We Are Now Conducting Surveys Without Surveying People

There is a sentence in HCLTech's newest wealth-management research that sounds completely ordinary until you look at how the research was done. Eighty-four percent of wealth-management firms, the company says, believe their operating models need a fundamental redesign to take full advantage of artificial intelligence. 

That is the kind of statistic we have been trained to read without stopping: big sample, global markets, senior decision-makers, a clean percentage ready for a chart.

Except the 1,066 respondents in this survey were not 1,066 wealth-management executives. They were AI-generated personas modeled to represent them. And that changes what the number means.

A Survey Without the Surveyed

HCLTech released its "Hidden In Pl(AI)n Sight" report describing it as a first-of-its-kind synthetic research study for the global wealth-management industry. The project was conducted with Evidenza, a company that builds synthetic audiences for market research. Instead of recruiting senior wealth-management decision-makers, asking them questions and aggregating their answers, the researchers created a synthetic audience designed to reflect the characteristics, behaviors, expertise and decision-making patterns of that population.

The questionnaire was then run across those modeled personas. HCLTech says human experts were involved in designing the personas, developing the research and validating the findings. So when the report says 84% of firms believe something, that figure is not a count of executives who selected an answer on a questionnaire; it is an aggregate of responses produced by a modeled population intended to behave like those executives would.

That may sound like a semantic distinction. It is not. For roughly a century, survey research has rested on a premise so obvious that we rarely bother to say it aloud: if you want to know what people think, you ask people. Synthetic research inserts a new step. 

You build statistical and computational representations of those people and ask the representations what the people would probably have said. The familiar survey statistic survives. The human respondent does not.

What Does 84% Mean Now?

This is where the methodology becomes more interesting than the findings. Readers already know how to interpret a traditional survey percentage, even if they do not know the mathematics behind sampling. If 84% of a properly drawn group of respondents chooses an answer, the percentage describes something that actually happened: people were asked a question and 84% of the sample answered a particular way.

A synthetic 84% is different. It is a property of a model run. That does not make it meaningless; models can reveal patterns, test assumptions, simulate scenarios and help researchers work with populations that are expensive or difficult to recruit. Senior executives are exactly the sort of respondents who are difficult to assemble by the thousand across many countries.

But the number is no longer simply an observed response rate. It is a prediction about response behavior rendered in the visual language of an observed response rate. The chart still says 84%. The reader now has to ask: 84% of what?

Not Fake People - Modeled People

It would be easy to make synthetic research sound ridiculous by reducing it to "AI pretending to be executives." That would also miss why companies are interested in it. Human market research has real constraints: respondents take time to recruit, narrow professional populations can be difficult to reach, executives are expensive, response rates can be low and global studies can take months. Synthetic samples promise speed, scale and repeatability without finding another human participant every time the sample grows.

Evidenza, the platform used for HCLTech's report, says it creates detailed synthetic personas from target-audience characteristics and existing research, then runs qualitative interviews and quantitative surveys against those modeled respondents. 

The company says it has performed more than 100 validation tests comparing synthetic and traditional research and reports an average similarity score of 88%. Those are Evidenza's own claims, and the broader research literature remains unsettled.

Academic work has found that large language models can sometimes reproduce broad survey averages surprisingly well. Other studies have found lower response variation than in real populations, sensitivity to prompt wording, changes over time as models change, and difficulty accurately representing some demographic groups or the relationships between variables. Synthetic research may be genuinely useful without being interchangeable with human research. The important part is knowing which one you are looking at.

The Most Dangerous Word May Be "Survey"

HCLTech is not hiding the synthetic nature of the project. The press release puts AI personas in the first sentence, and the methodology page explicitly contrasts synthetic research with traditional surveys that collect responses directly from individuals. That transparency matters.

But the resulting findings are still presented in the same rhetorical form readers have seen for decades: "84% believe." "98% are pursuing." "Nearly 80% believe." Percentages have authority because they imply measurement, and once synthetic respondents generate the same kind of percentages, the distinction between measuring an opinion and modeling an opinion can become visually invisible.

