Teaching Agile Requirements Engineering: A Stakeholder Simulation with Generative AI

1. Problem Statement

The paper addresses two primary challenges in modern software engineering education and practice:

• Persistent Stakeholder Involvement: In real-world Agile Requirements Engineering (RE), the active involvement of users and customers is a significant bottleneck. Students often lack the opportunity to practice eliciting requirements from diverse, realistic stakeholders during their studies.
• Misuse of Generative AI (GenAI): The rapid adoption of GenAI in higher education has outpaced the development of AI literacy and ethical guidelines. Students and practitioners frequently misuse these tools, ignoring limitations such as data bias, hallucinations, and the lack of contextual understanding, leading to superficial or ethically compromised outcomes.

The authors argue that while AI Personas (synthetic users) offer a solution for simulating stakeholders, their application in actual system development is limited by inherent biases in Large Language Model (LLM) training data. Therefore, they propose using AI Personas primarily as an educational and reflective tool to teach both Agile RE practices and the critical evaluation of AI tools.

2. Methodology

The authors developed and implemented a Stakeholder Simulation Teaching Case integrated into Agile Requirements Engineering courses (specifically at the University of Applied Sciences and Arts Hannover). The methodology follows a three-step pedagogical loop:

A. Requirements Discovery (The Simulation)

• Setup: Students act as Requirements Engineers. They are provided with a Meta-Prompt designed to configure a GenAI chatbot (e.g., ChatGPT) to simulate specific AI Personas.
• Scenario: The case study involves developing a mobile public administration app.
• Personas: Three distinct AI Personas are simulated:
  1. A public administration specialist (bureaucratic, risk-averse).
  2. An accessibility-focused citizen (needs inclusive design).
  3. A power user from a large city (efficiency-driven, tech-savvy).
• Interaction: Students conduct interviews with these personas to elicit requirements, navigate resistance, and handle conflicting expectations.

B. Requirements Specification (Agile Artifacts)

• Students translate interview data into standard Agile artifacts.
• Artifacts Used: User Stories, Story Maps, and Impact Maps.
• Constraint: The authors emphasize that while GenAI can generate text, the structuring and refinement of these artifacts (especially visual maps) must be done collaboratively by students to ensure deep understanding.

C. Reflection (Critical Analysis)

• A structured group discussion follows the simulation.
• Focus Areas:
  • Quality of Artifacts: Did the generated user stories have measurable acceptance criteria?
  • AI Limitations: Discussion of LLM biases (e.g., gender representation, neurodivergence), hallucinations, and the loss of context.
  • Ethics & Regulation: Data privacy (GDPR) and the EU AI Act.
  • Human vs. AI Communication: Analyzing how AI prioritizes efficiency over human-like nuance.

3. Key Contributions

• Pedagogical Framework for AI-Integrated RE: A structured approach to teaching Agile RE that uses GenAI not just as a productivity tool, but as a subject of critical inquiry.
• The Meta-Prompt Strategy: The authors propose using a Meta-Prompt rather than a fixed, provider-specific bot (like a custom GPT). This ensures:
  • Provider Independence: The simulation works across different LLM providers.
  • Flexibility: The prompt can be updated as LLM capabilities evolve.
  • Data Privacy: Students can choose local or alternative models if they have data protection concerns.
• Identification of "AI Traps" in Education: The paper highlights specific pitfalls where students accept AI output uncritically, such as:
  • Generating user stories that sound plausible but lack "Definition of Ready" criteria.
  • Creating Story Maps that are merely feature lists rather than user narratives.
  • Confusing "Stakeholders" (in the interview sense) with "Actors" (in Impact Mapping).

4. Results and Observations

The approach was tested over several terms (starting Winter 2024/25) with ~95 students. Key findings include:

• Technical Implementation Evolution:
  • Initial Attempt: Using a custom GPT hosted on OpenAI failed due to licensing limits (prompt caps) and data privacy concerns.
  • Current Solution: The Meta-Prompt approach succeeded, allowing students to use their preferred tools.
  • Prompt Optimization: An attempt to optimize the prompt for efficiency (using OpenAI's Prompt Optimizer) backfired, making the AI Personas too robotic and efficient, reducing the realism of the interview. The authors are now experimenting with prompts that enforce distinct personalities and interactive behaviors.
• Student Learning Outcomes:
  • Students successfully practiced stakeholder communication and requirement translation.
  • Critical Gap: Without explicit guidance, students often accepted GenAI-generated artifacts as final, failing to recognize the need for refinement (e.g., adding acceptance criteria).
  • Reflection Value: The structured discussion successfully highlighted the ethical and technical limitations of GenAI, turning the simulation into a lesson on AI literacy.

5. Significance

This paper is significant for the following reasons:

• Bridging Theory and Practice: It provides a scalable, low-cost method to simulate complex stakeholder interactions that are otherwise difficult to organize in a classroom setting.
• Redefining AI Literacy: It moves beyond "how to use AI" to "how to critically evaluate AI." It teaches students that AI Personas are synthetic constructs with inherent biases, making them unsuitable for final production requirements without human validation.
• Adaptability: The Meta-Prompt architecture allows educators to remain agnostic to specific LLM vendors, ensuring the teaching material remains relevant as the technology landscape shifts.
• Future Direction: The authors suggest that future work will explore Agentic AI (autonomous agents) and voice-based interfaces to further enhance the realism of these simulations, while continuing to evaluate learning outcomes systematically.

In conclusion, the paper demonstrates that GenAI, when used as a simulated stakeholder within a structured, reflective pedagogical framework, is a powerful tool for training the next generation of Requirements Engineers to be both technically proficient and ethically aware.