There are hardly any limits to its application—wherever companies want to understand the opinions or behaviors of their (potential) customers, digital twins can be used (Brand et al., 2023). However, as with any method, it is important to be aware of certain limitations when interpreting the results.

✅ Use Cases

• Brand Messaging & Communication Marketing teams can test which messages or advertising slogans resonate with different target groups. The AI personas react to slogans or campaign ideas and indicate what they think about them. This makes it easy to see, for example, whether a particular tone of voice works better with tech experts than with business generalists, or which product benefits should be highlighted.

• Product and Concept Testing Before launching a new product or feature, you can survey the digital target group: How is the product idea received? Which features are most important? Would they be willing to pay for it? This allows for early feedback without expensive test markets or focus groups. Even specific B2B contexts—like feedback from purchasing decision-makers on a SaaS offering—can be simulated in this way, which would otherwise be difficult due to small sample sizes.

• Opinion and Trend Forecasts How would my customer group react to market changes? What happens if a new competitor appears or a particular event occurs? AI can be used to play out hypothetical scenarios. Social trends or policy changes can also be anticipated in terms of public sentiment (Chu et al., 2023). For example: “How would our customers likely react to a 10% price increase?”—the simulated answers at least provide an indication of whether acceptance or dissatisfaction would prevail.

   

⚠️ Limitations to Consider

• Behavioral Change is Needed Using AI tools effectively requires a mindset shift within the team. It takes time to get familiar with prompt crafting, understanding how models respond, and iterating toward useful results. Without training or buy-in, any new approach might be underutilized.

• Generic Responses without Prompt Tuning Language models (especially ChatGPT) tends to give overly generic answers unless well instructed. One needs to fine-tune prompts or use pre-built frameworks to receive differentiated, persona-aligned feedback.

• Limited Contextual Awareness of Current Trends Unless enriched with timely background information, LLMs might miss out on very recent market movements. Either external updates (e.g. via RAG or manual input) or fine-tuned, newer models are required to reflect evolving contexts accurately.

• Pricing Predictions Are Not Reliable Yet Models are still weak in predicting realistic price thresholds or trade-off preferences, especially in B2B environments. Elasticity and willingness-to-pay often require empirical calibration that goes beyond linguistic simulation.

• Image Evaluation Works, but Needs Validation While AI can comment on visuals or ad creatives, the evaluation of images by virtual personas is still experimental and lacks scientific validation. Interpret these insights with caution until broader benchmarks are available.

• Challenges in Representing Diverse Human Perspectives. While LLMs can provide insightful and articulate responses across a wide range of topics, they may not always reflect the full diversity of human opinions—particularly those of specific demographic groups. Even when prompted to adopt certain viewpoints, these models can struggle to reproduce the nuance found in real-world public opinion (Santurkar et al., 2023). This highlights an ongoing challenge in treating LLMs as “silicon subjects”: their outputs are shaped by patterns in training data, which may not fully capture the breadth of societal perspectives. Care should be taken when interpreting their responses in contexts where demographic sensitivity or representational balance is important.
  
• Identifying Effective Conditioning Variables for Modeling Human Perspectives. While persona prompting can help large language models simulate a broader range of perspectives, its effectiveness is currently constrained by the limited explanatory power of the demographic, social, and behavioral variables commonly used. In many subjective NLP settings, these variables account for only a small fraction of the variation in human responses, which caps the potential gains from conditioning on them (Hu & Collier, 2023). This highlights an essential open question: identifying which types of variables—beyond standard demographics—are most informative for guiding LLM behavior in subjective or value-laden contexts. Without more meaningful conditioning signals, attempts to personalize or diversify model outputs may remain shallow or inconsistent.
  
• Lack of Methodological Rigor in Persona Generation Introduces Systemic Bias. The growing use of LLM-generated personas in fields such as public opinion modeling and social simulation offers compelling promise, but many techniques are methodologically underdeveloped. Most persona generation approaches rely on heuristics rather than principled, empirically validated frameworks, resulting in simulations that may reflect systemic biases rather than mitigate them. These limitations can produce substantial deviations from real-world population-level outcomes, as seen in applications like election forecasting and survey modeling (Li et al., 2025)