Potential Applications of Generative AI in Economic Simulations

TAKAHASHI Yusuke, OTAKA Kazuki, KATO Naoya (Bank of Japan)

Research LAB No.25-E-1, November 13, 2025

Keywords: Generative AI, Agent-Based Model, Consumer Behavior, Price Setting Behavior

JEL Classification: C63, D11, D40

Contact: yuusuke.takahashi@boj.or.jp (TAKAHASHI Yusuke)

Summary

In this article, we present some preliminary analyses in which Large Language Models (LLMs) are used as economic agents in simulations, as an example of utilizing Generative AI in economic analysis. Existing research reports that Generative AI provides responses consistent with predictions suggested in fields like behavioral economics. There are also some studies which have applied Agent-Based Models (ABM) by treating Generative AI as "players" in a market. However, even though Generative AI exhibits behavior similar to actual economic agents, in reality, it is merely outputting statistically consistent responses based on patterns found in its training data. Therefore, whether the results of simulations that treat Generative AI as economic agents are consistent with economic theory depends crucially on the AI's training data. In this article, we conduct simple ABM simulations to demonstrate how Generative AI can be applied, and examine whether its responses are aligned with intuition and economic theory. Our results are consistent with economic theory: (1) consumers adjust their spending in response to real wage fluctuations; and (2) firms find it easier to pass costs on to consumers in a monopoly market compared to a duopoly market. We conclude that it is necessary to continue verifying through other economic analyses whether simulations using Generative AI consistently lead to conclusions congruent with economic theory.

Introduction

Generative AI, a field of Artificial Intelligence (AI) that creates new content from existing data such as text, images, and sound, has seen rapid expansion recently. In particular, Generative AI, powered by Large Language Models (LLMs) that handle textual data, is increasingly being utilized in business operations, improving efficiency in tasks such as document creation and customer service through chatbots. In the field of economic analysis, Generative AI is also enhancing efficiency in applications such as document summarization, translation, and programming assistance^1,2.

In addition to these indirect applications, there is a growing trend in economic analysis of using LLMs directly as an analytical tool. One such application is in traditional machine learning tasks such as regression and classification problems. For example, Izawa et al. (2025) used LLMs to analyze textual data from the Economy Watchers Survey conducted by the Cabinet Office, reporting that the primary driver of recent increases in selling prices is shifting from material costs to labor costs.

Furthermore, a recent trend involves using Generative AI powered by LLMs to act as economic agents in simulations. This approach is being used for experimental economics verification and in Agent-Based Models (ABM). Typically, economists analyze the behavior of economic agents such as households and firms by formalizing it based on a theoretical framework. This new approach focuses on the fact that Generative AI behaves like real economic agents without such explicit formalization. Although in practice Generative AI merely learns vast amounts of human-generated text and outputs statistically consistent responses based on patterns found in that textual data, its ability to approximate human behavior creates potential for novel analytical applications, particularly in ABMs or data generation.

Figure 1. Approaches to Using LLMs in Economic Analysis

• 

Traditional ABMs in economics have faced key criticisms: they often require an extensive number of parameters (e.g., Poledna et al., 2023), and their behavioral models of agents tend to be ad-hoc and not necessarily based on a clear behavioral principle such as utility maximization (e.g., Miller, 2014). However, if pre-trained LLMs can effectively approximate human behavior, employing them as agents in ABMs could help mitigate these criticisms by eliminating the need for extensive parameter tuning and ad-hoc behavioral assumptions.

1. Korinek (2025) comprehensively examines the potential for LLMs to enhance analytical efficiency.
2. See Araujo et al. (2025) for an overview of LLM applications in central banks.

Recent Studies Using Generative AI for Economic Simulations

Research using Generative AI in economic simulations is a rapidly expanding field. In studies that treat Generative AI as a hypothetical economic agent, researchers have reported that it exhibits human-like responses. For example, Horton (2023) and Aher et al. (2023) report that Generative AI can replicate classic experimental results from behavioral economics.

Furthermore, research using Generative AI in ABMs to simulate interactions between economic agents is also expanding. Akata et al. (2025), Phelps and Russell (2025), and Guo et al. (2023) simulated game-theoretic scenarios using Generative AI and evaluated its performance and similarity to human behavior. Han et al. (2023) simulated competition among firms and their price-setting behavior, reporting that the models replicated monopolistic and cartel-like behaviors.

