Method and system for the simulation and analysis of consumer behavior by means of interacting ai instances

Abstract

The invention relates to a system and a method for the simulation and analysis of consumer behavior by means of interacting AI instances.

Description

[0001] Method and system for simulating and analyzing consumer behavior through interacting AI instances

[0002] The invention relates to methods for simulating and analyzing consumer behavior and experience through interacting AI instances and a computer-aided system for simulating and analyzing consumer behavior. Introduction and prior art

[0003] Market research serves to capture and analyze consumer behavior in order to provide companies with a sound basis for decisions regarding product development, brand strategies, and advertising campaigns. Traditionally, these methods are based on in-depth qualitative interviews, focus groups, and quantitative surveys. While qualitative methods provide profound psychological insights, they are time-consuming and expensive. Quantitative quick tests, on the other hand, allow for rapid data collection but often remain superficial and do not provide detailed explanations for consumer decisions.

[0004] To make market research processes more efficient, AI-supported systems based on large-language models have been developed in recent years. These systems are designed to simulate synthetic consumer interviews or answer standardized questions. However, existing AI-supported market research solutions have significant shortcomings that severely limit their application in demanding research scenarios.

[0005] A key problem lies in the lack of psychological depth and realistic consumer reactions. Standard AI models often react too rationally or generically and fail to reflect the emotional, intuitive, or conflicting decision-making processes of real consumers. While human consumers are shaped by unconscious motives, cognitive biases, and situational influences, existing AI models tend to provide simplified, idealized answers. For example, an AI-powered system might classify an advertisement as a positive message without recognizing that real consumers might perceive it as exaggerated or unbelievable.

[0006] Another shortcoming concerns the lack of consistency and individualization in surveys. Many AI models are unable to simulate stable consumer personalities with coherent decision-making behavior. When surveyed multiple times, they can produce contradictory or illogical answers due to a lack of a fixed psychological structure. For example, an AI might indicate in one survey that a consumer prefers a particular car brand, while in another survey it makes the opposite statement, without any comprehensible explanation. Furthermore, existing AI market research systems lack the capability for exploratory surveys and dynamically adapting questions. Conventional systems typically operate linearly with predefined questionnaires and are unable to ask follow-up questions or respond specifically to unexpected answers.While a human interviewer can flexibly respond to answers to, for example, explore consumption preferences or emotional reactions in more detail, many AI-supported systems rely on rigid scripts.

[0007] Another shortcoming lies in the lack of methodological validity of many existing AI-powered market research solutions. Numerous systems are based on purely statistical probability models, without incorporating sound psychological theories into the response generation. As a result, the generated responses may be computationally plausible, but not necessarily psychologically accurate or scientifically valid. For example, an AI-supported system might classify an advertisement as exciting simply because the word "exciting" frequently appears in similar contexts, without any actual analysis of the advertisement's emotional impact.

[0008] In addition to these structural weaknesses, many AI-supported market research solutions fail to consider nonverbal and affective factors. Human consumer decisions are strongly influenced by unconscious factors such as body language, facial expressions, tone of voice, and affective reactions. Existing systems that rely solely on text-based interactions are unable to capture or simulate these dimensions. While a human consumer might hesitate or react ironically to an advertising message, an AI typically provides only a neutral or highly simplified response.

[0009] Furthermore, existing AI systems lack the ability for iterative advertising optimization. While traditional market research teams adapt and retest advertising materials after initial trials, many AI-supported systems offer only a one-time evaluation without targeted suggestions for improvement. True optimization through AI would only be possible if the system could independently generate alternative advertising messages or designs and retest their effectiveness.

[0010] Object of the invention

[0011] There is a significant need for AI-supported market research technology that meets several requirements. Firstly, it must simulate consumer reactions based on sound psychological principles and realistic decision-making processes. Data collection must be dynamic and adaptive, allowing questions to adjust based on previous responses. Secondly, the system must enable in-depth market research scenarios that precisely analyze the impact of price changes, advertising, and packaging design on consumer behavior. Finally, it should be able to generate automated recommendations for marketing and product strategies based on realistic consumer reactions.

[0012] Technical solution

[0013] The present invention provides an AI-supported market research system consisting of several specialized AI instances. These include consumer AI twins, which, as digitally trained models, simulate realistic consumer reactions based on depth psychological principles. In addition, one or more moderator AI twins are provided, which employ adaptive questioning techniques to conduct exploratory in-depth interviews and dynamically guide targeted questioning. A software-supported communication protocol ensures structured interaction between the AI ​​twins, thus guaranteeing a consistent, realistic simulation of consumer behavior. Furthermore, the system features an iteratively learning optimization logic that automatically generates recommendations for action based on simulation results and, if necessary, suggests alternative advertising or product strategies.

[0014] The invention preferably comprises a modular AI architecture consisting of specialized agents interconnected via a server-side communication protocol. This architecture can be based on containerized AI models that can be operated in a distributed system environment. The consumer AI twin and the moderator AI twin can communicate with each other via an asynchronous messaging interface (e.g., WebSockets or RESTful APIs) to ensure low latency and high processing speed. This architecture can be further extended by an event-driven processing model that supports the processing of parallel simulation runs.

[0015] By combining these elements, the invention enables a highly automated, yet psychologically sound market research methodology that efficiently combines qualitative and quantitative research elements. The data collection according to the invention, performed by the moderator-Kl twin, preferably takes place as an automated workflow.

