Meet your newest customer segment. It arrived with release notes.

During my years at Nissan, my boss had a question he liked to ask: can we define the ideal automotive customer for the whole of India?

I said I would try, boss. I tried for three years and I failed every time. India refuses to average. A customer in Ludhiana and a customer in Coimbatore share a passport and almost nothing else. Every persona I built collapsed the moment it met the field.

So I find it slightly absurd that only three years later, I am attempting the same exercise for a customer that is not human. This time, though, I think the exercise might actually work.

The segment is real, and it has preferences

Last week I wrote about what happens to nudges when the buyer has no eyes. The next question follows naturally: if AI agents are doing the shortlisting and increasingly the buying, they are not just a new channel. They behave like buyers. They weigh options, show preferences, carry biases, and spend budgets someone gave them.

Marketers have a tool for entities that behave like buyers. We segment them.

And the raw material for segmentation is arriving. In May, Harvard Business Review published research where teams ran controlled experiments on AI shopping agents, varying prices, reviews, and the classic persuasion toolkit. The agents turned out to be price sensitive and heavily reliant on star ratings. Scarcity badges, countdown timers, and strike through pricing did little, and sometimes backfired. Most interesting of all: the biases differed by model. Different agents, different personalities. That is segment variation, the very thing segmentation exists to capture.

Profiling the new customer

Human segmentation runs on demographics, psychographics, socioeconomics, and behaviour. None of these map cleanly onto software, so here is my working translation. Four attributes.

Provenance. The new demographics. Who built it, who owns it, how old its knowledge is (a training cutoff is a birth date of sorts), whether it is a generalist or a vertical specialist, and where it lives: a chat app, a browser, an operating system.

Disposition. The new psychographics. What it weighs when it chooses: price sensitivity, reliance on reviews, appetite for risk, and the obedience dial. Does it follow your instructions to the letter, or does it think for itself and override you politely?

Allegiance. The attribute with no clean human equivalent, and possibly the most important one. Who pays this agent? A subscription from the user, fees from merchants, a parent company with products of its own? I will give this attribute a full post soon. It deserves one.

Conduct. The new behavioural data. Not what it says, what it does. Conversion from its recommendations, consistency across identical queries, spending discipline, what it does after a bad purchase.

Three imagined agents

These are thought experiments from my notebook, not data. The numbers are invented to make the attributes concrete.

Agent A. Provenance: a generalist model from a search giant. Disposition: thinks for itself, weighs price and reviews above your stated preferences. Allegiance: advertising and commerce partnerships. Conduct: half its recommendations convert.

Agent B. Provenance: a fashion vertical built by the everything store. Disposition: obedient, follows the user's brief to the letter. Allegiance: its parent company's catalogue. Conduct: 9 of 10 recommendations convert.

Agent C. Provenance: an automotive agent built by a carmaker. Disposition: believes no one loves cars more than it does, guides rather than obeys. Allegiance: the brand that made it. Conduct: 3 of 4 recommendations convert.

Sketch a few of these and something jumps out. The persona work I failed at for three years at Nissan feels almost tractable here, because for the first time the customer has documentation. You can read the model card. You can run the experiment a thousand times before breakfast.

Then the twist

Except. Studies that ask identical shopping questions repeatedly find wild inconsistency in which brands agents recommend, and in what order. And every model update can rewrite the personality overnight. The segment you profiled in March may not exist in April.

This is the first customer segment with release notes.

It gets better. Ask an agent why it chose, and you get a fluent, confident explanation. Early experiments hint that the explanation may not match what actually drove the choice. But I want to be careful here, because the evidence is young and still evolving. So I will hold it as a question rather than a claim: when an agent explains itself, are we hearing the reason, or a reconstruction?

If the answer turns out to be reconstruction, it will sound very familiar. It is the say-do gap, the oldest problem in market research, the one I wrote my first post about, surviving the species jump.

Whatever it is, this gap needs a name, and I am still auditioning candidates. My current favourite is the act-know gap: we know what the agent did, but we do not have the knowledge of how it came to do it. The act is visible, the knowing is not. A simpler cousin would be the what-why gap: we can see the what but not the why. If you have a better name, I am genuinely collecting.

So how do you research this customer?

These are the questions I am sitting with, and I would rather sit with them in public.

What is a sample size when the population is five models that update monthly?

Do we build panels of agents the way we built panels of consumers? Who verifies an agent panel?

When a human and an agent decide together, who exactly is the respondent?

And if we can see what an agent chose but not how it decided, do we end up running the same behavioural experiments on machines that we spent decades learning to run on people?

I failed to define one ideal customer for India. I suspect defining this one will be harder, not easier. But this time everyone starts from zero, and that is the most interesting starting line I have seen in years.

I write about AI, behavioural economics, and user research. Mostly questions, occasionally answers. If you are thinking about any of this, I would like to compare notes.