From Code to Culture: Aligning Large Language Models With Corporate Values

MCML Authors

Stefan Feuerriegel

Prof. Dr.

Principal Investigator

Artificial Intelligence in Management

Abstract

Organizations increasingly deploy large language models (LLMs) to support daily work, yet little is known about how these systems can embody and transmit corporate values. Hence, a central question for managers is how LLMs can be designed to internalize organizational values and how such systems subsequently shape employees' attitudes and behavior. We examine this question through four empirical studies conducted with a large German original equipment manufacturer (OEM). Studies 1A and 1B examine how LLMs such as OpenAI's GPT can be engineered to reflect corporate values. Specifically, in Study 1A, we compare common algorithmic approaches to incorporate organizational values into LLMs (e.g., by adapting the system prompt or fine-tuning LLMs on company knowledge). We find that a carefully crafted system prompt offers a cost-effective way to align LLM responses with corporate values, while, surprisingly, fine-tuning is less effective. Study 1B confirms these findings in day-to-day tasks (e.g., email writing). Studies 2A and 2B then assess how interacting with these systems influences employees. Study 2A finds that exposure to a value-aligned LLM does not alter employees' personal values. However, interestingly, Study 2B shows that employees interacting with such systems subsequently show behavior that is more closely aligned with corporate values. For organizational theory, this suggests that digital tools can offer a lightweight way to steer employee behavior in line with organizational behavior, even when the underlying attitudes of employees are unchanged. Together, our studies contribute to organizational science by demonstrating both how LLMs can be designed to internalize corporate values and how such LLMs can shape employees' behavior. This has important managerial implications for organizations seeking to embed corporate values into digital tools.

misc

---

Preprint

Oct. 2025

---

Authors

M. Kosmas • A.-S. Mayer • S. Feuerriegel • F. Bodendorf

---

Links

DOI

---

Research Area

 A1 | Statistical Foundations & Explainability