[0016] Process aspect of the invention

[0017] The invention relates to a method for simulating and analyzing consumer behavior using specialized AI instances. The method includes the initialization of an AI-supported agent system comprising at least one consumer AI twin and one moderator AI twin. The consumer AI twin is a trained large-language model that generates simulated consumer reactions based on real consumer responses and psychologically grounded meaning networks. The moderator AI twin acts as a virtual interviewer, conducting adaptive interviews in real time and dynamically controlling the responses of the consumer AI twins.

[0018] The procedure begins with a simulated survey process in which the moderator-client twin generates adaptive, dynamic questions. Rule-based decision mechanisms and semantic rules are used to adapt the questions to the response patterns of the consumer-client twins. The consumer-client twins generate answers based on psychologically modeled affect and motivation parameters, generated using stochastic probability models. A real-time comparison between simulated consumer responses and reference datasets allows for the identification of deviations in the modeling and the derivation of correction values ​​for the adaptive behavior of the consumer-client twins.

[0019] The process further includes iterative optimization of the consumer AI twins using machine learning. For this purpose, empirically obtained consumer interviews are transformed into vector-based semantic representations, which are then embedded into the large-language model. A fine-tuning process optimizes the neural network weights of the consumer AI twins through backpropagation based on validated consumer responses. Continuous self-evaluation is performed through a difference analysis between real and simulated responses, with automated rule adjustments to the model parameters.

[0020] To dynamically adapt the simulation model to the specific research context, psychological meaning networks are modified, activating specific consumer typologies and behavioral patterns depending on the research objective. The consumer AI twins undergo automated recalibration of response patterns through reinforcement learning, receiving feedback on their generated responses and dynamically adjusting their preference and decision logic.

[0021] The simulated consumer reactions are provided via a digital interface. The responses from the consumer AI twins are delivered in a structured output format, such as JSON, XML, or an API-compatible format for market research systems. An analysis unit aggregates the responses from the consumer AI twins and prepares them for data-driven marketing or product decisions.

[0022] System Aspect of the Invention: The invention further relates to a computer-based system for simulating and analyzing consumer behavior. The system comprises a network of specialized AI instances, including at least one consumer AI twin and one moderator AI twin. The consumer AI twin is a trained large-language model that generates simulated consumer reactions based on real consumer responses and psychologically sound semantic networks. The moderator AI twin simulates the behavior of a human interviewer, conducts adaptive interviews, and dynamically controls the reactions of the consumer AI twins.

[0023] The system features a software-based communication protocol that governs the interaction between the specialized AI instances. The moderator AI twin poses questions to the consumer AI twins, using an adaptive decision tree and semantic rules to optimize the questions. The consumer AI twins generate answers, taking into account psychologically modeled affect and motivation parameters and utilizing stochastic probability models.

[0024] The system comprises a multi-stage training pipeline that enables the gradual improvement of the consumer AI twins. This is achieved using exploratory, in-depth morphological interviews with real consumers as reference data. The collected responses are integrated into psychological meaning networks and semantic annotation models to specifically train the large-language model. Iterative fine-tuning based on real consumer reactions improves the simulation quality of the consumer AI twins.

[0025] To increase model fidelity, a comparison and calibration module is provided. This module compares the simulated responses of the consumer Kl twins with real consumer interviews and algorithmically adjusts the response patterns. Statistical similarity metrics assess the simulation fidelity, while a dynamic correction mechanism continuously optimizes the weighting of affective and cognitive parameters.

[0026] Simulated consumer reactions are provided via a real-time interface to market research. The API interface enables integration into existing market research systems by outputting structured data formats such as JSON, XML, or SQL-compatible formats. The aggregated responses are processed in an analysis unit and made available for data-driven marketing or product decisions.

[0027] Such a real-time interface can be implemented as a RESTful API or a WebSocket connection. The data can be provided in JSON or XML format and integrated via an asynchronous event processing system for further processing in business intelligence tools or data lake architectures. To ensure high scalability, processing can be carried out using a microservices framework that supports load balancing and parallel processing.

[0028] The real-time interface makes it possible to use consumer reactions in real time for data-driven decision-making processes. For example, advertisers can conduct live campaign tests by analyzing the impact of advertising messages on simulated consumers within minutes and making immediate adjustments. Similarly, product developers can test user reactions to prototypes by simulating different designs or product concepts in real time.

[0029] By providing simulated consumer reactions via a real-time interface, companies can make data-driven decisions. This interface not only allows integration into existing market research systems but can also be used for dynamic adjustments to marketing strategies or advertising campaigns. For example, an advertiser could simulate a target group's reaction to a new advertising message within minutes and directly adjust the design or communication strategy.

[0030] The real-time interface can therefore be used in various market research applications. For example, an advertising agency can simulate a target group's reaction to an advertising campaign in real time and make adjustments to design and communication. In product development, the system can capture consumer reactions to packaging designs or product concepts to support informed design decisions.

[0031] By combining these elements, the system provides an automated, depth-psychologically based market research methodology that efficiently combines qualitative and quantitative research elements and enables a realistic simulation of consumer behavior.

[0032] The terms used previously will be explained in more detail below.

[0033] Adaptive, dynamic question generation

[0034] Adaptive dynamic question generation refers to a method for guiding surveys in which questions are not processed linearly according to a fixed scheme, but rather adapt in real time to the answers of the consumer AI twins during the interview. The goal is to enable a deeper exploration of consumer motives, attitudes, and decision-making processes by flexibly responding to unexpected or particularly relevant answers. While conventional questionnaires or rule-based chatbots usually use a predefined sequence of questions, adaptive dynamic question generation is based on a semantic decision tree and machine learning algorithms. These analyze the given answers in real time and control the subsequent questions based on content-related, linguistic, or affective patterns.This allows for targeted follow-up questions, enabling a deeper understanding of consumer reactions.