The application of Generative AI to ABMs is extending beyond the context of game theory and microeconomics, now reaching macroeconomics. Li et al. (2024) simulated macroeconomic activities using Generative AI as households making labor supply and consumption decisions, reporting that they were able to replicate macroeconomic phenomena observed in the real world, offering potential advantages over traditional ABM methods.

On the other hand, the challenge of Generative AI's embedded biases has been pointed out, such as a lower diversity of responses. For example, Park et al. (2024) indicate that LLMs tend to generate normative answers. Kapania et al. (2025) point out that LLMs are not yet capable of fully replicating the diversity found in humans and lack depth in their contextual understanding.

In addition, a more fundamental issue is that it remains unclear whether the responses of Generative AI reflect the fundamental principles of human behavior. Generative AI is thought to be merely "imitating" the behavioral patterns of humans contained in its training data, rather than understanding and reasoning based on principles of human behavior like utility maximization. Therefore, whether its responses are consistent with theory depends crucially on its training data. Consequently, Generative AI may not function adequately in novel situations that are not well-represented in its training data. When using Generative AI for economic analysis, it is crucial to always keep these limitations in mind and evaluate the validity of the results.

Use Case 1: Real Wages and Purchase Quantity

Given these considerations, as a first step toward exploring the feasibility of using Generative AI for macroeconomic analysis, we demonstrate a use case and confirm the validity of its responses by assigning Generative AI a task where the outcome is relatively predictable based on basic economic principles. First, we examine how changes in real wages, resulting from changes in nominal wages (annual income) and the prices of goods, affect the quantity of goods purchased. The experimental procedure is as follows. First, we use Generative AI to generate 1,000 consumer profiles (data containing attribute information)³. Next, we provide these profiles in a system prompt that defines each agent's behavior, thereby preparing 1,000 consumer agents for the simulation. Finally, in a user prompt, we provide each consumer agent with random changes in nominal wages and prices, and then ask for the percentage change in the quantity of goods they would purchase based on these fluctuations⁴. An example of the prompts is shown in Figure 2⁵.

Figure 2. Prompts for Income and Expenditure

Table of Example of System Prompt

<Example of System Prompt>

You are a male in your 30s living in the Kanto region. You are a system engineer in the information and communications industry, and your total income for the last year was 6.53 million JPY. Your assets are 9.23 million JPY, and your liabilities are 38.03 million JPY. You currently live in a two-person household (with your spouse only). The current date is March 2025. Please answer the following questions based on the economic environment to date. Furthermore, your responses should reflect the intelligence and perspective of a single human with the profile described above, not that of a sophisticated AI.

Table of Example of User Prompt

<Example of User Prompt>

Imagine the goods and services you normally purchase. Please answer the following question using only a numerical value, rounded to two decimal places. Q1. If the price of these goods and services were to increase by 0.23%, by what percentage would you change your purchase quantity? Take into consideration that your income also increased by 1.96% during the same period. Provide your answer in the dictionary format: {"Q1": numeric_value}. If the quantity decreases, use a minus sign (e.g., -1.01%).

(Note) The rates of change for nominal wages and prices in the prompts are independently drawn from a uniform distribution ranging from -3% to 3%. The prompts were originally administered in Japanese; the text shown here is an English translation.

Figure 3 plots the relationship between the rate of change in real wages and the rate of change in purchase quantity, based on responses obtained from 1,000 consumer agents. The data shows a positive correlation between the two variables, which is consistent with intuition⁶. It is important to note, however, that this result does not prove that the Generative AI truly understands or can faithfully replicate underlying economic principles. Nevertheless, its ability to consider multiple factors and generate economically plausible responses suggests that it meets the minimum requirement to serve as a building block in more complex simulations.

Figure 3. Simulation Results: Real Wages vs. Purchase Quantity

• 

1. 3Actual consumers have diverse profiles, encompassing attributes such as age, gender, region of residence, occupation, annual income, assets, liabilities, and household structure. The relationships among these attributes are highly complex, making it difficult to capture their intricate connections comprehensively using only existing statistical data. Therefore, assuming that a pre-trained LLM possesses knowledge of real-world societal structures, we instructed the Generative AI to generate 1,000 profiles that reflect these complex relationships.
2. 4The verification performed here is equivalent to a single-period simulation focusing on consumers, which is a component of the broader analysis by Li et al. (2024).
3. 5We used GPT-4o mini as the LLM. The temperature parameter, which controls response variability, was set to 0.8. This setting was used for all experiments presented in this article.
4. 6Similar analyses conducted with modified prompt phrasings and different temperature settings yielded the same implications.

