There is an agency I work with. Small team, digital marketing. Earlier this year they had a role budgeted, office coordinator, somebody to take the slack off a team that was spending most of its week orchestrating. Chasing, scheduling, formatting, keeping things moving. Then they started using Orca, which is the thing I build, and a chunk of that work stopped needing a person. They were growing at the time, so what happened next was not a redundancy. They just grew into the hours that came back. The role never got opened.
Nobody lost a job. That is the bit I keep coming back to. There is no redundancy figure anywhere with that role in it, and if you went and asked them, which I have, they would tell you they have had a good year. They would be right.
Bill Gates published an essay on Wednesday about AI and work, and in the middle of it there is an idea he is calling Human Reserved. Jobs machines could do perfectly well, that we decide to keep for people anyway. He compares it to a nature reserve, land you could build on and choose not to because the loss would be too great. He got there through the carers who looked after his dad through Alzheimer’s, and he says what they did was irreplaceably human, which, having watched something fairly similar, yeah.
Then he does the thing almost nobody does, which is say out loud that he has not got the answers. His words: who gets to decide what we reserve for humans, what criteria should we use. And then, those will need to be worked out in public.
So that is what I have spent the last day on.
The test he is working with is whether something is irreplaceably human. I would tighten that slightly and say the real test is whether you have to be in the room. And that covers a lot more than care work, which is the useful part of it. You have to be in the room to work out what is wrong with the wiring in a hundred-year-old house. A model cannot know what a century has done to a wall by looking at a photo of it. Same way a carer does not need an algorithm to work out that the bloke in the bed wants another pillow.
But be honest about it and it does not stretch to people who take meetings for a living. The coordinating layer, the chasing and scheduling and formatting and summarising that most junior professional jobs are actually made of, fails that test completely. None of it needs a body in a room. Which is exactly why it goes first.
And that is the problem, because that is also where people learn.
An office coordinator does not stay an office coordinator. That job is how you end up in operations, and then running operations, or sideways into project management. What makes somebody good at that is not on a course. It is knowing why things run late, why some weeks move and others do not, why a plan survives one client and falls apart with another. You learn it by being the person the mess goes through.
We hired at the garage in June. Vehicle technician, MOTs and servicing. Apprentices there start at the bottom, sweeping up, making the tea, oil changes, the boring repetitive stuff. On paper that is the most automatable work in the building. In practice it is the bit that teaches the trade. The repetition, and making mistakes on things where a mistake does not cost much, is what lets somebody handle a hard job three years later.
I have built the alternative to that, which is part of why I am sceptical of it. Last year I built the simulation tools at Enterprise Skills, a business I have since exited, and they were meant to teach exactly these capabilities through scenarios you could fail safely. They worked better than nothing. But you do not sweat in a sandbox. You can fail a simulation and feel absolutely nothing, because nothing was at stake. Competence you can read for. Judgement needs consequence.
So the criterion I would put next to his is this. Reserve the work that is the only route to the judgement above it. And the way you would actually use that on a real job is to ask whether the competence at the next level up can be learnt any other way. If it can, automate it, you have lost nothing but the cost. If it cannot, that rung is load-bearing, and taking it out does not save you a salary, it takes out a stair.
Which leaves who decides. Not a new committee, and not government drawing up a list of protected occupations. It should be whoever is already accountable for producing the next generation of judgement in a profession, and in most professions that body already exists. The Law Society defines what counts as qualifying work experience. The accountancy institutes sign off practical experience. Royal colleges set supervised hours. Apprenticeship standards say what an apprentice has to have actually done. Every one of those is already a decision about which experience counts, already written down, already enforced.
So they do not need a new power. They need to answer one question and publish the answer. Which parts of qualifying experience can now be produced by a machine and still count, and which cannot.
There is one more thing, from the GeekWire interview he did alongside the essay, that I have not been able to put down. He has been working the idea through with Claude, at three in the morning apparently, and he talks about getting the share of work reserved for humans up to 40 per cent, using shorter working days and earlier retirement to spread what is left around.
Both of those move work about between people who have already got it. Earlier retirement takes people off the top of the ladder. Shorter days thin every rung a bit. Neither one makes a training position. So you could get to 40 per cent and end up with a big, generous, well-meant reserve staffed entirely by people who got their experience before the reserve existed, and no way of producing any more of them.
I should be careful with the numbers, because the ones that get quoted are not as clean as people would like. Adzuna had UK entry-level roles down by about a third since ChatGPT launched. When that came out, the joint chief executive of the Institute of Student Employers said it was too early to blame AI, and that a flat economy plus the national insurance rise were doing most of the work. Indeed’s economist said much the same thing.
They are probably right, and I think that makes it worse rather than better. If the bottom rung is already thinning for ordinary boring economic reasons, the hole is already open. What AI changes is the repair. A downturn ends and firms start hiring juniors again because they need the hands. Automation does not end.
Then there is the stuff I cannot answer. Enforcement, mainly. Any boundary you draw gives somebody a reason to relabel work so it sits on the profitable side of the line, and what you would get is supervision that is nominal, logbooks signed for work a model did, roles that are junior in the title and empty in the middle. You can photograph a building going up in a nature reserve. You cannot photograph whether anybody learnt anything. I have not got a mechanism for that, and I am not certain I have got the shape of the problem right yet either.
I have written the long version up properly, with the quotes and the sources in it, and it is here: simmance.ai/the-next-rung/human-reserved
If you have got a view on the criteria question I would like to hear it. He said it needs working out in public. This is me working it out in public.
The Next Rung is out on 19 January 2027 with Practical Inspiration Publishing.