[0035] The moderator AI twin takes over the adaptive control of the survey. Compared to conventional chatbots or rule-based questionnaires, it enables a significantly more realistic interaction. While classic systems usually use linear or predefined survey structures, the moderator AI twin can react in real time based on the answers, generate alternative questions, and thus explore consumer behavior. This leads to greater depth in the analysis and makes it possible to identify previously unknown correlations or patterns in consumer reactions.

[0036] Adaptive control is achieved through several mechanisms. Semantic analysis identifies key terms or concepts in the responses of the consumer AI twins and derives relevant follow-up questions from them. For example, if a consumer AI twin answers the question about their perception of an advertising campaign with the term "too artificial," the next question could specifically address which aspects are perceived as artificial or what characteristics an authentic advertising campaign should have.

[0037] Additionally, affective analysis can be used to capture emotional reactions. This involves analyzing linguistic indicators such as word choice, sentence structure, and tone of voice to determine whether a response is positive, negative, or ambivalent. For example, if a consumer-Kl twin gives a neutral or evasive answer, the system can ask targeted follow-up questions to elicit a more detailed response.

[0038] Another element of adaptive question generation is the rule-based decision logic, which ensures that all relevant topics are systematically covered during an interview. This logic defines certain core questions that must be asked regardless of the course of the conversation to guarantee comparability of the results, while supplementary questions adapt flexibly.

[0039] By combining these mechanisms, adaptive, dynamic question generation enables an interview process that closely resembles a real human interview in its structure. While classic, linear questionnaires often yield superficial or generic answers, this method allows for a deeper analysis of consumer behavior and perception by specifically uncovering particular motives, contradictions, or decision conflicts. Rule-based decision mechanisms and semantic rules form the basis for structured yet flexible control of the interview process within the AI-supported market research platform. They serve to guide the interaction between the moderator-AI twin and the consumer-AI twins by ensuring that the questions asked are both thematically relevant and context-sensitive.

[0040] Rule-based decision mechanisms define the conditions under which specific questions are asked or certain response patterns of the consumer AI twins are explored in greater depth. These mechanisms operate according to a structured set of rules based on predefined criteria. Examples include decision trees, in which various follow-up questions are available depending on a given answer. Such a decision structure ensures that specific research priorities are covered, while simultaneously allowing the survey to respond flexibly to individual responses.

[0041] An example of a rule-based decision-making mechanism is the control of queries based on predefined keywords or topic categories. If a consumer AI twin answers the question about their perception of a brand logo with the term "modern but impersonal," a rule could stipulate that a more in-depth question about the perceived emotionality of the design is asked. Alternatively, the system could aim to identify which visual elements are perceived as particularly impersonal.

[0042] Semantic rules complement these mechanisms by enabling a content-based analysis of the given answers. While decision trees are based on predefined answer options, semantic rules use linguistic models to interpret open-ended responses. They help to extract the meaning of an answer, establish relationships between concepts, and identify relevant aspects for further questioning.

[0043] A key feature of semantic rules is their ability to analyze responses not only based on individual keywords but also to consider their context. For example, a statement like "The advertisement seems professional, but somehow too polished" can be broken down into its constituent parts using semantic rules. The system recognizes that the adjective "professional" describes a positive attribute, while "too polished" suggests a potential criticism. A rule-based decision logic could then generate a targeted query to determine whether the perceived "polishedness" is due to the design, the language, or the content of the advertisement.

[0044] Furthermore, semantic rules enable the linking of information across different interview phases. If a consumer AI twin indicates at the beginning of a survey that they value authenticity in brand communication, the system can store this information and selectively access it in later phases of the interview. If a new advertising message is presented later, the system can automatically formulate a question that investigates whether the new advertising message is perceived as more authentic than previous versions.

[0045] The AI ​​twins primarily use Latent Semantic Analysis (LSA) to recognize semantic structures in responses by analyzing terms and their semantic spaces. Principal Component Analysis (PCA) is preferred for data reduction and cluster analysis, enabling the efficient processing of high-dimensional response patterns and their integration into statistical decision models.

[0046] By combining rule-based decision-making mechanisms with semantic rules, an adaptive survey strategy is created that is both structured and flexible. While rule-based decision-making mechanisms ensure a certain consistency and comparability of the surveys, semantic rules ensure content-based and individually tailored interviewing. This guarantees that the survey is not only reactive but also exploratory, allowing for deeper insights into consumer behavior.

[0047] Psychologically modeled affect and motivation parameters

[0048] Psychologically modeled affect and motivation parameters are central elements of the consumer AI twins, enabling the simulation of realistic consumer reactions. These parameters form the basis for the AI ​​twins' decision-making and ensure that their reactions are not based solely on superficial language patterns, but rather represent a psychologically grounded simulation of human consumer behavior.

[0049] Affect parameters are used to model emotional responses to various stimuli, such as advertising messages, product designs, or brand communication. They encompass both consciously perceived and unconscious emotional processes that influence consumer behavior. These include fundamental emotional dimensions such as arousal, valence, and dominance, which determine whether a response is positive or negative, whether advertising has an activating or calming effect, or whether a brand is perceived as dominant or submissive. Additionally, specific emotional states such as joy, skepticism, anger, or surprise can be simulated, which vary depending on the context and consumer type.

[0050] Motivational parameters describe the underlying drives and needs that govern purchasing behavior. These are based on psychological models that explain why consumers prefer certain products or why they are particularly sensitive to specific marketing messages. The modeling of these parameters can be based, for example, on need theories such as Maslow's hierarchy of needs or self-determination theory. This approach distinguishes between different levels of motivation, ranging from basic physiological needs and social belonging to self-actualization.

