AI is changing knowledge work forever

I started my career doing the sort of work nobody wants to do. I was in my mid-twenties, fresh off a stint travelling around the world largely disconnected, and I took a sales job at a PR agency in Stockholm. It was commission only. Everyone around me said it was a terrible idea, and for the first few months, my bank account agreed with them. But I met Fredrik Bage there. He told me I would be good at sales, and I believed him enough to put my head down. My days consisted of staring at a blank screen, writing terrible pitch emails, sending them to people who did not want to hear from me, and getting rejected. Then I would rewrite them and try again. It was painful, slow, and tedious. But that is how I learned. I built a sense of what worked by producing absolute rubbish and slowly refining it. That was the classic apprenticeship. You earned your stripes by doing the heavy lifting.

The heavy lifting is largely gone now. Today, I live in Marbella, a move from Sweden in 2018 with my family that was supposed to last two years and somehow became permanent. Most of my time is spent building and investing. With Anchor, we are building AI for alienated fathers navigating the family court system. With Substrat, we are providing infrastructure for AI builders. Because of this, I spend my days watching exactly what these models can do. The blank page is dead. You do not start from zero anymore. You ask a screen for a legal summary, a Python script, or a commercial strategy, and it spits it out in three seconds. It is usually decent. Sometimes it is entirely wrong, but it is always fast, and it sounds confident.

The shift from producing to judging

The fundamental unit of knowledge work has changed completely. When Fredrik and I co-founded Mediapilot, we grew it over roughly seven years and created more than 100 jobs. Back then, the bottleneck for our growth was pure production. We needed people who could physically write the code, draft the copy, and build the campaigns. Production was expensive and slow. Now, production is practically free. The bottleneck has moved. The job is no longer to produce the work; the job is to judge it. You have to look at a wall of AI-generated text or code and decide if it is brilliant, mediocre, or a massive liability. You have to be an editor.

We are turning junior staff into editors before they have ever learned how to write.

Judging work requires a well-calibrated sense of taste. The problem is that you only acquire taste by producing a massive volume of bad work and figuring out why it failed. Later in my career, I worked with Happeo on the commercial side. The best salespeople there were the ones who had made every mistake possible. They knew a bad email because they had personally sent five hundred of them and heard nothing back. They knew when a pitch was too aggressive because they had lost deals by pushing too hard. They had an internal alarm bell built through painful, analogue repetition. If an AI writes the email and builds the target list, the twenty-two-year-old junior never feels the sting of getting it wrong. They just read a plausible-sounding draft and click approve.

The broken ladder

We have entirely broken the traditional path for training junior staff. For decades, the implicit deal in any office was simple. The senior people gave the junior people the boring, repetitive, low-stakes work. The junior did it poorly, got corrected, and slowly learned the craft. That was the ladder. If artificial intelligence is now perfectly capable of doing all the boring, repetitive work in a fraction of the time, what do we give the juniors? Many founders I speak with are ignoring this entirely. They are thrilled to cut costs, run leaner teams, and let the software handle the drafting. They are optimising for this quarter's margins, completely blind to the fact that they are cutting the bottom rung off the career ladder. In five years, they will try to promote these juniors to senior roles and realise they have hired a team of people who lack basic commercial judgement.

The blind spot

I will be honest with you. I have absolutely no idea how we fix this. Every year since around 2010, I have written and updated a life plan. I sit down and rate my family and friends, my health, my finances, my work, and my personal time out of ten. It forces me to look at what is broken and what needs attention. Right now, when I look at the work section, this gap in training is the thing that bothers me most. I am actively building AI companies, and I still do not know how we will train the next generation of workers to have taste. I play tennis down here on the coast badly, but often, and I talk about this with the other guys on the court. None of us has figured it out. We might have to invent artificial sandboxes where juniors are forced to produce work manually just to learn the mechanics, even if a machine could do it faster. Until we find a real alternative to the old apprenticeship model, we are flying blind, hoping experience can be downloaded. I suspect it cannot.