AI in Product Innovation: How Accurately Can an Artificial Target Group Simulate Real Customers?

• Scientist Monika Imschloß on the Simulation of Purchasing Behavior
• Developer Paulo Salem on Interaction in Groups of AI Personas
• Innovator Julian Mangold on the Use of Generative AI for Prototype Testing
• Five aspects to take away

Use AI to create a digital twin of your target group to better understand their needs, test ideas earlier, or evaluate prototypes more realistically. This is exactly what expert Joe Pulizzi recommends in the episode "Unconventional Content Marketing Strategies" of the podcast This Old Marketing, aired at the end of 2024. Pulizzi is a luminary in the field, with over 250,000 followers on LinkedIn.

Pulizzi is supported by around 3,000 market researchers from 14 countries. They predict that, soon, it won’t be humans testing your products, but their AI-generated counterparts. By 2027, "more than half of traditional market research could be conducted using AI-created synthetic personas," write the authors of the Market Research Trend 2025 report.

But how well does this so-called silicon sampling work? And how realistic is this forecast? ISPO has investigated what’s behind the trend, how you can already benefit from it today, and why using synthetic target groups can give you a clear competitive advantage.

Monika Imschloß is a Professor of Marketing at Leuphana University Lüneburg. She has been researching multisensory marketing for years and is now exploring this field using artificial intelligence.

ispo.com: What does science know about synthetic personas?

Monika Imschloß: We are investigating whether large language models (LLMs) can reproduce effects known from marketing research. For example, we used ChatGPT 4.0 to test whether human response behavior can be replicated in synthetic surveys. One example is the "Exciting Red and Competent Blue" effect, where brands with red logos appear more exciting, and brands with blue logos appear more competent.

When we present the logos separately to ChatGPT, the results are less clear. However, if both logos are shown to ChatGPT at the same time, the effect, which we also observe in human test subjects in previous marketing studies, becomes more visible.

What is the conclusion of this experiment with the two logos?

A key question is: How do I prompt correctly? Sensory marketing shows that AI can recognize consistent patterns, but the right prompt strategy is crucial. Comparative prompts, such as "Here are two logos. Which one looks more competent?" perform better than individual evaluations, at least based on current knowledge.

So to what extent can AI simulate target groups?

That depends on the research question. For example, we investigated whether the effects of price reductions, which we know from studies with human participants, are also reflected in synthetic samples. One method is to create a synthetic persona using data from a human sample, such as age, income, or product preferences, and then compare the response behavior of this artificial twin sample with that of the real one. The main effects were mostly preserved, but some variables, such as purchase history, influenced human behavior more strongly than they did AI simulations.

Are there cases in which AI can replicate a target group in a deceptively real way?

This works well with associative learning, such as color associations. There are also promising approaches in qualitative surveys, where AI embodies personas with predefined characteristics.

And what are the limits of AI simulation?

There are psychological effects that are difficult for AI to reproduce, for example, when many unobservable factors influencing human purchasing behavior are not represented in the AI’s text corpora. This remains an open field of research.

If you prompt the AI precisely, you can, in some cases, simulate realistic AI representations of your target group. But what if someone wants to interview more than one person? Paulo Salem, PhD in Computer Science, works as a Senior Data & Applied Scientist. In 2024, he introduced the idea of a person-based multi-agent system, first presenting it to his colleagues at Microsoft, and later releasing it online as an open source version.

ispo.com: TinyTroupe is a framework that simulates AI-driven personas with their own personalities, interests, and goals – designed to replicate human behavior for business scenarios. 

Paulo Salem: Yes. We want to test: Can simulated personas be useful in brainstorming sessions? Do they reflect real customer opinions? Do they provide valuable insights for product developers?

Your tool takes a playful, simulation-based approach. What is the strength of this method? 

Simulations enable rapid experimentation. Companies can play through different scenarios, test assumptions, and iterate on ideas – without using real focus groups. And: mistakes are allowed in the simulation. They help improve products and strategies early on before they enter the real world. 

TinyTroupe consists of two core components: TinyPerson, AI-powered personas with pre-trained personalities – and TinyWorld, a simulated environment where they interact. 