[0051] Additionally, motivational psychology concepts such as hedonistic vs. utilitarian buying behavior can be considered. While some consumers make purchasing decisions based on emotional, pleasure-oriented motives, others prioritize the rational utility of a product. By implementing such parameters, consumer K-linked twins can exhibit differentiated preference patterns and thus simulate realistic consumer profiles.

[0052] An example of the application of psychologically modeled affect and motivation parameters is the evaluation of an advertising campaign by a consumer twin. If an advertisement represents a luxury brand, the consumer twin can generate their reaction based on specific parameters. For example, a consumer twin with a high affinity for prestige and status might evaluate the advertisement positively, while a consumer twin with a stronger focus on authenticity and value for money would be more likely to react skeptically. The combination of affect and motivation parameters thus makes it possible to simulate realistic consumer reactions that go beyond purely linguistic probability models.

[0053] The use of psychologically modeled affect and motivation parameters ensures that the responses generated by the consumer AI twins are not only linguistically plausible but also psychologically consistent. This enables a differentiated analysis of consumer behavior and perception that goes far beyond the possibilities of conventional AI-supported market research.

[0054] Iterative optimization of consumer AI twins using machine learning

[0055] The iterative optimization of consumer AI twins using machine learning describes a continuous learning process in which the consumer AI twin models are incrementally improved based on new empirical data and validated simulation results. The goal of this method is to increase the precision of the simulated consumer reactions, minimize inconsistent or unrealistic response patterns, and ensure the adaptability of the AI ​​twins to different research scenarios.

[0056] The optimization process takes place in several phases, which are linked together in a cyclical learning process. First, empirically obtained consumer interviews, collected, for example, through in-depth morphological interviews or real market research studies, are transformed into vector-based semantic representations. These representations serve as training data for the consumer AI twins by being integrated into the neural network of the large-language model used.

[0057] A fine-tuning process is then performed, in which the weights within the neural network are adjusted based on validated consumer responses. This is done using backpropagation, a fundamental method of supervised learning, in which errors in predicting the AI ​​response are detected and the network layers are optimized so that future simulations more closely reflect real-world consumer reactions. This mechanism minimizes biases in response generation and improves the simulation fidelity of the consumer-AI twins.

[0058] Another key element of iterative optimization is the self-assessment function of the consumer K-K twins. Simulated responses are compared with real consumer interviews, and a difference analysis is performed to identify deviations or patterns in the response structures. This difference analysis is based on statistical similarity metrics that calculate the degree of agreement between simulated and real responses. If significant deviations are detected, the model parameters are automatically adjusted so that future responses are more accurately aligned with realistic consumer reactions.

[0059] In addition to supervised learning, reinforcement learning is also used as a method of iterative optimization. Here, the consumer AI twins receive feedback on their response quality by evaluating simulated consumer reactions according to specific criteria, such as consistency, psychological plausibility, or agreement with existing market research findings. Based on this feedback, the decision logic of the consumer AI twins is dynamically adjusted by reinforcing positive reactions and weakening inappropriate response patterns. A particular advantage of this iterative optimization method is that it allows not only a one-time improvement of the AI ​​models, but also continuous adaptation across different research cycles.This allows the consumer AI twins to be used in various market research scenarios and continuously developed further without requiring manual reprogramming. Furthermore, iterative optimization enables the personalization of the AI ​​twins for specific target groups by aligning them with segmented consumer data over the long term.

[0060] This process ensures that the consumer AI twins generate not only linguistically coherent but also psychologically valid and realistic answers. At the same time, continuous improvement increases the AI ​​twins' adaptability to new research topics and market changes, enabling them to be used as a precise tool for data-driven market research.

[0061] Preferably, this is implemented using a comparison and calibration module, which enables the algorithmic adjustment of the response patterns of the consumer Kl twins by comparison with real consumer interviews.

[0062] Modifying psychological meaning networks

[0063] Modifying psychological meaning networks describes a key mechanism for adapting and further developing consumer-client twins to tailor their responses as realistically as possible to different research questions, consumer groups, and market scenarios. Psychological meaning networks consist of semantic and conceptual structures that model consumer perceptions, decision-making processes, and emotional reactions. Modifying these networks ensures that consumer-client twins exhibit differentiated responses to specific contexts and target groups, rather than generating standardized or generic answers.

[0064] Psychological meaning networks consist of several interrelated levels. The first level comprises basic semantic structures that represent concepts, associations, and categories. This includes, for example, links between a product and specific attributes such as quality, prestige, or sustainability. A second level integrates psychological concepts, such as cognitive biases, motivational structures, or perceptual heuristics, which influence how consumers interpret a brand, product, or advertising message. The third level links these elements with empirical market research data that provides statistical probabilities for specific response patterns.

[0065] Modifying these meaning networks takes place in several steps.

[0066] First, the existing structure is reviewed based on new empirical findings from consumer surveys or market analyses. This involves analyzing whether existing associations remain valid or whether new semantic relationships need to be considered. Subsequently, the semantic and psychological connections within the network are updated by adjusting the weighting of concepts or integrating new terms and associations.

[0067] One example of such a modification is the adaptation of a consumer-Ki twin to changing consumer trends. For instance, if sustainability plays an increasingly important role in a particular market segment, the psychological meaning network can be modified so that terms like "ecological," "fairly produced," or "climate-neutral" are more strongly associated with positive emotions and purchasing decisions. At the same time, it can be analyzed whether terms like "value for money" or "brand loyalty" are losing importance in this segment.

