What the model has never seen

I've argued that as production becomes free, the scarce input in marketing isn't taste. It's proximity — knowing what's true of a specific population, which is information that lives in people rather than in text.

I want to push that further, because proximity turns out to be useful for something much larger than choosing between angles.

It's how you find out why a market isn't moving.

The thing you can't see from a desk

I've been developing a framework I call the Five Constraints of Market Formation. Every emerging market faces five things that stop people taking part: Understanding, Access, Trust, Momentum, Coordination. Usually only one or two are binding at any moment, and most money gets spent on the others.

What I've noticed, running this on real markets, is that the five are not equally visible.

Understanding you can measure from a distance. Give people an assessment and find out whether they can explain what the thing is and how it works. Uncomfortable, but tractable from an office.

Coordination is also largely desk-visible. You can map an ecosystem, find the interoperability gaps, work out who would need to be in a room together. It's structural, and structures show up in documents.

The other three are different.

Access looks fine from a desk almost every time. The app exists. The account can be opened. The rails are documented. You only find out that it takes eleven steps and sixty-three percent of people give up at the third one by watching somebody try. I have never once found a real access constraint by reading about it.

Trust is worse, because people don't report it accurately. Nobody says I'm afraid of being defrauded and too embarrassed to say so. They say the app is confusing. You learn what people are actually afraid of by being close enough that they'll tell you the real reason, which is usually the fourth thing they say rather than the first.

Momentum is invisible in any snapshot, because it's about what happens on the second and fifth and twentieth occasion. You can't see a loop in a single observation. You have to be around long enough to watch people either come back or not.

Three of the five constraints — and in my experience, the three most often binding — are only reliably found by presence.

That is what I mean by proximity as a diagnostic instrument rather than a creative one.

Building as a way of finding out

There's a stronger version of this, and I've written about it elsewhere in relation to OPay.

Before it was a payments company, OPay was running a motorcycle ride-hailing business in Lagos. That business could not function without payments that actually settled — the rider has to know they've been paid, the platform has to know the transaction happened.

They didn't reason their way to the binding constraint from a deck. They found it because their own operation broke without it.

Sometimes you can't see the constraint from outside, and building something is how you find it. Which is an uncomfortable thing for someone who sells diagnosis to admit, and I'd rather say it than have it pointed out.

Twenty structured checks are cheaper than building. They are also not sufficient every time.

Why this matters more than the AI conversation

Most of the discussion about AI and marketing assumes the problem is a marketing problem. That the market exists, people understand the product, and they simply aren't responding to the communication.

When that's true, the AI-native model does work. Fewer people, more output, faster execution.

I want to be careful about calling it a good business, though, because I don't think it stays one.

When production costs collapse across an entire industry, the saving does not stay with the producer. It goes to the buyer. It always does. The first movers enjoy a windfall for a year or two; then competitors reach the same cost base, clients — who have the same tools and can see roughly what they cost — reprice accordingly, and the margin transfers. Programmatic did this. Website build did this. Design tooling did this.

And it is worse than margin compression, because the unit shrinks. Work that was a substantial project becomes a modest one. You need several times the client count simply to stand still.

Meanwhile the two historical reasons a client used an agency for production — capacity and specialist skill — are being eroded inside the client's own marketing team at the same time. So the addressable market contracts while the price falls.

Efficiency is not a moat when everybody acquires the same efficiency in the same quarter.

So I'd put it differently. The AI-native production business is a good business for about eighteen months. After that it is a commodity with better tools — unless the firm has something the tools don't.

Which is the same argument as before, arriving from the other direction.

And a great deal of the work I find interesting isn't production at all.

What if people don't understand the category? What if they understand it and don't trust it? What if they trust it and can't physically reach it? What if they take part once and never return? What if the thing they'd need to participate in doesn't exist yet?

You cannot fix any of those by asking a model for better copy. Not because the model isn't good enough, but because copy isn't the intervention. Sometimes the intervention is a rail, a protection mechanism, an agent network, a learning journey, an incentive, a standard agreed between competitors.

And you find out which one by being close enough to see what's actually stopping people — which is the same faculty that produced the good angle, pointed at a larger question.

The uncomfortable part

If proximity is the scarce input, then the value of a firm sits in specific people's access to specific populations.

That is hard to scale. It doesn't transfer cleanly between markets. It walks out of the building at six o'clock, and occasionally it resigns.

You can systematise around it — a fixed diagnostic protocol, named evidence for each finding, predictions registered in advance so the record can be checked. All of that helps, and I'd argue it's necessary. But it doesn't remove the dependency. It just makes the dependency legible.

The discipline is systematic. The input to it isn't.

I'd rather build a business that admits that than one that pretends the model can do the whole job.