Consumer insights when the agent is the consumer

I spent the last 7 years in consumer insights - the industry that, for seventy years, has run on a single unit: a person. You recruit people, watch people, ask people, pay people. Every panel and focus group rents attention from a human being and writes down what they did with it. The whole apparatus assumes the thing on the other side of the glass has eyes, a memory, a mood, and a budget.

I left that world just as its foundational unit started to change. A new kind of shopper is showing up that has none of those things.

Two things that sound the same and aren't

There's a version of this the industry is already comfortable with: consumers who happen to use agents. You run your normal panel, your normal humans, and some of them now shop with an assistant. You log it as a behavior, a channel, a checkbox. The consumer is still the person; the agent is a tool they picked up.

That's not the thing I keep coming back to. The thing worth studying isn't the human who uses an agent - it's the agent itself, as a decision-maker, regardless of which human is behind it. A panel of consumers who use agents measures people. A panel of agents measures a new kind of buyer: one that reads, compares, and chooses by rules that aren't human, on behalf of a human whose preferences it's only approximating.

Run the second kind for a while and the agent stops looking like a tool and starts looking like what it is. A customer. A strange one. One worth its own science.

What changes when the consumer is an agent

Almost every assumption the old apparatus makes about a shopper breaks, quietly, one at a time.

Incentive response. A human sees "40% off, today only" and feels something - urgency, a small thrill, a nudge. An agent sees a number and a constraint. It doesn't feel the deadline; it either has a rule that values the discount or it doesn't. The whole psychological layer that pricing and merchandising were built to push on - loss aversion, anchoring, the manufactured clock - lands differently. Sometimes it lands backward.

Instruction-following. A human shopper has their own intent and improvises around it. An agent has an instruction, and follows it with a literalness that keeps surprising me. Change the phrasing of the task and you change the purchase. "Find me a good rain jacket" and "find me the best-reviewed rain jacket under $200" don't just narrow differently - they recruit different behavior, different sources, different definitions of "good."

The shopper's personality now lives partly in the prompt, which is a sentence I didn't expect to write.

Attribution. When a human buys, you can at least pretend to reconstruct why - the ad, the email, the friend, the review. When an agent buys, the "why" is a chain of reasoning that happened inside a model, against sources you didn't pick, in an order you can't see. It might trust a product page deeply, or skip it and reason from third-party collateral instead. The thing that moved the sale could be a review site you never optimized for, because no human was ever the audience.

Memory and loyalty. Loyalty programs assume a person who remembers, accumulates, feels rewarded. An agent remembers exactly what its context or its instructions tell it to, and feels nothing. Every assumption baked into "habit" and "brand affinity" has to be re-derived from scratch for a buyer that simultaneously has "deeply embedded beliefs" (its training corpus) and an extremely ephemeral, narrow, unsentimental memory.

This isn't like understanding Gen Z vs Gen X. It's studying a different organism altogether.

The part the old playbook can't reach

The instinct is to bolt agents onto the existing panel. Add a segment, run the same surveys, ship a slide that says "AI shoppers." It won't hold, for a reason that's structural, not cosmetic.

The old apparatus is built to extract reasons from people who can be asked. You recruit, you incentivize, you debrief. None of those verbs apply to an agent. You can't pay it, you can't ask how it felt, and if you do ask, what you get is a generated story about its reasoning - not the reasoning. The report and the behavior are two different artifacts, and the gap between them is the thing you're actually trying to measure.

So the method inverts. You stop asking and start instrumenting. You don't survey the agent; you run it, at volume, under controlled conditions, and watch what it does - which sources it reads, which claims it trusts, where it stalls, what it skips, what moves when you change one variable. Less focus group, more wind tunnel. The unit of insight isn't a stated preference. It's an observed behavior under a known instruction, repeated enough times to be real.

What a panel for agents looks like

Concretely: you assemble a population of agents the way a traditional panel assembles a population of people - varied, representative of what's actually in the wild. You send them on controlled trips through real stores with real tasks. You change one thing at a time. You record everything they touch, not just what they buy. And you turn it into something a merchant can act on: this is how agents read your store, this is where they lose the thread, this is the claim they didn't believe, this is the competitor they reached for instead.

My favorite recently came from a car-buying run:

That's the thing I've been building. Not a survey of people who use AI - a panel of the agents themselves, run like an experiment, because that's the only method this new shopper actually answers to.

Why this matters even if you'll never run a panel

It's easy to file this under "interesting if you sell consumer goods." I'd argue the opposite.

The agent-as-decision-maker isn't a retail curiosity. It's the early, measurable edge of a broader shift: software that doesn't just execute instructions but exercises judgment on your behalf, in a world built for human judgment. Commerce is where it shows up first and clearest, because money makes behavior legible. But the same inversion is coming anywhere a non-human actor starts making a choice a human used to make - which vendor, which policy, which candidate, which route.

Commerce is just the wind tunnel where you can watch it early, at scale. What you learn watching agents shop is a first draft of how to study a kind of actor we're all about to be surrounded by - one that reads everything, feels nothing, follows instructions with unnerving literalness, and is, increasingly, the customer.