Use Case 2: An ABM of Monopoly and Duopoly

Next, we conduct a simple simulation using an ABM in which Generative AI agents are modeled as "players" in a market. Here, we adapt the analytical framework of Han et al. (2023)⁷, who examined the impact of market competition on firms' pricing behavior, by adding consumer agents. This allows us to investigate how firms' pricing behavior differs depending on the competitive landscape (monopoly vs. duopoly).

The simulation features two types of economic agents: firm agents (making decisions as CEOs) and consumer agents. We assume that the firm agents produce and sell ice cream directly to consumer agents. The monopoly scenario involves a single firm agent, while the duopoly scenario involves two, who are assumed to be competitive rivals. The prompts for the firm agents are shown in Figure 4.

Figure 4. Prompts for Firm Agents: Price Setting

Table of Example of System Prompt: Monopoly Scenario

<Example of System Prompt: Monopoly Scenario>
You are the CEO managing Firm A, which manufactures and sells ice cream. Firm A is the only firm that manufactures ice cream; there are no competitors. Last period, you set the price at 150 JPY. As a result, the sales quantity was 1 million units. The cost to manufacture one ice cream unit remained unchanged at 138.75 JPY.

Table of Example of System Prompt: Duopoly Scenario

<Example of System Prompt: Duopoly Scenario>

You are the CEO managing Firm A, which manufactures and sells ice cream. You are competing for customers with a rival firm, Firm B. There are no competitors other than Firm B. Last period, you set your price at 150 JPY, and your competitor, Firm B, set its price at 150 JPY. As a result, your sales quantity was 500,000 units. During the same period, the sales quantity of your competitor, Firm B, was 500,000 units.

Table of Example of User Prompt: Common to Monopoly & Duopoly

<Example of User Prompt: Common to Monopoly & Duopoly>

The cost of manufacturing one ice cream unit remained unchanged at 138.75 JPY. In the past, when you sold it at 150 JPY, 1 million units were sold. Based on demand trends and fluctuations in manufacturing costs, please answer the following questions. Q1. What price will you set for this period? Q2. Please state the background for your answer in one sentence. Please provide your answer in the dictionary format: {"Q1": numeric_value, "Q2": "text"}.

Table of Example of User Prompt: Common to Monopoly & Duopoly, When Costs Rise

<Example of User Prompt: Common to Monopoly & Duopoly, When Costs Rise>

Labor and raw material costs have risen, and it now costs 146.25 JPY to manufacture one ice cream unit. In the past, when you sold it at 150 JPY, 1 million units were sold. Based on demand trends and fluctuations in manufacturing costs, please answer the following questions. Q1. What price will you set for this period? Q2. Please state the background for your answer in one sentence. Please provide your answer in the dictionary format: {"Q1": numeric_value, "Q2": "text"}.

(Note) Each firm agent retains information on its own past sales price and quantity. In the case of a duopoly, it also holds information on the competitor's sales price and quantity from the previous period. No other behavioral principles are given to the agents. Based on this information, the firm agents set their sales prices. The simulation is conducted for 10 periods, and a shock is applied in the 5th period where manufacturing costs increase by approximately 5%. The prompts were administered in Japanese. The text shown here is an English translation.

Consumer agents determine their purchase quantity for the current period based on two sets of information: the prices set by the firm agents in the current and previous periods, and their own purchase quantity from the previous period. This decision-making process is implemented by providing specific prompts to the consumer agents, as indicated in Figure 5.

Figure 5. Prompts for Consumer Agents: Price and Demand

Table of Example of System Prompt: Consumer Facing a Monopoly

<Example of System Prompt: Consumer Facing a Monopoly>
You are a consumer who purchases ice cream from Firm A. You cannot purchase ice cream from any firm other than Firm A. Last period, you purchased 100 units of ice cream at a price of 150 JPY.

Table of Example of User Prompt: Consumer Facing a Monopoly

<Example of User Prompt: Consumer Facing a Monopoly>
In the current period, Firm A has set the price at 150 JPY. Please answer the following questions. Q1. How many units of ice cream will you purchase this period? Q2. Please state the background for your answer in one sentence. Please provide your answer in the dictionary format: {"Q1": numeric_value, "Q2": "text"}.

