the third paragraph.
None of it was written in a book. I learned all of it on the job.
Put another way: if I were an employee in a company, I would be the person who is hardest to replace and, at the same time, hardest to get to hand over what they know.
Later, I wrote it down and turned it into rules for AI.
Writing was not the hard part.
The hard part was that I could not explain the reasons behind many of my judgments. I could tell at a glance that a translation was wrong, but it took a long time to work backward and say exactly why. Sometimes, after thinking it through for ages, the rule I finally wrote down was much cruder than the instinct I had started with.
That loss is real. What gets written down is always less than what the mind knows.
But the smaller part that is written down can be used. The larger part that stays in my head only works when I am there.
Maybe he is not refusing to write it down because he is lazy
So when you send someone from the front line to learn AI and they still cannot get it moving afterward, I am not surprised at all.
Ask someone who has done the job for eight years to document their judgment, and they may stall, give you a shallow version, or say the work is too complicated to explain clearly.
It really is hard to explain. I lost a lot of detail myself when I tried to write mine down.
But there is another thing they will not say out loud: after writing all that down and handing it to the company, what will they have left?
The bargaining power they have built up over eight years is in the things they have not written down.
This is not a moral issue. It is a very practical calculation. Anyone in that position would make it.
![[raw/images/e90329025dd6b51c83f905b70ece1ddc_MD5.png]]
There is another kind of employee. They have already written it down.
So far, I have described people who cannot write their knowledge down, and people who are unwilling to. There is a subtler case.
Some people have already done it.
They use AI at work, and they are very good at it. They have figured out what to do first, what to feed it, and what counts as an acceptable result. They have built a system of their own: a prompt, a skill, a workflow only they know how to tune.
They get the work done twice as fast as before, and nobody can find fault with what they hand in.
The company sees an employee who is incredibly efficient.
What the company does not see is that the efficiency is not in the employee’s hands. It is in the system they have built.
If you now ask them to organize that system, document it and hand it over to the company, they will smile—and nothing more will come of it.
Because this is no longer just experience. It is the result of repeated experimentation, a tool they built themselves. Asking them to hand it over for free is like asking someone to donate the machine they built to the company and keep working for the same salary.
There is no argument that makes this fair. They are under no obligation to do it.
I have a system like that myself. I have turned the steps of translation that do not need a person into skills, and they run the work for me. I spent a long time tuning them. Now they are my production line.
Would I hand it over for nothing? I would not.
![[raw/images/b9dfb5d30e021c878739d68aa3dac9fc_MD5.png]]
So this hurdle cannot be solved as a technical problem
Who should go to learn, how many days they should spend, which tools they should learn—those are all easy questions to answer.
And many companies are not actually starting from zero. They are starting from a place where someone is already using AI, but the output is credited to that person alone. The boss just does not know it.
The real hurdle comes later: is anyone willing to hand over the system they have built?
And you cannot produce that willingness with a motivational speech in a meeting.
I wrote earlier that there are two kinds of AI users: those who act like bosses and those who act like operators. This is where the paths split.
If someone hands over their system and the company responds, “Great, AI will do this job from now on. You go do something else,” they will never hand over anything again. They will also tell everyone around them not to do it.
But if the response is, “You will maintain these rules from now on. AI will follow what you wrote, and you will fix it when it goes off course. If others can use it too, that counts in your favor,” then the employee becomes the boss of the system instead of its operator.
The same situation, handled in two different ways, leads to opposite outcomes.
The person willing to hand over their system must first believe that doing so benefits them, too. That benefit has to be real—the kind you can put in writing.
Zhishen once put together a list of obstacles to putting AI to work in companies: business, data, organization, technology, cost and compliance, with a few dozen issues across those areas. This article picks just one.
I chose it because it looks least like a technical problem and is the hardest to solve with money.
![[raw/images/7ed6750065d24515136fac8947b7f595_MD5.png]]