Not long ago, I helped a client set up an automated publishing system for a WeChat Official Account. It wrote articles, designed covers, added illustrations and handled the layout. From understanding the requirements and deploying the system to making revisions and getting it running for real, the whole job took a week.
AI did the analysis. AI did the planning. AI did the production.
I did only two things: figure out what the client actually needed, and keep AI from drifting off course.
Those are the two things I will not hand over to AI.
Clients say they want AI. What they want is never AI.
![[raw/images/fffb8b80423314a8e386216ad82a3132_MD5.png]]
The first is understanding the real need.
For the past two years, nearly every client who has come to me has opened with some version of the same sentence: We want to use AI to cut costs and improve efficiency, and find more customers through digital marketing.
The words AI were added later, because everyone was talking about it. What was really weighing on them was finding customers.
I have tried handing the requirements analysis to AI. I fed it all the materials the client had sent and asked for an analysis.
It could do the job, and it looked convincing: target audiences, content direction, publishing cadence—everything was neatly reasoned and hard to fault at first glance.
The problem was that it could not read the person behind those materials.
A traditional manufacturer whose owner wants to start a WeChat account might want downstream customers to see that the company is changing. Or he might want competitors to stop thinking the company is stuck in the 2010s. Or perhaps someone asked during a tender whether the company was doing anything with digitalization.
Those three motives would lead to completely different writing.
And none of that information was in the documents he sent me. It was in two hours of conversation when we met, and in the way he described why a project had fallen through the year before.
If the information needed to make a judgment is not in what you give AI, you cannot hand it that step.
It has not simply made a bad guess. It has to give you an answer without that information—and it has to make the answer sound convincing, because it will not say, “I don’t know.”
That is why the most time-consuming part of this work has never been producing things. It is sitting down and listening to people.
AI listed ten options. Not that many were right.
![[raw/images/e8522cd5cc992e3770fe27f58848d9ec_MD5.png]]
The second thing is keeping it on track.
I do not mean proofreading typos. Every time it gives me a draft, I have to judge whether it can be used, where it falls short, whether it needs revision, and how much revision is enough.
Once, while researching products to source, I asked it to scan a category and give me a list of options worth pursuing.
It listed ten.
I went through them one by one: three were genuinely viable; four were optional—you would not lose money by doing them, but you would not miss anything by skipping them; the remaining three were overreach, and following the advice could actually get us into trouble.
If I had let it carry on by itself, it would have pursued all ten.
It is not stupid. It does not bear the consequences, so it cannot make trade-offs. It is not AI’s money sitting in the warehouse, and it is not AI that loses face when a client rejects a proposal. Someone who bears no consequences will always decide that doing everything is the safer choice.
Only someone who has something to lose can make a real trade-off.
It is even more obvious with branding. A client’s brief says premium, young and bold—three words that say almost nothing. Give AI those words and it can write you an entire brand manual. The language will be polished; it just will not sound like that company.
Someone has to say: this line does not work; this paragraph is too airy; the client’s distributors will not understand this phrasing.
The person who says that has to be the one who will have to redo the work if the draft comes back rejected.
Two yardsticks
The reasons for keeping these two things are different, but they boil down to two questions you can ask at any time.
First: Is the information needed to make a judgment actually in what you give AI?
If it is not, do not hand over that step. AI will not tell you it does not know. It will invent a plausible answer, and you may not realize it is made up.
Second: Who bears the cost if this goes wrong?
If the person who bears the cost is not part of the process, someone still needs to watch over that step. They do not have to supervise every moment, but a person must make the final call.
Draw a clear line, and you can let go
![[raw/images/f5ae2c90b4f73b8753636f513f73cd13_MD5.png]]
Once I kept those two things in my own hands, I handed over the rest completely.
As long as the material I feed it is sufficiently complete, the first draft is usually ready in one go. I read it through and make a few small edits. AI handles the structure, illustrations and layout; I do not need to supervise each piece. Even an illustration prompt can be settled in a single sentence.
Drawing a clear line actually makes it easier to let go.
When things are vague, people worry in both directions: they do not dare hand everything over, but they are not willing to start from scratch themselves either. So every step is done halfway.
Doing every step halfway is the most expensive way to work. You pay twice—once for the machine and once for the person—and get a result that neither has fully delivered.
So if you want to know whether a company is really using AI, do not count how many tools it has adopted.
Look for the line it has drawn: which parts are handed over completely, and which parts are never handed over at all.
If the line is not there, it probably has not really started.
When I put together a proposal for a client, the most time-consuming part is never listing what we could do. It is persuading them to cut some things.
Because the things that need cutting are often exactly the things the boss is most excited about.
![[raw/images/5a24de6e0604588df9fec63a88f047d1_MD5.png]]
My own notes took me by surprise
Building a knowledge base that talks back | Karpathy’s LLM Wiki, built and customized