Our co-founder Anne Scherer wrote for Swiss Insights on the four methodologies behind AI customer twins, and why the setup, not the technology, decides whether the insights are actually valid.
In this piece for Swiss Insights, the platform of Switzerland's professional community for market and social research, our co-founder Anne Scherer makes the case for what genuinely valid AI customer twins look like, and why most teams are using a version that produces generic, unreliable results.
The argument is concrete. AI twins are not a single technology. They span four distinct methodologies, ranging from the basic level (a simple persona described in a prompt, where the model generates plausible but generic answers) to the most rigorous setup, which combines fine-tuning on company data with retrieval-augmented generation (RAG) that pulls in real-time market and study data. The validity gap between the cheapest and most rigorous methods is enormous, and most companies do not realize where on that spectrum they sit.
Anne argues that for AI twins to deliver insights worth acting on, three things need to be in place: real-time adaptability through RAG, broad and high-quality data integration, and a deliberate choice between commercial models like GPT-4 and open-source alternatives like LLaMA or Mistral. Skip any of those, and the insights are plausible at best, misleading at worst.
The piece is in German, written for the Swiss research community, and is the most direct statement we have published on what good customer simulation actually requires.
Read the full article on Swiss Insights (in German)
