3 minutes
There is a quiet assumption running under most of the conversation about synthetic data that if a machine can simulate a customer, the researcher who used to talk to that customer becomes less necessary. It is rarely said out loud, but it’s shaping almost everything written about this technology right now, including the wave of thought pieces warning organizations not to over rely on AI-powered simulations. The warnings are not wrong. But they are answering the wrong question. The real question is not how much synthetic data an organization can tolerate before it becomes a problem, but rather what happens to the people who do research, once part of that research can run itself?
Our answer is that their job gets bigger, not smaller.
Let’s start with what this is, because the term “synthetic data” is a disservice. What we build are bespoke simulations, models trained on real human research, not fabricated data points, that are shaped to reflect a real segment or persona and extend the value of the work that already happened. A researcher spends months building a rich picture of a customer type through interviews, surveys, and behavioral data. Historically, that picture lived in a deck, got presented once, and slowly went stale in a shared drive. A simulation allows that work to stay alive. Product teams can stress test a concept against it at 11pm before a Tuesday deadline. Marketing can evaluate test messaging without waiting weeks for a new study. Rather than a replacement for talking to real people, it is a way of making sure the talking you already did keeps paying off.
That reframing matters, but it is not the most interesting part. The most interesting part is what it does to the researcher’s role inside an organization.
Insights teams have always been the people an organization turns to when it needs to know something true about its customers, not simply something plausible. That standing does not go away with AI technology. If anything, it becomes more valuable, because insights teams are the only people equipped to build and maintain something reliable enough to trust. Anyone can upload a folder of documents to a general-purpose AI tool and ask it to talk like a customer. That takes only an afternoon and produces something that sounds convincing, which is exactly the risk. What it does not produce is intentionality about what went into the model, or judgment about what is missing.
The work that actually makes a simulation trustworthy looks a lot like the work insights professionals already do, aimed at a new kind of output. This translates to deciding which findings are load bearing and which are noise. It means noticing the gap between what you know about a segment and what you are being asked to simulate, and commissioning fresh primary research to close that gap rather than letting the model guess. It means understanding that a document is not just a document, it is a set of claims with different levels of confidence and different shelf lives, and that the model needs to know the difference. And it means doing something a chatbot left alone will never do on its own: actively deprioritizing data that has gone stale, so the simulation does not quietly turn into an echo chamber that just repeats last year’s version of the customer back to you with more confidence than it deserves.
None of that is a technology problem. It is a research discipline problem, applied to a new format. And the people best equipped to solve this problem are the people who have been solving versions of it for their entire careers.
This is also where the democratization argument gets even more interesting, and worth being specific about rather than waving at. When a simulation is built well, a brand manager who has never sat in on a research readout can ask it a direct question and get an answer grounded in actual customer truth, without waiting on a research team’s calendar. That is real and it is valuable. But it only stays valuable if someone is maintaining the model being queried, deciding when it needs new input, and deciding who gets to touch it. Access without governance is just a faster way to be confidently wrong. The insights team is not sitting on the sidelines while data gets democratized. They are the ones who make democratization safe to do at all, by shaping the build, keeping the content fresh, and controlling who can use it and how.
There are a lot of tools in this space right now, and most of the conversation about them is either breathless or defensive. We would rather be neither. We think the honest case for this technology is not that it replaces research, and not that it is dangerous and must be contained, but that it gives insights professionals a genuinely new job to be done, one that sits inside the full research cycle rather than bolted onto the end of it as an activation gimmick. The organizations that get the most out of this will be the ones that put their research function in charge of it, not around it.
That is the bet we are making. Human truth in, human judgment throughout, and a tool at the end that is only as good as the discipline behind it.