The agency model doesn't have to die. But it does have to become useful again.

A lot of agency machinery existed because producing the work required lots of people. That reason has gone. What replaces it is less obvious than it looks.

I saw a post arguing that AI will eat the old agency model before it eats anything else.

The argument was simple. A forty-person agency full of account managers and project managers is now competing against a very small number of experienced people who can use AI to do an extraordinary amount of work.

I think there’s something in that, partly because I spent about six years inside agencies — and even from the inside, I sometimes struggled to see exactly where all the value was coming from.

There were talented people everywhere. Writers, strategists, creatives, designers, account directors. There was also an enormous amount of machinery around the work. Briefs went in. Meetings happened. People managed the people doing the work. Presentations were built. Feedback passed backwards and forwards. Eventually something came out of the other end.

For a long time clients needed most of that machinery, because producing the work itself required a lot of people.

That justification has largely gone.

The tools are not the moat

So the response I find least interesting is bolting an impressive-looking proprietary AI platform onto the existing model.

The awkward problem is that clients have access to extremely capable AI as well. Nobody is paying a premium for access to a model their own team can subscribe to on a Tuesday.

If the tools aren’t the moat, the thinking has to be. Which sounds like a platitude until you work out what it implies about headcount.

What I’d actually want

Not an agency sitting outside waiting for briefs. One embedded in the evidence — customer conversations, sales calls, product strategy, research, market changes, internal expertise, performance data. The strange thing somebody said in a meeting three weeks ago that suddenly matters because something shifted yesterday.

I’ve argued before that the most valuable editorial thinking falls into the gap between the agency and the in-house team, so I won’t relitigate it. The relevant part here is the staffing consequence.

That agency needs fewer people moving work around and fewer people producing straightforward copy. It needs more people who can walk into a complicated organisation and work out what it actually knows, what matters, what’s missing, what customers are telling it, what competitors haven’t noticed, and what it should have an opinion about.

Part strategist, part journalist, part researcher, part editor, part operator. Small, senior, and close enough to challenge the client rather than service them.

That’s the model most people arrive at. I want to make two arguments against it, because I don’t think either gets said.

The pyramid was doing something

The obvious version of this says you no longer need large numbers of junior people producing first drafts, because AI produces first drafts.

True. But producing first drafts was never only about producing first drafts.

It was how people learned. Somebody wrote something mediocre, an editor took it apart, and eventually they got better — badly and expensively, but it happened. The agency pyramid was, among other things, the training institution for an entire industry. Journalism had one, and it went. Agencies still have one, just about.

So “we need a small group of extremely capable senior people” is a business model with a supply problem built into it. Those people were produced by the structure the model is proposing to remove. It works fine for the first cohort — there are plenty of experienced operators available now — and I have no idea what it does in fifteen years.

There’s a personal version of this that undercuts my opening.

I spent six years unable to see exactly where all the agency value was coming from. Those six years are also where I learned to work. How a brief gets mangled between the client and the person actually doing the job. How decisions really get made, as opposed to how the process says they do. How to read what an organisation wants rather than what it has asked for. When I moved in-house, I already knew what the other side of the wall looked like — and most people who spend a career on one side never find out.

That is a large part of what I bring to a company like Sage, and it sits underneath most of Editorial Intelligence. Knowing both machines is the thing, not knowing either one.

I didn’t learn it from the interesting bits. I learned it from the machinery.

So when I say the pyramid was doing something, this isn’t a general point about training somebody else’s juniors. I’m the output of it.

I’ve written about this gap before from the in-house side and I still can’t solve it. Every version of the argument that senior judgement is the scarce thing has this hole in it, including mine.

Embedding removes the thing you were buying

The second objection is more uncomfortable, because it cuts against the model I’ve just described wanting.

A lot of what an agency actually sells has nothing to do with production capacity.

It can say things an insider can’t. It carries no internal political history. It can be pointed at a problem and then fired. It absorbs blame. It arrives without needing to protect a relationship with the person whose strategy it is about to question. Its distance is not a bug in the model — for a certain class of problem, the distance is the product.

“Get closer to the business” trades that away. An embedded partner who attends the meetings, knows the politics and depends on the relationship becomes subject to the same pressures as the in-house team — which is exactly why the in-house team couldn’t say the thing in the first place.

I don’t think that kills the embedded model. It does mean the strongest version of it has to be deliberate about keeping some distance rather than treating proximity as the whole answer. Close enough to know what the organisation knows. Far enough to still say it.

Consultancies have the same problem, harder

If your value proposition is gathering information, synthesising it, and turning it into a polished deck of recommendations, clients can now do a great deal more of that themselves.

Again the answer isn’t better AI tooling. It’s judgement, domain expertise, original evidence, interpretation, and the ability to see something the organisation couldn’t see by itself.

The difference is that consultancies charge more for the synthesis layer, which is the layer that got cheapest fastest.

Which might be healthy

AI isn’t necessarily killing agencies or consultancies.

It is making it much harder to hide the difference between doing a lot of work and creating a lot of value. Volume used to be evidence of effort, and effort was a reasonable proxy for cost. Neither survives contact with a model that produces the volume for nothing.

That distinction was probably overdue. It just arrives with two problems attached — where the next generation of senior people comes from, and what happens to independence when everybody gets closer to the client — and I notice the confident versions of this argument tend not to mention either.