What a good agency actually sells now that production is free

A retainer I saw last year had a line for "asset production: 14 versions" billed at studio rates. The client had asked for the campaign in nine aspect ratios across four languages, and a year earlier that genuinely was a fortnight of someone's life, a designer cutting and re-cutting in a quiet room. By the time the invoice landed, a mid-level designer was generating the variants with a tool over a long morning and spending the rest of the week on something else. The hours on the invoice were real on paper. The work behind them had quietly stopped being two weeks of anything.

McKinsey put a number on the size of that gap: generative AI could lift marketing productivity by the equivalent of 5 to 15 percent of total marketing spend, worth roughly $463 billion a year. Read as an operator, that is a repricing of an input rather than a productivity headline. The thing a designer or a copywriter or an editor used to sell by the hour just got cheap, and most of an agency's old margin was sitting on top of exactly that input.

The agencies feel it before the clients say it out loud. The 4As found that 75 percent of US agencies were using generative AI in 2025, up from 61 percent the year before, with 89 percent naming employee productivity as the benefit. Everyone reached for the tool to do the same volume of production with fewer people. Almost nobody changed what they put on the invoice. So the cost of making the work fell inside the building while the price of making the work held on the page, and the difference is a tension waiting for the first client who does the arithmetic.

The part that did not get cheaper

A founder I work with paid two firms for the same thing in the same quarter. Firm one delivered forty social assets, clean, on-brief, fast. Firm two delivered twelve, after a week of arguing him out of the campaign he asked for and into the one his customers would actually respond to, and those twelve outperformed the forty by a wide enough margin that he dropped firm one. He was not paying firm two for the twelve files. He was paying for the week of argument, the part where someone who had seen this go wrong before stopped him spending money on the obvious thing.

That week did not get cheaper, and neither did anything in it. Knowing which of the forty variants is worth running, reading the account when the numbers move, deciding that the right move is to make less and aim it better, building the measurement so you can tell the difference at all: the tool does not touch any of that, and the tool getting better does not touch any of it either. A model will generate the headline, and it will do it well, but it will not tell you the headline is solving the wrong problem, because it does not know what the business is afraid of. Judgement is still a person who has been wrong enough times to recognise the shape of a mistake before it costs anything.

What the invoice should say now

The honest move is to stop selling the output and start selling the system that decides what the output should be. That is a real reframe and it is uncomfortable, because the output is legible and easy to count and the judgement is neither. A client can see forty files. They cannot see the decision not to make the wrong forty, and an agency that has only ever charged for visible things has to learn to charge for the invisible one.

In practice it looks like this. You take the cheap production for granted, the way you take electricity for granted, and you build the thing on top that turns it into a result: the brief that is actually a strategy, the test that tells you which version earns its place, the measurement that closes the loop, the taste that says no to the forty and yes to the twelve. The assets fall out of that work as a by-product. When a client asks what they are paying for, the answer is the reason those assets exist and not the obvious ones.

The agencies in trouble are the ones whose pitch is still "we make a lot of things, quickly, well." That sentence used to be a moat and it is now a description of a tool subscription, and the client will work that out, because the same model sitting on your desk is sitting on theirs. The ones that hold their fee will be the ones whose pitch was never really about volume. They were always selling the judgement, and the production was the thing they wrapped it in because the production was easy to bill. Now that the wrapping is free, they have to sell the judgement on its own, name it, price it, and stand behind it. That was always the actual job. It is just that the invoice used to hide it.