Let’s talk about the twenty-six-year-old who was supposed to be running that place in ten years, and the fact that nobody hired her.
Picture a commercial law firm in Sandton, 2019. Twelve associates. The first two years of their lives are dull in a way that’s hard to describe to someone who hasn’t done it. Data rooms at eleven at night. Reading the same indemnity clause in four hundred different contracts, marking up where the limitation of liability sits, flagging the one paragraph that’s been quietly changed. Bad coffee. Bad pay per hour. Everyone complains.
That was the curriculum.
Now picture the same firm this year. A model reads the four hundred contracts before lunch, pulls the clause variations, produces a redline, and flags the eleven that need a human look. It’s better at the dull part than the associates ever were. Faster, cheaper, doesn’t need to be talked into staying for the weekend.
So the firm cuts the intake. Of course it does. The partner who made that call got a bonus for it, and he was right on the numbers in front of him.
That’s the thing I keep circling. He was right. The numbers were real. And the decision was still wrong, because the numbers he was looking at stopped at the end of the financial year and the cost doesn’t show up until 2034.
Here’s the mechanism, and it’s not really about AI.
Every profession hides its school inside its worst jobs. Medical interns do the admissions paperwork. Article clerks sit in the discovery. Junior developers fix the broken import at two in the morning and learn, painfully, how the system actually fits together — not how the architecture diagram says it fits together. Audit graduates tick and tie. Site engineers stand in the sun counting rebar.
None of that work is productive in any meaningful sense. Partners could have done it. They did do it, once — badly, slowly, with mistakes that got caught by somebody more senior who had made the same mistake five years earlier.
The mistake was the tuition.
And here’s why nobody ever called it tuition: the firm paid you while you learned. There was no invoice. The cost was buried inside a graduate salary and a bad utilisation rate and a partner grumbling that juniors these days couldn’t draft. The whole apprenticeship system ran on the fact that the learning was a byproduct of the work, so it never had to be justified as a line item. It looked free because it was attached to something that produced revenue.
Detach it, and it becomes a cost somebody has to defend. And nobody in a partnership defends a cost centre with an eight-year payback.
That’s the whole story. It’s an externality with an unusually long fuse. The person who cuts the rung is not the person who pays for the missing seniors, and the gap between cause and consequence is longer than any review cycle, any tenure, any bonus structure. So the cut is always locally rational and collectively stupid, which is the definition of a problem that markets do not fix on their own.
Now — the obvious objection, and it’s a fair one. Juniors with a model do ramp faster on some things. A sharp graduate with a good agent can produce in a week what used to take a month, and that’s real. I’ve watched it happen.
But producing is not the same as being responsible, and I don’t think people have caught up to that distinction yet.
The difference between someone who has shipped and someone who has generated is a memory of consequence. Not knowledge. Memory. The specific, physical recollection of being the person whose name was on the thing when it broke. The model will never hand that back to you at three in the morning. It has no three in the morning.
People say the junior can just learn from the model. That misses what a job is. A tutor and a job are different teachers. A tutor explains. A job makes you answerable for something that went out into the world with your name on it, and the fear of that is what makes a person careful. You cannot get careful from a chatbot. You get careful from having been careless once and having to sit in the room while it was discussed.
Now the part that should worry the business owner more than the partner.
You can’t supervise work you’ve never done. That’s the actual reason the apprenticeship existed — not charity, not tradition. It was how the supervision layer got built from the bottom up. A senior partner’s value is not that he reads faster than a model. It’s that a trained reflex fires in him before any analysis happens. He looks at a document and something itches. He couldn’t tell you why for another twenty minutes, but the itch is right.
That itch took eleven years to build. There is no version of it that arrives from reviewing model output alone, because output looks correct. That’s the entire point of output. Correctness is what a model produces by default; being wrong in a way that looks fine is the failure mode you’re trying to catch.
So the risk isn’t that you’ll have fewer seniors. It’s that you’ll have seniors who can’t check. And a senior who can’t check is not a cheaper senior. He’s a liability with a title.
I want to be honest about the scale of this, because I don’t have a number for you and I’m not going to invent one. I don’t know how many firms have already hollowed out the bottom. What I know is the shape: the tasks disappeared faster than the replacement learning arrived, and the replacement is being sold by the same people who made the tasks disappear.
The second-order effect lands on the buyer. The client whose contract was reviewed by a person who has never negotiated one. The patient whose scan was read by a radiologist who trained on a system that pre-labelled everything. The passenger. The borrower. They don’t know the difference, and they won’t find out until the one case where the machine was confidently wrong.
So what do you actually do about it, if you run a team.
Keep the rung. Change what’s on it.
Give the junior the review seat instead of the production seat. Make her the first check on machine output, and make her write down what she caught and what she missed. Pattern recognition is built by comparing a machine’s answer against reality, and she’ll see more edge cases in six months than the old associates saw in three years.
Make her explain it out loud, to a person, once a week. If she can’t explain why the clause matters in plain language, she doesn’t own it yet. Fluency is the trap, remember — a model is always fluent. You’re training the thing fluency can’t fake.
Put her on call for something real. Something with a pager, and a client who will be annoyed, and a consequence that isn’t hypothetical. Consequence is the only teacher that doesn’t get cheaper and doesn’t scale. That’s precisely why it’s still worth paying for.
And budget the years as research, not headcount. Because what you’re buying is a person who in 2034 can stand in a room and say “this is wrong” with enough conviction that the room stops. That’s not a cost. That’s the only insurance policy that ever actually paid out.
The uncomfortable part is that you’ll be asked, in about four years, why your team costs more than the benchmark. And the honest answer — “because I’m paying for judgement that doesn’t exist yet” — will sound like an excuse to whoever is holding the spreadsheet.
Hold the line anyway. Somebody has to. The alternative is a building full of very fast systems and nobody in it who can tell when they’re lying.
There’s a line I keep coming back to, and I’ll give it to you plain. You can rent intelligence by the token. You cannot rent eleven years of somebody’s judgement. Competence has a birth date. Every attempt to compress the years has produced someone confident and hollow, and confident and hollow is the worst combination to have standing between a machine and a consequence.
So hire the kid. Give her the boring work and the pager and the blame. Ten years from now she’ll be the only one in the building who can see the thing the machine got wrong — and you’ll be very glad you paid for that when it was still cheap.
The machine can do the work. It cannot become the person who knows when the work is wrong.
That takes a kid, a bad year, and somebody willing to pay for both.