The Price of Execution Is Collapsing. Most Agencies Are Priced on Execution.

There is a comfortable story the agency industry is telling itself about artificial intelligence, and it goes roughly like this: AI is a productivity tool, we will adopt it, we will do the same work faster, and we will keep the difference. It is a story with an obvious appeal, because it requires nothing to change.

It is also wrong, and it is wrong in a way that is easy to demonstrate. Productivity gains are only retained by the people who produce them when the thing being produced is scarce. When the tool that produces it is available to everybody — to the agency, to the client, to the client’s competitor, and to a freelancer in another country — the gain does not accrue to the producer. It is competed away, and it is passed to the buyer. That is not a prediction. That is what happens to the price of anything that becomes easy to make.

What actually got cheap

It is worth being precise about what has changed, because the imprecision is where the panic lives. Generative models have collapsed the cost of a specific category of work: the production of competent, unremarkable, on-brief output. Twenty variants of an ad. A landing page. A quarter’s worth of social copy. A first-draft strategy deck that hits all the expected beats. A campaign concept that is fine.

This category is enormous. It is, if we are honest, the majority of what the marketing services industry has been selling — and it has been priced not on its difficulty but on the number of hours it took a person to do it. The billable hour was always a proxy for scarcity. When the hours disappear, the proxy stops working, and the client notices immediately, because the client can now produce the same output themselves in an afternoon.

What did not get cheap is the part nobody was itemising on the invoice: knowing which of the twenty variants is the right one, knowing whether the campaign worked, knowing what to do when it did not, and being accountable for the answer.

AI collapses the cost of execution. It does not collapse the cost of judgement. The advantage moves to whoever can tell, quickly and honestly, whether the machine was right.

The uncomfortable middle

The agencies in the most danger are not the small ones and they are not the large ones. They are the ones in the middle: too big to be a specialist, too small to own a measurement or data advantage, and structurally dependent on selling volume of output at a headcount-derived price. Their entire commercial logic is that producing marketing material is hard. It is no longer hard.

Their clients know this, and the client-side conversation has already changed. It is no longer “can you produce this”. It is “why am I paying you to produce this”. There is no good answer to that question if the honest answer is that you own a piece of software the client could also buy.

Three things that survive

If execution is not defensible, what is? We would argue three things — and notably, all three are harder to buy than a model licence.

• Judgement under uncertainty. Choosing what not to do, which market to enter, what a brand must refuse. This is not a prediction problem; it is a decision problem with incomplete information, and models are structurally bad at it because there is no training signal for a road not taken.
• Measurement you can defend. The ability to say, credibly and against a control, that a thing worked — and the willingness to say it did not. This is a capability, not an opinion, and it takes years to build.
• Accountability. A model cannot be responsible for an outcome. It cannot be fired. It cannot stake a reputation. Someone still has to, and buyers pay for that more than they admit.

The perverse consequence: proof gets more valuable, not less

Here is the second-order effect that most of the industry has not priced in yet. When output becomes free, output becomes abundant. When output becomes abundant, it stops being a signal. A beautifully produced case study, a polished deck, a confident claim on a website — none of these mean anything any more, because a competitor with no clients and no track record can produce all three before lunch.

In a market flooded with cheap, plausible, well-written assertion, the only thing that carries information is verifiable proof. Not a claim about a result: a documented result, with a method attached, that a sceptical buyer could interrogate. That has always been valuable. It is now the only thing that is.

This has an unglamorous implication for how an agency should actually spend the next few years, and it is not “build an AI tool”. It is: get rigorous about measurement, document outcomes properly, and become the kind of firm that can prove things. That work is slow, it is boring, and it cannot be generated.

What we are doing about it

Fabulous.Media is a network of specialist marketing agencies, and the network is being built deliberately rather than assembled quickly — two agencies operating today, the rest in formation. That pace is not an apology; it is the thesis. A network whose value rests on judgement, measurement and accountability cannot be stood up in a funding round, because none of those three things can be bought.

We use AI heavily, and we are candid about where: prediction, allocation, forecasting, anomaly detection — the places where the model produces a number that reality will shortly grade. We are equally candid about where we do not rely on it: deciding what a company should be, and telling a client something they do not want to hear.

The agencies that survive the next decade will not be the ones that adopted AI first. Almost everyone will adopt it, and quickly. They will be the ones who were selling something other than the thing AI made free.