How AI Actually Changed: From Novelty to Infrastructure

The interesting thing about the last few years of AI is not the capability jump. It is the change in where the technology sits. It stopped being something you opened and became something that runs underneath what you were already doing.

That is a quieter shift than the headlines suggest, and a more consequential one.

The first wave was about astonishment. You typed something, it produced something, and you showed a colleague. Almost none of it survived contact with real work, because the output was impressive in isolation and unusable in context — wrong tone, invented facts, no awareness of the project it belonged to.

A lot of people formed a permanent opinion during this phase and never revisited it.

The second wave made it useful for drafts. First-pass copy, transcripts, summaries, code you then corrected. The value was real but bounded: it shortened the distance from nothing to something, and a person still did the last eighty per cent.

This is where most individuals still are, and for a lot of work it is a perfectly reasonable place to stay.

The current shift is that AI has moved into workflows rather than windows. It sits inside a system, triggered by an event, doing one narrow job repeatedly — tagging enquiries, drafting a reply for approval, generating captions from a transcript, reconciling a report.

In this phase you mostly do not see it, which is the point. The measure stopped being how impressive the output is and became how much of somebody's week it gave back.

Judgement. The model still has no idea which of its plausible answers is the true one, and no stake in the outcome if it is wrong. Everything consequential still needs a person between the output and the world.

Nor did it change the value of knowing your own domain. AI raised the floor for generic work and left the ceiling exactly where it was — which has widened, not narrowed, the gap between people who understand what they are producing and people who do not.

The opportunity is not replacing people. For a three-person studio there is nobody to replace. It is that the repetitive half of the work — the passes between a shoot and a delivery, the admin between an enquiry and a quote — can now be handled by a system, which is what gives a small team the output of a larger one.

Start with the task you do most often and enjoy least. That is almost always where the return is.

I use AI across research, production and the systems I build. I do not use it to write anything involving a real person's story, and nothing reaches a client without a person having read it.

That is not caution for its own sake. It is that the parts of this work worth paying for are the parts that require someone to have actually decided something.