Imagine the statistic separated from the methodology page and reposted in a presentation, a LinkedIn graphic, a sales deck or a news story: "84% of wealth-management firms believe..." How many readers will naturally finish that sentence with the invisible clause, "according to 1,066 AI personas designed to predict what senior decision-makers would say"? That missing clause may become one of the most important clauses in modern market research.

The Methodology Is Already Hard to Summarize Cleanly

There is even a small but revealing discrepancy in HCLTech's own public material. The press release says the 1,066 personas were modeled across 17 global markets, while the report methodology page currently describes 1,000-plus modeled perspectives across 15 geographies. That may be nothing more than a publishing inconsistency; traditional research reports have typos and mismatched summaries too.

But synthetic research raises the stakes of methodological clarity because readers cannot inspect the respondents in any ordinary sense. There is no roster of executives who participated, no response rate in the conventional meaning and no fieldwork story involving actual people deciding whether to answer. 

The sample frame, model construction, validation procedure and prompt design are not background details. They are the instrument. When the respondents are generated, explaining how they were generated becomes inseparable from explaining what the findings mean.

The Strange New Category Between Data and Prediction

We already have words for observed data, and we already have words for forecasts. Synthetic research sits somewhere awkwardly between them. It is not simply made up: the personas are constructed from information about real populations, and their outputs can be tested against human responses. But it is also not observation in the traditional empirical sense.

The modeled executive did not wake up, open an email, consider the question, worry about how their company would look, interpret the wording through years of experience and click an answer. The model generated a response that the research system judged representative of someone who might. That distinction gets especially interesting when the topic is unstable, private or newly emerging.

A model can only simulate from what it has learned and how it has been instructed to represent the target population. People, meanwhile, surprise researchers. They contradict their demographic profile, misunderstand questions, change their minds, hold inconsistent beliefs and react to events nobody predicted. Human messiness is not always noise in a survey. Sometimes it is the thing being measured.

Useful Tool, Dangerous Shortcut

None of this means synthetic research should be dismissed. It may be excellent for testing questionnaires before fielding them, generating hypotheses, running scenarios quickly, exploring hard-to-reach professional populations or identifying areas worth investigating with humans. It may eventually become good enough in some settings that refusing to use it looks as quaint as refusing computer-assisted analysis.

The danger comes when simulation quietly inherits the authority of observation. If a synthetic study produces a number that looks exactly like a conventional survey number, readers need a way to tell the difference immediately. "84% of AI-modeled wealth-management personas responded..." is clunkier than "84% of firms believe..." It is also more accurate.

Market research spent decades teaching the public to recognize sample sizes, margins of error, weighting and polling methodology. Synthetic research is going to require another layer of literacy: model provenance, persona construction, validation benchmarks, prompt stability, versioning and whether findings were ever checked against humans from the actual target population. Those details sound technical. So did margin of error once.

The Question Has Changed

The most interesting thing about HCLTech's report is not whether wealth managers need to redesign their operating models for AI. They probably have plenty of reports about that already. The interesting thing is that a company wanted to know what a difficult-to-reach group of executives thought and decided it could build 1,066 computational stand-ins, ask them instead and publish the resulting distribution as research.

That is a remarkable change in how evidence can be manufactured, interpreted and circulated. We used to ask whether a survey sample accurately represented the people we wanted to understand. Now we may also have to ask whether the sample contained people at all.

And that leads to a much stranger question than anything in the wealth-management report: when a model of a person gives an answer, at what point do we begin treating the prediction as an opinion? Maybe synthetic respondents will become one of the most useful research tools of the AI era. But if they do, readers will need to learn one new reflex very quickly.

Before asking what the survey found, ask who, or what, actually took it.

SOURCE NOTES

• HCLTech via PR Newswire, Sept. 28, 2026 — “HCLTech launches first-of-its-kind AI-based synthetic research study for the global wealth management industry”
• HCLTech — Capital Markets and Wealth Management (full “Hidden In Pl(AI)n Sight” report)