When AI Learns to "Impersonate a Customer": Market Research Is About to Be Rewritten

A few days ago, I had tea with a founder who runs a consumer-goods company.

He vented to me about user interviews: expensive, slow, exhausting. Bring in ten people, one hour each, transcribe the recordings, write it all up into a report — the whole ordeal takes a month. And the conclusions at the end? Not necessarily anything that actually guides a decision.

"Tell me," he sighed, "when does this ever get easier for people?"

I didn't answer him directly. But I had a hunch.

It's getting rewritten. By AI.

Not gradually tweaked. Uprooted, the whole thing.

One: First, Four Tiers

The people who study this stuff have already mapped the path. AI is rewriting market research along four tiers.

Which four?

Tier one: Make the old work faster, cheaper, easier to scale.

Tier two: Use AI-generated "synthetic data" to stand in for part of the real data.

Tier three: Fill in the insights that used to be out of reach or out of budget.

Tier four: Generate a kind of data that simply didn't exist before.

Let's go tier by tier.

Two: Efficiency — Pulling People Out of the Drudge Work

What do people who do research hate most?

They hate digging through recordings.

One interview transcript: two hours, tens of thousands of words. You sit there, line by line, picking it apart. By the time you're halfway through, your eyes are glazed and you can't remember what they said at the start.

It's different now.

Someone ran the numbers: of the practitioners already using AI, 62% put it to work on the exact same task — summarizing long interviews. Of the rest, about half use it to analyze data, and half to write reports.

In plain terms, AI took over all the repetitive, time-sink work that still needed a human in the loop.

A founder working in this space said something I keep coming back to. He put it like this: when the only thing AI is doing is "chatting with a person," people open up completely, and the AI almost never hallucinates.

Sharp.

Because researchers used to face a forced choice — do interviews, and get depth, but can't scale the sample size; or send out surveys, and get breadth, but the answers stay shallow. One head of user research told me you finally don't have to choose. Now you get both.

That's tier one. Take every old pain point, and grind it down, one by one.

Three: Substitution — When AI Starts to "Pose as a Customer"

This tier is bolder.

What's "synthetic data"?

In plain terms, it's having AI generate a batch of synthetic personas, feed them personality, age, income, and preferences — then have them fill out questionnaires and answer questions in place of real people.

Sounds far-fetched?

One company actually did this. They took an annual brand survey normally given to CEOs, handed the questionnaire to AI, and had it generate over a thousand matching "synthetic CEOs." Then they let those synthetic CEOs fill it out.

They compared the results — against the real CEOs' answers, the conclusions matched 95%, with strong correlation, and some of the numbers were nearly identical.

When I heard that, my reaction was: wow.

But hold on before you get too excited.

Of that same group of respondents, only 31% rated the AI-generated data as "valuable" — the lowest score on the list.

Why?

Because synthetic is still synthetic. It can imitate "how most people would answer in most situations," but it can't capture the sudden, irrational, off-script real thing. People change their minds out of nowhere. People buy things for utterly baffling reasons. That kind of stuff, AI can't learn.

So the right way to use this tier is as an accelerator, not as the real thing. Run the synthetic data first to scan the broad trends, then spend real money on a small, real-world study to confirm.

Four: Augmentation — The Moment of Decision Finally Has Data

This is the tier I'm most bullish on.

Think about it: how do most decisions inside a company actually get made?

By gut.

Not because people don't want to use data — because there isn't time. By the time you spend two months producing a research report, the window has already closed.

But AI is different. It's always on.

At 3 a.m. you're agonizing over whether to change a price, and it can hand you an answer grounded in historical data plus scenario modeling. No waiting, no budget approval, no scheduling.

Someone ran a survey: 81% of practitioners are either already using AI to monitor market movements, or planning to. 30% are already deploying it on decisions that "previously would never have touched data."

The role AI plays here is sparring partner. It doesn't make the call for you — it walks you through your assumptions, pressure-tests your options, and flags the traps you might step into, ahead of time.

Five: New Data — When a Customer Gets "Copied"

The last tier is the most sci-fi.

It's called "digital twins."

What's a digital twin?

You take a real customer's data — everything they've bought, searched, clicked, complained about — feed it all into AI, and have AI replicate a "virtual them."

And then? Then marketers can test a campaign on the digital twin before the real customer ever gets bothered.

What's the upside of a digital twin?

It doesn't get tired, doesn't get annoyed, and doesn't blow up at you for asking the same question ten times. It's just always there, letting you run the experiment again and again.

Already 40% of practitioners are experimenting with this.

One university research team is doing something even bolder. They're building 2,500 digital twins, each mapped to a real person — and every one of those real people first completes a battery of psychological, behavioral, and cognitive tests as the "foundation." Then AI uses that foundation to construct a matching virtual person.

Once built, can those virtual people stand in for the real ones in research?

A research team tried it. They had the real people answer the same questions again two weeks later — and when compared, the digital twins' answers reproduced the real people's responses with 85% accuracy.

85%.

Is that enough?

Depends on what you're using it for. For trial and error, sure. To replace real research — not even close.

Six: But Don't Deify It

Having said all that, I have to throw some cold water on this.

AI's problems in research are anything but minor.

First, it has bias. Its training data is what it is, so it leans toward people who look like that data. Researchers have found that mainstream models express opinions that sound more like liberal-leaning, highly educated people — and less like adults over 65, or people who are religious.

Second, it's bad at predicting sudden shifts. What it learned is the past. So it can tell you "how people used to choose." But "where the next disruptive inflection point lands" — that, it doesn't actually know.

Third, its answers drift. Take the same batch of virtual people and ask them again three months later — the answers will have shifted. Not because they learned anything new, but because they're inherently unstable.

Fourth, it's extremely sensitive to how a question is asked. Same question, different wording or a different order of answer choices, and the answer moves with it. So the old rules of survey design still apply. Randomize the options. Keep questions neutral.

One team tried having AI simulate "if the price changes, will people still buy?" The demand curve it drew didn't just differ from what real people produced — it didn't even hold up economically.

This is where AI can't pass for human.

Seven: So, What Do You Do?

Back to the founder's question at the start: when does this ever get easier?

My answer: it already has — just not in the way you think.

AI isn't going to make the decision for you. What it does is:

Compress what used to take a month of "reference input" and put it on your desk in ten minutes.

After that — whether to trust it, whether to bet real money on it, whether to go all in — that's still on you.

Its role is closer to a co-pilot who's always online, never loses patience, and has read a library full of case studies.

But the steering wheel stays in your hands.

I keep coming back to a line lately: when a tool can impersonate a customer and chat with you, what becomes scarce isn't data — it's judgment.

Data is everywhere. AI picks it up for you.

But which of those pieces are gold and which are sand — that, AI can't do for you.

Only you can.