[0068] The modification of psychological meaning networks can be performed both manually through expert intervention and automatically using machine learning methods. Automated adaptation involves analyzing large amounts of text and behavioral data to identify changes in consumer perception. These changes are then incorporated into the semantic connections of the AI ​​twins, allowing their reactions to adapt to changing market dynamics over time.

[0069] This process ensures that the consumer-Kl twins exhibit consistent yet flexible decision-making structures in the long term. This enables them not only to react to current market developments but also to realistically anticipate future changes in consumer perception.

[0070] In a preferred embodiment, the system comprises a study leader AI twin who takes over the control of the interaction processes between the consumer AI twin and the moderator AI twin by:

[0071] Question lists for the moderator-class twin are generated and managed, with the selection of questions based on a rule-based decision logic and / or

[0072] compares the answers of the consumer-Kl twins with reference data sets and calibrates the simulated answers using statistical similarity metrics.

[0073] The method according to the invention preferably comprises a study director AI twin that generates question lists for the moderator AI twin and / or compares and calibrates the answers of the consumer AI twins with reference datasets. Integrating a study director AI twin into the agent system offers several crucial advantages, particularly regarding the optimization of question creation, the consistency of the interviews, and the quality of the generated data. These advantages include the centralized control of the interaction between the consumer AI twin and the moderator AI twin. The study director AI twin acts as a higher-level authority, ensuring that the interaction between the participating AI agents remains methodologically consistent. This prevents the moderator AI twin from asking random or inappropriate questions by ensuring that all relevant topics of the market research study are covered.

[0074] Creating questionnaires using rule-based decision logic ensures that interviews are structured. This allows for both standardized and exploratory questions relevant to the specific research scenario. Automating question creation reduces the manual effort for market researchers and increases the comparability of results across different surveys.

[0075] The study leader's twin interrogator verifies the responses of the consumer twin interrogators against reference datasets. This allows for the calibration of the generated responses, ensuring they remain as realistic as possible. Any deviations or inconsistent response patterns detected can be automatically corrected or further investigated through targeted follow-up questions.

[0076] Integrating a central control module ensures that different surveys are conducted under similar conditions. This improves the comparability of study results and enables reliable analysis of long-term market changes.

[0077] Because the principal investigator-client twin manages questionnaires and analysis procedures, the system can be easily configured for different research contexts.

[0078] For example, a consumer survey on fashion products can have different question priorities than a study on food or financial services, without requiring manual reprogramming.

[0079] Preferably, the study director's AI twin performs the following procedural steps: Adaptive real-time adjustment of the questionnaires; and / or automatic weighting of consumer responses; and / or creation of a research and analysis report; and / or generation of recommendations for market research purposes; and / or implementation of an algorithmic scoring procedure.

[0080] In the preferred embodiment of the system, the study leader AI twin is assigned the following functions: adaptive real-time adjustment of the question lists, based on the analyzed response patterns of the consumer AI twins, to ensure individually optimized questioning; and / or automatic weighting of consumer responses by performing difference analyses between simulated and real responses to minimize model biases; and / or the creation of a structured research and analysis report, which is accessible via a digital interface (e.g.,

[0081] JSON, XML, API); and / or the generation of recommendations for action for market research purposes by deriving probability models for consumer decisions from the aggregated responses of the consumer K-twins; and / or the implementation of an algorithmic scoring procedure that enables further statistically representative analyses based on a large number of synthetic interviews.

[0082] The study director's twin can thus assume a range of additional functions alongside its control tasks, making the system more flexible, efficient, and adaptable. These enhancements ensure dynamic control of market research processes by optimizing the survey strategy in real time, minimizing biases in simulated consumer responses, and enabling deeper data analysis.

[0083] A key advantage lies in the adaptive, real-time adjustment of the questionnaires. Instead of using rigid questionnaires, the researcher-client twin can dynamically adapt the questionnaires based on the responses of the consumer-client twins. This allows the system to specifically address unexpected answers or particularly relevant aspects without a fixed set of questions restricting the dialogue. This flexibility enables a deeper exploration of consumer perceptions and significantly increases the validity of the collected data.

[0084] Another advantage is the automatic weighting of consumer responses, which reduces biases in the model and increases statistical accuracy. Through continuous comparison between simulated and real responses, the researcher-K-twin can identify when certain consumer-K-twins systematically exhibit atypical or unbalanced reactions. In such cases, the system can automatically make adjustments to ensure the results remain representative. This leads to a more realistic simulation of consumer decisions and more precise market research.

[0085] Additionally, the study leader's twin can generate structured research and analysis reports. The collected consumer feedback is automatically converted into standardized formats such as JSON, XML, or API-compatible structures, allowing for direct integration into market research systems or company databases. This reduces the manual effort for market researchers and ensures that the results can be processed efficiently.

[0086] Another significant advantage of this system extension is the generation of actionable recommendations for market research purposes. Based on aggregated consumer reactions, the study director's AI twin can create probability models for consumer decisions. For example, it's possible to predict the likelihood of a new product's success or which advertising messages will elicit the strongest emotional response. This automated data analysis enables companies to optimize their marketing strategies and product development without the need for time-consuming manual evaluation.

[0087] The functionality is rounded out by the implementation of an algorithmic scoring system that allows for the standardized evaluation and prioritization of simulated consumer reactions. By generating statistical assessments of market trends, target group preferences, and advertising effectiveness based on numerous synthetic interviews, companies can optimize their decision-making processes using data. This not only improves the comparability of market research results but also provides deeper insights into the factors influencing purchasing decisions.

