During the Canton Fair I spent a few days at a booth receiving customers. Almost everyone who sat down eventually asked: can you do this with AI?
I usually asked what problem they were trying to solve first. Most of the time what they wanted was not AI. It was two fewer hires, or a quotation that comes out faster.
Afterwards I kept thinking about what AI will change, and what it will not.
A history of outsourcing
Compress economic history and it looks like a history of outsourcing.
Before the industrial revolution, value came from muscle. Then humans outsourced muscle to machines and stepped behind them to design the machines. The information revolution outsourced process to software: people wrote code, code ran the process, process directed the machines.
What AI is doing now is outsourcing the last layer — judgement, writing, analysis. These used to require a person. Now a model can produce a first version and a person edits it.
So I do not think “will AI replace people” is a good question. Every previous round of outsourcing replaced one layer of labour and pushed people up a layer. The real question is: what does the layer above require, and do I have it?
From an individual’s point of view there is another pattern: what gets outsourced is the repeatable, standardisable part. What stays is judgement, communication, and carrying the consequences.
What year of the internet is this
Hard to say. But there is a useful test: find a direction where, in the future, not having it would feel like not having e-commerce today.
When the internet appeared, nobody imagined food delivery or online shopping. In 1995 people used it to read news. They probably did not picture buying groceries, hailing taxis or arguing with strangers on it ten years later.
AI today is in the same position. The models exist, but nobody can see clearly what “e-commerce on top of AI” will be.
The test has a precondition: you have to be deep enough in an industry to see what is missing. From outside you only see the noise.
I spent some time working around recruiting, so when I look at AI tools for hiring I can usually tell which ones solve one concrete step and which ones are just an old feature with a new name. That kind of judgement only comes from being inside.
What work fits AI
My view is fairly conservative. Work that fits AI usually shares a few traits.
- It is repetitive, with a clear feedback signal — change one thing and you can tell whether it got better.
- The data is structured, not scattered across chat logs and screenshots.
- Mistakes can be verified and rolled back, so one bad output is not irreversible.
- People still make the key decisions; the model prepares material.
Conversely, scenarios with slow feedback, dirty data or expensive errors will adopt AI slowly, and often end up as demos rather than systems.
My latest test is simpler: could an intern learn this in a week? If yes, it probably fits AI. If no, AI probably will not manage either.
Implementation is harder than the model
In my last internship I worked on AI implementation. The biggest takeaway: the model is the most standardised part of the whole system.
The difficulty is elsewhere. Which system does the data come from, how are permissions handled, who reviews the output, what happens to colleagues used to the old process, how do you roll back when it fails. None of it is cool. All of it decides whether AI moves from a demo to something used daily.
A concrete case: a customer wants AI to generate quotations. The model can write. But a quotation needs the company price list, customer discounts, exchange rates, lead times. That data is not in one system; some of it is in Excel. Just connecting the data takes longer than tuning the model.
A ten percent better model might improve the final result slightly. Cleaning up the process, connecting the right data and changing people’s habits might double it.
What I learned in an internship
In my last internship I did two things: help customers choose AI tools, and connect those tools to their existing processes.
Choosing tools was not hard. The second part was. The customer’s old process had accumulated over ten years: Excel, group chats, paper forms. To put AI in, you first have to untangle all that, then decide which steps to hand to the model.
The most memorable case was a request that looked small and ended up changing how a whole department worked together. That was when I understood: the hardest part of an AI project is people, not technology.
Most companies do not lack models. They lack someone who can wire a model into the business process.
My own position
I am not worried about being replaced by AI. I am worried about the moment a customer asks “can you do this with AI?” and I cannot answer.
So I am doing two things now. One: understanding the processes in front of me, so I know which step can be automated. Two: building tools myself, even ugly ones.
Building tools is harder than I expected. Getting a script to run reliably means handling all kinds of strange inputs. But once it works, the time saved every day is real.
Being on the side that uses AI is not the same as being able to use it. You have to break a few tools first. I am, roughly, a beginner.
AI is a bigger round of outsourcing. What I can do is stand on the side that knows how to use it.
I do not know where this road leads. But every tool I finish makes my understanding of it a little more concrete.
That is also what I am catching up on lately: fewer launch events, more things built by hand. Using a tool takes an afternoon. Wiring it into a business takes months of dirty data and uncooperative process. Those are different skills.
