Creative direction in the age of AI

For most of media's history, making the thing was the hard part. You paid for the hours it took to write the line, cut the film, set the type. The cost of production was the moat, and taste rode along inside it, bundled with the labor.

That bundle has come apart. Agents now make a competent version of almost anything in minutes, at a cost that rounds to nothing.¹ The work got cheap. What did not get cheap is the decision about which work is worth doing, and which should never have been made.

When production is free, the only thing left to pay for is judgment. The studio of the next decade does not sell hours. It sells the refusals.

Taste is a series of refusals.

It is a fair question, and the honest answer is not the one an agency wants to give. You do not need an agency to make more things. You can do that yourself now, at two in the morning, for the price of a subscription.

What you need is harder to buy. You need someone who has seen ten thousand versions of this decision and knows which one holds. Someone to set the direction, hold it across every channel, and say no to the ninety pieces that almost worked. The agency that survives is not the one with the most hands. It is the one with the clearest point of view.

The old line was simple: to make more, you hired more. Output and headcount climbed together, and the second one set the ceiling on the first. Every brand that wanted to do more eventually ran into the cost of the people it took.

Agents break that link. Output can climb while headcount stays flat, because the volume moves to the system and the people stay on the judgment.² The two lines, once locked together, finally come apart.

This is the part the demos skip. A model can imitate a point of view, convincingly, on demand. What it cannot do is hold one. It has no stake in the outcome, no memory of the times it was wrong, no reason to prefer the harder, truer line over the one that tests well.

A point of view is not a style you can prompt. It is an accumulation of choices made by someone who has to live with them.

Strip out the person, and you do not get a neutral point of view. You get none.

The work becomes fluent and forgettable, which is the precise texture of most of what the machine makes when no one is steering.

For a while, the human in the loop was framed as a hedge. A safety net. The thing you mention quickly so no one worries the machine is unsupervised. That framing has it backwards.

When anyone can generate the work, the person who decides what is worth making is not the apology for the system. They are the reason to choose it. The human read is the luxury good. It is the scarce thing in a market suddenly drowning in the abundant one, and scarce is what luxury has always meant.

Judgment is the scarce thing now.

In practice it is quiet. A brief goes in. Agents draft the campaign, the content, the pages, the cutdowns, faster than any team could. Then a person reads all of it, and most of it does not survive the read. What ships is the small part that cleared the bar, in one voice, on the record.

It does not look like a robot doing a human job. It looks like a studio that finally has the time to care about the decisions, because it stopped spending its hours on the production. The machine made the work possible. The person made it worth looking at.