[0088] The algorithmic scoring method can be based on a multidimensional evaluation model that analyzes consumer reactions using statistical significance tests, latent semantic analysis (LSA), and regression models. The reactions can be transformed into multidimensional feature vectors and processed using principal component analysis (PCA) to extract the most relevant response patterns. Alternatively, a classifying evaluation algorithm based on a decision tree or a neural network can be used to cluster consumer reactions into different categories (e.g., "positive," "neutral," "negative"). This method can enable the automatic prioritization of advertising-relevant content and identify patterns that correlate with a high probability of purchase.

[0089] In the invention, stochastic probability models are used to probabilistically simulate the decision-making processes and reaction patterns of the consumer AI twins. Instead of deterministic, fixed answers, they enable a dynamic and varied response to surveys, allowing the AI ​​twins to react to stimuli not only more consistently, but also more flexibly and realistically.

[0090] Stochastic probability models simulate consumer reactions by using probabilistic decision structures. They calculate probability values ​​for different possible responses. The Kl twins make their decisions based on a weighted response model that takes into account linguistic semantics, previous responses, and individual consumer profiles.

[0091] A key element of stochastic modeling is the use of probabilistic decision structures, which calculate probability values ​​for different possible answers. The AI ​​twins can thus make a weighted selection from several plausible answer options based on previously defined psychological affect and motivation parameters. Various influencing factors are incorporated, including the semantic meaning of the question posed, previous answers, and the individual characteristics of the simulated consumer profile.

[0092] The probability distribution is designed so that frequently occurring or logically obvious answers receive a higher weighting, while improbable or inconsistent answers are generated with a lower probability. This prevents the AI ​​twins from giving either completely random or overly rigid, deterministic answers. Instead, an adaptive response emerges, balanced between stability and variability. An example of the application of a stochastic probability model is the evaluation of new product packaging. Suppose a consumer AI twin with a high affinity for sustainable products is asked how they feel about plastic packaging.Based on its psychological parameters, the model could generate a rather critical response with a probability of 70 percent, offer a nuanced perspective with a probability of 20 percent, and produce a neutral or even positive reaction with a probability of 10 percent. This weighting ensures that the simulated consumers do not behave in a completely predictable or one-dimensional manner, but rather that their responses reflect a certain range of realistic opinion variations.

[0093] Besides generating individual responses, stochastic models also play a role in the long-term consistency of consumer K-Kinsey twins. Iteratively adjusting the probability distributions based on previous responses ensures that past statements influence future response behavior. This allows for long-term simulations of consumer reactions, which, for example, investigate how brand perception changes across multiple survey waves.

[0094] The use of stochastic probability models ensures that the consumer-Kl twins act consistently and with psychologically sound reasoning in their responses, while also exhibiting natural fluctuations and nuances of opinion typical of real consumer surveys. This method allows for deeper insights into consumer behavior without the Kl twins providing merely mechanical or repetitive answers.

[0095] The algorithmic scoring method can also be used to evaluate the consistency and validity of the simulated responses. The results of this method are incorporated into the iterative optimization model by weighting or correcting responses with low consistency or high deviation from real consumer reactions.

[0096] In summary, these preferred features of the study leader-AI twin enable flexible adaptation of the survey strategy, greater accuracy and relevance of the simulation results, and direct decision support for companies through data-driven forecasts. The combination of these functions makes the system a powerful tool for modern, AI-supported market research that goes far beyond traditional survey methods.

[0097] In another preferred embodiment, the moderator-Kl twin performs the following functions: the dynamic control of the questions; and / or the rule-based control of the interview process; and / or the recognition of semantic patterns in the answers.

[0098] The moderator-client twin plays a central role within the system, dynamically controlling the interview, methodically organizing the conversation, and analyzing the content of the generated responses. Its specific characteristics significantly improve the quality of the simulated consumer interviews by making the interaction between the consumer-client twins and the system more efficient, flexible, and analytically precise.

[0099] A key advantage arises from the dynamic control of the questions. While conventional AI-powered chatbots usually use predefined questions, the moderator AI twin can adapt the questions posed in real time. It analyzes the answers of the consumer AI twins and decides, based on this, whether a more precise follow-up question is needed or whether a new direction in the interview should be taken. This leads to a greater depth of exploration, as the questions can be tailored to individual consumer reactions. This allows, for example, emotional reactions, unclear statements, or unexpected answers to be specifically followed up.

[0100] Another advantage lies in the rule-based control of the interview process. The moderator-KiT twin ensures that all relevant topics are covered during the interview. It uses predefined decision trees, semantic rules, and heuristic models to keep the dialogue methodically structured yet flexible. This ensures that the interview remains both exploratory and reproducible. For example, in market research for a new product, the system can decide whether to focus more on an advertising effectiveness analysis or on price evaluation.

[0101] Additionally, the moderator-K-twin offers the possibility of recognizing semantic patterns in the responses. By employing linguistic analysis methods, the system can identify key themes, implicit meaning structures, or emotional tendencies in the responses of the consumer-K-twins. This enables a deeper analysis of the data by considering not only the direct responses but also their semantic and affective nuances. For example, the system can recognize whether a response contains indirect indications of skepticism toward a product, even if the wording appears neutral.

[0102] These features significantly improve interview management, the quality of data analysis, and the methodological consistency of the simulated surveys. While simple AI-supported market research approaches often use standardized question-and-answer patterns, the moderator-AI twin introduces intelligent, adaptive control into the system, enabling both exploratory and structured interviews. As a result, the insights gained from the simulated consumer interviews become not only more precise but also more meaningful and practically relevant for data-driven decision-making processes in market research.

[0103] The moderator-client twin preferably uses a rule-based control model consisting of semantic decision trees, predefined rule sets, and heuristic models. This model ensures that all critical topics are covered during an interview. Additionally, semantic rules are applied to automatically eliminate irrelevant or redundant questions.

