Why the Richest Data Set in the World Still Can't Be Simulated

On paper it's a simple transaction, an asset changing hands. In reality it's got nothing to do with money. It’s about which sibling will take it personally. What the neighbors will assume. Whether letting go of the house means admitting the family isn't "that family" anymore, not really, not the way it used to be. None of that shows up in a press profile. None of them shows up in a CRM record or a LinkedIn post either. It exists in exactly one place: a conversation with someone the family actually trusts, and nobody outside that room will ever hear it.

If you asked an AI model to predict that decision, what would it even be learning from?

That's the debate right now in HNWI market research. Synthetic respondents and AI-generated data are moving fast, and honestly, it's not hard to see why. They're quick, they're cheap, and for a lot of categories synthetic data for market research is genuinely good enough. Ask a model to act like someone picking a streaming service, or buying a pair of sneakers, and it'll do fine. It has millions of real examples sitting behind it.

Have it taken a decision to run at the beach-house instead, and there's almost nothing there. Which is the whole problem.

No corpus, because none of it got written down.

While we use AI at Jasper Colin- deliberately, responsibly, and within clear compliance guardrails- we also believe firmly that primary research on UHNWIs has to stay human-led. Here's why.

What synthetic data actually is, plainly: information generated statistically resemble real respondent behavior, minus the real respondents. Fine when the underlying record is deep. Not fine when it isn't. That's the entire story of synthetic data in research: a model is only a mirror of what it was trained on. For most consumers that's not a problem. For the ultra-wealthy, it is, because the public record is thin and a little strange. Mostly rich lists and press profiles, plus the occasional interview about a yacht. It's a record of what this audience is comfortable being seen doing, which has almost nothing to do with what they're actually doing.

It won't tell you how a family office really negotiates a second home sale. It definitely won't tell you what shifted someone's philanthropic priorities two years before anyone outside the family caught on.

Push a model to reason from that thin record anyway, and it doesn't fail loudly. It gives you something fluent. Well-structured. Confident. It reconstructs the surface convincingly and misses everything underneath, which, for this crowd, is basically the whole point. And that's really the crux of the whole argument here: the failure mode isn't a bad answer, it's a good-sounding one, and the two are much harder to tell apart than people assume until they've watched it happen on a real account. A model doesn't know what it doesn't know. It just keeps talking.

The real reason remains unsaid.

People rarely give the real reason when you ask them directly about a big decision. You get the version that sounds reasonable instead. Legacy, family politics, the quiet influence of a private office, wanting to be seen a certain way by one very specific peer group. That stuff doesn't get said plainly, doesn't get typed into a chatbot, doesn't show up in transaction records. It surfaces in one place: a trusted, one-on-one conversation, once a researcher has earned enough rapport to get past the rehearsed answer.

Where AI actually pulls its weight

None of this makes us AI skeptics. We use it constantly as part of our own AI-driven B2B research: synthesizing large volumes of qualitative material, speeding up analysis, spotting a pattern across quant and qual faster than a person could manually. It's increasingly part of how we approach AI-driven HNWI buyer segmentation, too, once the underlying primary data actually came from real conversations in the first place.

What we don't do is let it generate primary findings, or stand in for a real person, on this audience specifically. That's a deliberate line; not caution we'll eventually grow out of. Give us another decade of wave-on-wave transcripts and proprietary data on this exact segment, and predictive modeling might genuinely earn its place, built on our own depth, not borrowed from a public record that was never built to support it. We're not there yet. Pretending otherwise just means handing a client a well-formatted answer to a question the data can't actually answer.

To be fair to technology: there are real benefits of synthetic data in B2B research. Filling gaps in lower-stakes categories, mostly, or pressure-testing a questionnaire before it goes anywhere near real respondents (there's a decent argument for modeling directional scenarios too, if speed matters more than precision that day). Used that way, synthetic data for B2B insights saves real time without touching the findings a client will actually act on. This isn't a trust issue. Just a matter of knowing where the tool is reliable and where it isn't.

Convincing AI, not bad AI

Nobody's worried about synthetic data for HNWI studies producing something obviously broken. That would be easy to catch. The real risk is the opposite: something polished, and wrong underneath. Most categories can absorb that; the market corrects for it over time. At the UHNWI tier, where entire strategies rest on a handful of qualitative signals instead of statistical weight, there's no averaging-out effect to catch the mistake. A wrong answer here doesn't just miss. It misleads, and the mistake is usually expensive by the time anyone traces it back to where it started.

Same logic holds for anyone doing AI in wealth management research more broadly. Advisors, private banks, family offices, they're all circling the same question about where AI tools can safely sit in their own client’s work. The honest answer doesn't change depending on who's asking, let it speed up the edges, and keep the human conversation right where it's always been, at the center, doing the part it's always done, the part nobody's figured out how to automate yet and maybe never will.

The wealthiest, most private people in the world don't show up in a spreadsheet. They show up slowly, privately, in conversation, rarely in exactly the same words twice.

How we do it at Jasper Colin

Our contribution isn't refusing to use AI. It's knowing exactly where to put it.

Every engagement starts the same way: real conversations, with real UHNWIs and family offices, conducted by researchers who've spent years building the kind of trust that gets past the rehearsed answer. That's the raw material. Nothing synthetic touches it, and nothing skips it.

AI enters after that, not instead of it. We use it to process the volume- hours of qualitative interviews, years of proprietary panel data, quant and qual sitting side by side- and to surface patterns a person would take weeks to find manually. It's how a single conversation about a beach house becomes a signal we can track across an entire segment.

The result is a body of primary data that's genuinely ours: built specifically on this audience, refreshed wave over wave, deep enough that when we do apply AI-driven segmentation or synthesis, it's reasoning from something real, not from a public record built for rich lists and press profiles.

That's the model. Human-led where it has to be. AI-accelerated where it earns its keep. And a depth of proprietary UHNWI data that gets harder for anyone else to replicate every year we keep at it.