Method, data, and limits

Not by prompting a language model to role-play a persona. We start from census and client data, build statistically weighted agent profiles, layer in media diet and context, then run a multi-step reasoning process on top. The population layer is math-grounded and traceable. The reasoning step is where interpretation happens, not the population itself.

We do not quote a fixed accuracy number, since it depends heavily on category and study design, and we would rather show you than tell you. The way to know is a parallel test: the same question fielded to a synthetic population and to a real panel, compared side by side on your own study.

As a public example, our synthetic population predicted a Swiss federal referendum result before the vote, whereas the official polls did not. See the full prediction →

Two layers. The population layer starts from public sources, census data and other demographic datasets, which sets the baseline structure of the agents. On top of that, if you bring your own customer or study data, it feeds directly into building the agent profiles, so the population reflects your actual base rather than only public averages. No client data is ever mixed into another client's population.

No. Client data is used only for that client's own studies. It is never used to train shared systems and never shared across accounts.

Studies typically run into the thousands of agents rather than a few hundred. You can inspect any individual synthetic respondent's answer, not only the aggregate summary, so you can check the reasoning behind a result rather than take the topline on faith.

Yes, without refielding. Once a population is built for a study, you can return and ask new questions of the same group or a subgroup without commissioning a new wave.

Anything outside the scope of the data and context they were given for that study. We would rather be upfront about where that boundary sits than let a study appear to answer everything.

Both. A digital twin is a synthetic population enriched with your own proprietary data. If you have historical customer data to feed in, the agents get closer to a twin of your real base. If you do not, you still get a synthetic population built from census and public data. Same engine, different depth of input.

A chat model answers as one voice, whatever persona you ask it to hold. Foresenta builds a weighted population first, thousands of distinct agents grounded in demographic and client data, then runs reasoning across that population. You get a distribution of responses shaped by real population structure, not one model's best guess at a single consumer.

Quantitative. Every run gives you a structured, ranked output across a population, not an open-ended conversation. If you need in-depth interviews or open discovery, that is a different tool.