[0104] Further embodiments of the invention are described below. These embodiments represent extensions of the systems and methods described above and can be implemented individually or in any combination.

[0105] In a further embodiment, the described system is implemented as an AI agent system comprising at least one consumer AI twin and one moderator AI twin. The moderator AI twin is configured to provide survey stimuli via a software-controlled communication interface. The survey stimuli are processed by the consumer AI twin, which comprises a Large Language model modified according to the invention and configured to generate consumer responses. The processing of the survey stimuli and the generation of the consumer responses are performed within the AI ​​agent system by a simulation and decision engine.

[0106] The consumer responses generated by the Large Language Model modified according to the invention can subsequently be transformed into structured decision data. Such decision data can include, for example, preference values, rankings, selection decisions, or rating scales. The structured decision data can be aggregated across a large number of consumer AI twins and then analyzed using algorithmic methods. In a further embodiment, the system can be operated in different survey modes. A distinction can be made between a qualitative survey mode and a quantitative survey mode.

[0107] In qualitative interviewing mode, the moderator-client twin generates questions adaptively and contextually. The consumer-client twin then generates open-ended textual responses. These responses can subsequently be processed using semantic analysis methods. For example, semantic similarity metrics can be calculated or vector representations of text elements can be created. Text elements can be transformed into semantic vector representations, which enable the algorithmic calculation of semantic similarities.

[0108] In quantitative survey mode, structured questionnaires with predefined answer options are provided. These options can include, for example, rating scales, Likert scales, multiple-choice answers, or rankings. This allows for the simulation of standardized market research studies. It is possible to efficiently conduct a large number of surveys using this method, e.g., 1000 interviews in one hour. This ensures a statistically representative survey logic.

[0109] In another embodiment, a large number of consumer K-twins are generated, which together form a synthetic consumer population. Each consumer K-twin can have an individual consumer profile. By simulating a large number of such consumer K-twins, a synthetic sample can be generated, which is used to conduct quantitative simulation experiments. The consumer K-twins can be run in parallel, thus enabling scalable simulation of large consumer populations.

[0110] This enables the simulation of a group discussion, the so-called focus group, through which diverse consumer twins, who differ, for example, in their attitudes and behaviors, can be brought into dialogue with one another in a discussion room. This allows for a moderated consensus process, e.g., in the optimization of marketing concepts.

[0111] In another embodiment, stimuli can be provided to multiple consumer AI twins. Such stimuli can include, for example, product concepts, brand messages, price variations, or visual representations. The consumer AI twins react to these stimuli with simulated decisions. The generated decisions can then be converted into structured decision data.

[0112] In another embodiment, generated responses or decisions, as well as emotional and rational classifications, can be transformed into semantic representations. Here, vector representations can be created within a semantic embedding space. The semantic representations can be stored in a high-dimensional embedding space, thus enabling algorithmic analysis of the generated content.

[0113] In another embodiment, the system can provide a reset mode. In reset mode, previously stored interaction data of the consumer-Kl twins are deleted or disregarded, so that subsequent simulations are carried out independently of previous interactions and the conditioning and influencing effects that otherwise occur in conventional interviews are avoided.

[0114] Additionally or alternatively, a memory mode can be provided. In memory mode, interaction data from the consumer AI twins is saved and reused in later simulations. This allows for time-dependent analyses of consumer preferences. The saved interaction data can be used, in particular, to analyze changes in preferences over time. Furthermore, additional survey items can be introduced into the survey process at a later point.

[0115] In another embodiment, visual stimuli, particularly moving images, can be processed. Here, visual representations of people, such as advertising protagonists, testimonials, or brand ambassadors, can be provided to consumer AI twins. Feature representations can be generated from these visual stimuli using algorithmic feature extraction. This feature extraction can be performed using image analysis or computer vision algorithms. Such feature representations can, for example, represent parameters such as facial expression, gaze direction, posture, gestures, or physical appearance. The described embodiments can be used individually or in combination to conduct various simulations of consumer reactions, preference structures, or market reactions.

Claims

Claims 1. Computer-aided method for simulating and analyzing consumer behavior through interacting AI instances, comprising the following steps: 1.1 Initialization of an AI-supported agent system, consisting of: 1.1.1 At least one consumer AI twin, which, as a trained Large Language Model (LLM), generates simulated consumer reactions based on real consumer responses and psychologically sound meaning networks; 1.1.2 A moderator-client twin who asks questions in real time, adaptively and depending on the response patterns of the consumer-client twins, based on a predefined list of questions. 1.2 Conducting a simulated interview process, in which: 1.2.1 The moderator-client twin performs adaptive, dynamic question generation, using decision trees and semantic rules to influence the response behavior of the consumer-client twins; 1.2.2 The consumer-Kl twins generate responses based on psychologically modeled affect and motivation parameters and are generated using stochastic probability models; 1.2.3 A comparison between simulated consumer responses of the consumer-Kl twins and reference datasets is performed to identify deviations in the modeling and to derive correction values ​​for the adaptive behavior of the consumer-Kl twins. 1.3 Iterative optimization of the consumer-Kl twins using machine learning, by: 1.3.1 Transfer of empirically obtained consumer interviews into vector-based semantic representations, which are integrated as embeddings into the LLM of the consumer-Kl twins; 1.3.2 A fine-tuning process in which the neural network weights of the consumer Kl twins are optimized by backpropagation based on validated consumer responses. 1.4 Dynamic adaptation of the simulation model to the specific research context, through: 1.4.1 Selection and modification of psychological meaning networks based on the respective research subject (e.g., consumer behavior in specific product categories, political decision-making behavior, advertising effectiveness tests); 1.4.2 Adaptation of the response behavior of the consumer AI twins through reinforcement learning, based on feedback mechanisms through a real-time evaluation of response quality. 1.5 Provision of simulated consumer reactions via a digital interface, wherein: 1.5.1 The responses from the consumer AI twins will be provided in a structured output format (e.g., JSON, XML, or API-compatible format for market research systems).