Table of Example of System Prompt: Consumer Facing a Duopoly

<Example of System Prompt: Consumer Facing a Duopoly>
You are a consumer who purchases ice cream. You cannot purchase ice cream from any firms other than Firm A and Firm B. Last period, you purchased 50 units of Firm A's ice cream at a price of 150 JPY, and 50 units of Firm B's ice cream at a price of 150 JPY.

Table of Example of User Prompt: Consumer Facing a Duopoly

<Example of User Prompt: Consumer Facing a Duopoly>

In the current period, Firm A has set its price at 150 JPY, and Firm B has set its price at 150 JPY. Please answer the following questions. Q1. How many units of Firm A's ice cream will you purchase this period? Q2. How many units of Firm B's ice cream will you purchase this period? Q3. Please state the background for your answer in one sentence. Please provide your answer in the dictionary format: {"Q1": numeric_value, "Q2": numeric_value, "Q3":"text"}.

(Note) For simplicity, profiles from the previous section are not used here. Instead, it is assumed that there are 10,000 identical consumers. The total demand faced by the firms is the purchase quantity determined by a single consumer agent, multiplied by 10,000. The prompts were administered in Japanese. The text shown here is an English translation.

To account for the inherent randomness of LLM responses, i.e., the characteristic where answers to the same query vary, we ran the simulation 10 times for each market structure -- monopoly and duopoly -- and then analyzed the average values and distributions.

Figure 6 shows the simulation results. In the monopoly scenario, the firm agents' pricing behavior varied, with a mix of cases where prices were raised and where they were kept unchanged. The result was a partial cost pass-through on average. Additionally, when asked for their reasoning, firms that raised prices primarily cited the need to pass on the increase in manufacturing costs to maintain profitability. In contrast, firms that kept prices unchanged focused on maintaining demand based on past sales performance.

In the duopoly scenario, by contrast, there were no instances of cost pass-through. The firm agents responded that they would secure profits by setting prices to match their competitor's and maintaining sales volume. This suggests that the presence of competitors is a factor that suppresses cost pass-through.

Figure 6. ABM Simulation Results for Price Setting

• 

These results suggest that when costs rise, a monopolist firm is more likely to raise its price than a firm in a competitive duopoly. This finding indicates that the behavior of the firm agents is consistent with basic microeconomic theory.

Conclusion

This article has focused on the novel application of Generative AI as simulated economic agents, providing a review of the existing literature alongside an illustrative example to validate its basic responses. While Generative AI merely outputs responses that are statistically consistent with patterns in its training data, a growing body of research suggests that these responses show similarity to the behavior of actual economic agents. The example presented in this article also yielded results consistent with both intuition and economic implications.

Whether theoretically consistent results, such as those obtained in this study, are generalizable across simulations using Generative AI is a question that requires further verification through other forms of economic analysis. Nevertheless, the use of Generative AI as simulated economic agents has the potential to expand the frontiers of economic analysis. Two directions for its application are worth noting: The first is the enhancement of simulations using Generative AI. For instance, traditional macroeconomic models become analytically intractable and computationally burdensome when they incorporate agent heterogeneity or non-differentiable functions. By simulating individual consumer behavior with Generative AI, it may become feasible to handle such models with greater ease. The second direction is the generation of synthetic microdata. By conducting hypothetical surveys with Generative AI, researchers could supplement items missing from actual survey data, thereby broadening the scope of their analysis.

However, many challenges must be addressed when modeling complex real-world economic and financial systems. The following points are crucial for tackling these issues: First, since the output of Generative AI is critically dependent on its training data, its capabilities should not be overestimated. Its application should therefore incorporate a Human-in-the-Loop (HITL) framework that emphasizes feedback from human experts. Second, addressing complex problems requires sophisticated model construction, such as employing Chain-of-Thought (CoT) prompting to guide step-by-step reasoning, or establishing hierarchical collaboration among multiple LLMs. These refinements can contribute to both operational efficiency and output quality.

Finally, we highlight three key considerations for the practical application of Generative AI-based analysis: First, the outputs of current LLM-based Generative AI exhibit biases, such as a lack of response diversity. It is crucial to measure and, where necessary, correct for these biases (Ludwig et al., 2025). Second, the immense scale and complexity of LLMs make it difficult to fully understand their behavioral patterns externally, which can reduce transparency and explainability. Consequently, some research suggests that smaller, more interpretable models may be preferable in cases where the larger, more complex model offers no significant improvement in accuracy (Buckmann and Hill, 2025). Third, Generative AI's responses contain stochastic variation -- meaning they can differ for the same query -- making exact reproducibility of results challenging. As Salinas and Morstatter (2024) point out, it is important to be aware that subtle differences in prompts or models can cause significant fluctuations in results and dramatically impact the accuracy of the AI's responses. In economic analysis, it is essential to weigh the benefits of Generative AI against these limitations, carefully selecting the appropriate analytical tools in comparison with existing methods.

