Generation got cheap, and most of the panic that followed aimed at the wrong target. What collapsed was not the work. It was the description of the work. A retoucher is not finished because a model can clean a bottle in six seconds. Those six seconds moved somewhere else, and where they moved is the only interesting question about AI jobs.
We produce for fragrance and jewellery houses. The engine now returns more usable material before lunch than a crew used to return in a week. That should have emptied the room. It did the opposite. Every hour the engine hands back is an hour that needs a decision attached to it, and decisions do not scale the way rendering does.
The phrase in circulation is that AI replaces jobs. On the floor it behaves more like a change of instrument. A crew of ten does not become a crew of two. It becomes a crew of ten pointed at a different part of the problem, and the part it is pointed at is the part that was always underserved: the brief, the cut, the finish.
So the roster did not shrink. It changed shape. Fewer people moving pixels, more people deciding which pixels deserve to exist.
What actually got automated
Look closely at any creative job and you will find two layers stacked on each other. There is the execution layer: masking, tracking, cleanup, versioning, the ninth pass on a bottle against white. Then there is the judgement layer: is this the right frame, does it belong to this house, would the founder recognise herself in it.
Generation ate the first layer and left the second untouched. It could not do otherwise. A model has no opinion about whether an amber highlight reads as expensive or as cheap, because it has never had to defend one to a client at four in the afternoon.
There is a quick test for which layer a task sits in. Ask whether two competent professionals would arrive at the same answer. Masking a bottle: yes, identically. Choosing which of six ambers belongs to a house: no, and the disagreement is the value. Everything in the first category was always going to be automated, in whatever decade the tool arrived.
The uncomfortable part is that plenty of careers were built entirely inside the execution layer, and those people are now asked to develop something nobody ever paid them for. That is a real cost and it lands on real people. It is also learnable, which is the half of the story that keeps getting cut.
Taste stops being a soft skill
For twenty years taste was the thing you mentioned last, after the software list. Reverse the order.
When output is scarce, leverage sits with whoever can produce it. When output is effectively infinite, leverage moves to whoever can throw most of it away. We generate wide and cut hard. The cutting is the job.
That skill is unglamorous and it is trainable. It comes from looking at a great deal of work while somebody senior tells you why the second frame beats the first, repeated until you can tell yourself. It is an apprenticeship, and the future of creative work now looks more like an apprenticeship than it has in a generation.
The old apprenticeship ran on access to equipment. Nobody needed to teach you why a frame was wrong, because you could only afford to shoot the right ones. That economy is gone. The teaching has to be explicit now, or it does not happen at all.
Rules travel further than skills
Ramón Béjar's library runs on one rule: No Faces, Only Storytelling. Not a preference. A constraint that settles a hundred later questions before anybody asks them. Architecture sets the frame, motion moves through it, light is built for material, and every scene arrives on surface and colour instead.
Anybody on that project can generate. Not everybody can hold the rule. Holding it across a whole library, then knowing which single frame is allowed to bend, is the senior job now. It looks almost nothing like what a senior artist did five years ago, and it is much harder to fake.
This is why the studio line is that we do not generate, we direct. It reads like positioning. It is closer to a job spec.
Where the careers actually go
Three things, in this order. The ability to write a brief that a person and a machine can both follow. The ability to look at forty outputs and say precisely why thirty-nine are wrong. And finish: colour, light, the last pass that moves an image from plausible to premium.
None of that is prompt engineering. Prompting is a syntax, and syntax ages fast. Artificial intelligence careers inside a creative studio are built on older skills, and the tool has not touched a single one of them.
There is a second-order effect worth naming. When execution is cheap, more of the work becomes defensible in a room — you can show the founder the version you rejected and explain the rejection. Junior people are in those rooms years earlier than they used to be, because the thing that kept them out was the hours the execution took.
The people doing best around here were not the earliest adopters. They were the ones who already had a point of view and found a faster instrument for it.