2. Method according to claim 1 , characterized by that a study leader-client twin is integrated into the agent system, which takes over the control of the interaction processes between the consumer-client twin and the moderator-client twin by: 2.1 A question list is generated and managed for the moderator-client twin, with the selection of questions based on a rule-based decision logic; 2.2 The list of questions is transmitted to the moderator-Kl-twin and 2.3 Receives the interview output from the moderator-client twin in the form of protocols and automatically evaluates it according to predefined categories and process steps.

3. Method according to claim 2, characterized by that the study director-Kl-twin carries out the following procedural steps: 3.1 Adaptive real-time adjustment of the question lists; and / or 3.2 Automatic weighting of consumer responses to minimize model biases; and / or 3.3 Creation of a structured research and analysis report, provided via a digital interface; and / or 3.4 Generation of recommendations for action for market research purposes; and / or 3.5 An algorithmic scoring method is implemented which, based on a large number of synthetic interviews conducted, statistically enables further representative analyses.

4. Method according to one of the preceding claims, characterized in that a reset mode is provided in which stored interaction data of the consumer-Kl twins are deleted.

5. Method according to one of claims 1 to 3, characterized in that a memory mode is provided in which interaction data is stored and reused in subsequent simulations.

6. Method according to one of the preceding claims, characterized in that a plurality of consumer-Kl twins are produced which together form a synthetic consumer population.

7. Method according to one of the preceding claims, characterized in that a quantitative survey mode is carried out in which structured questionnaires with predefined answer options are provided.

8. Method according to one of claims 1 to 6, characterized in that a qualitative survey mode is carried out in which the moderator-client twin transmits context-dependent generated questions to the consumer-client twin and the consumer-client twin generates answers from answer options.

9. A computer-based system for simulating and analyzing consumer behavior, comprising: 9.1 A network of specialized AI instances, comprising at least: 9.1.1 At least one consumer AI twin that, as a trained Large Language Model (LLM), generates simulated consumer responses based on real consumer answers and psychologically sound meaning networks; 9.1.2 At least one moderator AI twin that simulates the behavior of a human interviewer, conducts adaptive questioning, and dynamically controls the reactions of the consumer AI twins. 9.2 A communication protocol for AI interaction that enables software-controlled data processing between the specialized AI instances, wherein: 9.2.1 The moderator-client twin poses questions to the consumer-client twins based on an adaptive decision tree and semantic rules; 9.2.2 The consumer-client twins generate answers based on psychologically modeled affect and motivation parameters and generated using stochastic probability models. 9.3 A multi-stage training pipeline that performs a stepwise improvement of the consumer AI twins by: 9.3.1 Using exploratory morphological in-depth interviews with real consumers as reference data; 9.3.2 Psychological meaning networks and semantic annotation models are integrated; 9.3.3 An iterative fine-tuning of the consumer-Kl twins is carried out based on real consumer reactions. 9.4 A comparison and calibration module that compares the simulated responses of the consumer Kl twins with real consumer interviews and performs an algorithmic adjustment of the response patterns, wherein: 9.4.1 Statistical similarity metrics can be used to evaluate the simulation fidelity; 9.4.2 A dynamic correction mechanism is integrated that adjusts the weighting of affective and cognitive parameters. 9.5 A real-time interface to market research that: 9.5.1 Interaction with the system is made possible by testing marketing or product decisions based on simulated consumer reactions; 9.5.2 Automated recommendations for action are generated by calculating probability values ​​for certain consumer decisions.

10. System according to claim 9, characterized by that it includes a study leader-client twin who takes over the control of the interaction processes between the consumer-client twin and the moderator-client twin by: 10.1 A question list is generated and managed for the moderator-class twin, with the selection of questions based on a rule-based decision logic; 10.2 The questionnaire is transmitted to the moderator-client twin and.3 The interview output is received from the moderator-client twin in the form of a protocol and automatically evaluated according to predefined categories and process steps. System according to claim 10, characterized by that the study director's twin brother performs the following functions: .1 Adaptive real-time adjustment of the question lists, based on the analyzed response patterns of the consumer AI twins, to ensure individually optimized questioning; and / or .2 Automatic weighting of consumer responses by performing difference analyses between simulated and real responses to minimize model biases; and / or .3 Creation of a structured research and analysis report, provided via a digital interface (e.g., JSON, XML, API); and / or .4 Generation of recommendations for action for market research purposes by deriving probability models for consumer decisions from the aggregated responses of the consumer K1 twins; and / or .5 Implementation of an algorithmic scoring procedure that enables further statistically representative analyses based on a large number of synthetic interviews conducted. System according to one of claims 9 to 11 , characterized by that the moderator-Kl-twin assumes the following functions: 12.1 Dynamic control of the questions by reformulating or adapting the questions in real time based on the answers of the consumer AI twins to gain deeper insights; and / or 12.2 Rule-based control of the interview process by using predefined decision trees, semantic rules, or heuristic models to ensure that all relevant topics are covered during the interview; and / or 12.3 Recognition of semantic patterns in the responses by using the language processing of the moderator-client twin to analyze content-related relationships, key themes, or implicit meaning structures in the statements of the consumer-client twins. System according to one of claims 9 to 12. characterized by that the modules data collection, data analysis and / or reporting are executed as an automated workflow (workstream).