From 6 to 100 personas: the (im)possibilities of a synthetic panel

“You're not crazy if you hear a hundred voices”

Thanks to AI, we see a promising innovation in this area: personas are brought ‘to life’ . You can interact with them — they respond to questions, think along with you via an interactive platform, or have a conversation with each other in a virtual focus group.

The introduction of these digitally accessible and interactive personas opens up a world of new possibilities — precisely because of their scalability .

Because why limit yourself to 6 personas, 8 Mentality environments or 14 GeoTypes? Wouldn't it be cool to have a hundred synthesized personas at your disposal - each representative of 1% of the Dutch population? This way you can work much more finely in one go, by combining general demographic characteristics with domain-specific situations and personal motivations.

I envision a virtual panel of digital characters — based on real data, domain-specific knowledge, and behavioral profiles. AI agents that respond from their unique perspective when you present them with a research question. As if you were talking to 100 people at once. Not as a gimmick, but as a valuable addition to your research process.

In this article I will show you how to build such a synthetic panel, what challenges come with it, and most importantly: why now is the time to experiment with it. The potential is great — but the practice is still in its infancy.

How it works: from microsegment to AI agent

The basis of the synthetic persona panel is a set of 100 micro segments, based on a large-scale representative consumer survey. We cluster the respondents into 100 micro segments with a clear profile — each with its own behaviors, preferences and context. Each micro segment represents 1% of the Dutch population and gives color and depth to the customer landscape.

Each AI agent is given its own identity

For each microsegment, we then build an AI agent — a digital representative that “thinks” and “responds” from the perspective of that specific segment. The agent is given a unique prompt, based on the data from the research. This prompt includes information about demographics, context, and personal preferences, as well as tone of voice, language, and attitudes toward the topic.

This means that if you ask the panel an open question , you will get 100 different answers — each from a different perspective. And if you ask a closed question , you will get a frequency distribution of the answers, which gives an indication of how different types of people in the Netherlands might respond.

Accuracy grows with targeted input

The accuracy of these answers and frequency distributions depends largely on the distance of the topic from the original data. Questions that rely heavily on measured behavioral or attitudinal information yield robust patterns. For topics that are far removed from the basic data, the results should be seen more as inspiring input — a qualitative simulation, not a representative judgment.

Over time, you can continue to increase the accuracy of the panel by explicitly investigating new themes . By conducting additional research on specific topics and feeding this data to the AI agents, you enrich their perspective. This creates an increasingly broadly applicable panel that grows with the questions that arise.

First practical example: working together with KPN

In a pilot with KPN , we are building a panel of 100 AI agents, based on segmentation across three dimensions: demographic profile, telecom situation and personal preferences. The result is a set of digital conversation partners that clearly represent different perspectives.

KPN can use the panel to test initial ideas, discover nuance in customer experience or refine communication. It is especially valuable in the exploratory phase — when you are still feeling out what works, for whom, and why.

AI agents are just like humans (but not quite)

Of course, AI agents are not real people. They base their responses on data and context, but they remain simulated representations. Especially for topics that are somewhat removed from the original input data, it is important to be careful with hard conclusions. So don't see them as a replacement for quantitative or qualitative research, but as a complementary tool — with which you can collect input from the target group in an accessible way.

We are only at the beginning. The technology is there, the first pilots are running. That is why I am inviting you: let's do a pilot together . No theory, but practice. No promises, but discovering together what works — and what doesn't.

So, will you hear a hundred voices when you ask a question? Don't worry, you're not crazy — you're just using a synthetic panel as it was intended.

‍