The foundation is powerful Large Language Models, AI systems trained on massive text data. Our framework complements these models with an infrastructure that makes working with them easier. Instead of writing code, a user can define: "I need a 35-year-old marketing manager from Munich who shops price-consciously." TinyTroupe then generates a suitable persona and lets them act within a simulation. The personas also remember past interactions, which provides a form of memory. 

One use case is prototype testing. Are there companies already using it? 

Yes, a company in the education sector is currently testing TinyTroupe for brainstorming new software features. The AI simulates customer discussions to determine which ideas resonate. Early results show that these simulated conversations often closely resemble real meetings. But this is still an experimental phase – there is no scientific validation yet. 

Another use case is advertising analysis?

We took real Bing ads and presented them to TinyTroupe personas, who then gave their opinion on them. The principle: advertisers test ads in the simulation before spending real money on A/B tests. In addition to marketing and product development, we see potential in market research, service design, innovation processes and analyzing how customers use digital products.

A major challenge: many unconscious factors of human behavior are difficult to model.

This is a central problem. LLMs are optimized to be polite, fair and helpful - this leads to distortions. An AI persona will tend to be nice, even if a real customer would be annoyed. We are working on making the behavior more realistic by further developing the interaction rules of the personas.

The next steps in the development of TinyTroupe? 

First: We are working on even more realistic behavior models. Second: We want to make it easier to simulate larger target groups so that users don’t have to define each persona individually. Third: TinyTroupe should become more flexible – with its own simulated tools for AI personas. The AI personas will then, for example, have their own email clients, calendars, or to-do lists. 

What is your vision?

People don’t always act rationally. They change opinions spontaneously, act based on emotions, or are influenced by personal experiences. We are just at the beginning of replicating these nuances in AI personas. But in a few years, these simulations could come much closer – and perhaps even provide surprising insights into human behavior. 

Interim conclusion: We are witnessing the beginning of a new development. But what are the experiences of the first users in industry? Julian Mangold can answer that.
He works at ZEISS Digital Business Innovation, which develops new products in collaboration with internal business units, ranging from medical technology to semiconductor production. The team relies on generative AI and synthetic personas, among other tools.

ispo.com: What does your AI-supported innovation process look like?

Julian Mangold: We combine real customer insights from qualitative interviews with AI-supported simulations. Our research team regularly interviews customers and users, for example, doctors. We use this data to create realistic, synthetic personas. The AI then takes on a kind of role-playing function: it reacts to product ideas as if it were a real customer.

Can you give me an example?

Take a new medical product, such as a holographic display for the operating theater. Without AI, we would directly ask surgeons what they think of it. Today, we’ve also created a panel of ten virtual doctors. These personas simulate different levels of experience, from junior to senior professionals. We then pitch the idea and get immediate feedback: Which features are relevant? Where are there concerns? And where should we follow up with real users?

What target group insights do you gain with AI?

The above example is particularly insightful: We observe differences between user groups. Younger users want more visual support, while experienced doctors prefer greater control. Others prioritize patient interaction or efficiency. AI helps us identify and address these segments more precisely.

Science reports: AI cannot map unconscious factors of human behavior.

AI doesn’t replace real customer interviews. We still conduct traditional user research because people often have needs they don’t explicitly express. In human interactions, we observe behaviors that wouldn’t come up in conversation. One example: A nurse was manually documenting the same procedure in three different places, part of her routine. She wouldn’t have mentioned this as a problem, but we uncovered it through AI-assisted analysis. Our solution is a hybrid approach: human interviews for deeper insights, AI for faster hypothesis testing.a

One criticism of AI is that it’s too polite to reflect real customer opinions. How do you address that?

My colleague and developer Devran "Cosmo" Ünal tested this – using his own AI twin. He was able to teach it to be more critical and adopt its own perspective. We learned a lot from that. After all, it’s not the customer who should adapt to our product, it’s the product that should adapt to the customer.

And in the future?

We plan to make AI even more critical and diverse. Right now, it still thinks in very German-European terms. We’re also working on a scalable platform so that other teams can use it – not just innovation departments, but also marketing and internal communications.