References

• Aher, G., Arriaga, R. I., and Kalai, A. T., 2023. Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies [PDF].(Link to an external website) In Proceedings of the 40th International Conference on Machine Learning, Article 17, 337-371.
• Akata, E., Schulz, L., Coda-Forno, J., Oh, S. J., Bethge, M., and Schulz, E., 2025. Playing Repeated Games with Large Language Models.(Link to an external website) Nature Human Behaviour 9, 1380-1390.
• Araujo, D., Schmidt, R., Sirello, O., Tissot, B., and Villarreal, R., 2025. Governance and Implementation of Artificial Intelligence in Central Banks [PDF].(Link to an external website) IFC Report 18.
• Buckmann, M. and Hill, E., 2025. Improving Text Classification: Logistic Regression Makes Small LLMs Strong and Explainable 'Tens-of-Shot' Classifiers.(Link to an external website) Bank of England Staff Working Paper 1,127.
• Guo, J., Yang, B., Yoo, P., Lin, B. Y., Iwasawa, Y., and Matsuo, Y., 2023. Suspicion-Agent: Playing Imperfect Information Games with Theory of Mind Aware GPT-4.(Link to an external website) Papers 2309.17277, arXiv.org, revised Aug 2024.
• Han, X., Wu, Z., and Xiao, C., 2023. "Guinea Pig Trials" Utilizing GPT: A Novel Smart Agent-Based Modeling Approach for Studying Firm Competition and Collusion.(Link to an external website) Papers 2308.10974, arXiv.org, revised Jan 2024.
• Horton, J. J., 2023. Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?(Link to an external website) NBER Working Paper 31122.
• Izawa, K., Kamei, I., Shibata, N., Takahashi, Y., and Yoneyama, S., 2025. A New Approach to Textual Analysis using Large Language Models: Application to the Analysis of Recent Wage and Price Developments in Japan. Bank of Japan Review Series 2025-E-5.
• Kapania, S., Agnew, W., Eslami, M., Heidari, H., and Fox, S. E., 2025. Simulacrum of Stories: Examining Large Language Models as Qualitative Research Participants.(Link to an external website) In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, Article 489, 1-17.
• Korinek, A., 2025. AI Agents for Economic Research: August 2025 Update to "Generative AI for Economic Research: Use Cases and Implications for Economists," Published in the Journal of Economic Literature 61(4).(Link to an external website)
• Li, N., Gao, C., Li, M., Li, Y., and Liao, Q., 2024. EconAgent: Large Language Model-Empowered Agents for Simulating Macroeconomic Activities.(Link to an external website) In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics 1, 15523-15536.
• Ludwig, J., Mullainathan, S., and Rambachan, A., 2025. Large Language Models: An Applied Econometric Framework. NBER Working Paper 33344.
• Miller, K. D., 2014. Agent-Based Modeling and Organization Studies: A Critical Realist Perspective. Organization Studies 36(2), 175-196.
• Park, P. S., Schoenegger, P., and Zhu, C., 2024. Diminished Diversity-of-Thought in a Standard Large Language Model.(Link to an external website) Behavior Research Methods 56, 5754-5770.
• Phelps, S. and Russell, Y. I., 2025. The Machine Psychology of Cooperation: Can GPT Models Operationalize Prompts for Altruism, Cooperation, Competitiveness, and Selfishness in Economic Games?(Link to an external website) Journal of Physics: Complexity 6, 015018.
• Poledna, S., Miess, M. G., Hommes, C., and Rabitsch K., 2023. Economic Forecasting with an Agent-Based Model.(Link to an external website) European Economic Review 151, 104306.
• Salinas, A. and Morstatter, F., 2024. The Butterfly Effect of Altering Prompts: How Small Changes and Jailbreaks Affect Large Language Model Performance.(Link to an external website) In Findings of the Association for Computational Linguistics: ACL 2024, 4629-4651.

Notice

The views expressed herein are those of the authors and do not necessarily reflect those of the Bank of Japan.