ߣߣ��Ƶ

For many students, the convenience of using AI is difficult to ignore. Why work through a complex problem when a solution is available within seconds?

Rather than treating AI solely as a productivity tool, we explored how customised AI agents could be designed to support active learning, reflective thinking and authentic engagement. 

Guided by the principle of “pedagogy first, AI second”, we developed a series of pedagogically designed AI agents for undergraduate students. Our intention was that these agents would encourage students’ participation, reasoning and enquiry, rather than passively feeding them the answers.

Our approach was guided by three principles:

1. Build with purpose: AI agents should align with clear learning outcomes rather than technological novelty. 
2. Design for thinking: Interactions should encourage reasoning, reflection and active participation. 
3. Keep it human: Educators remain central in guiding, scaffolding and contextualising learning experiences. 

These principles informed the design of our customised AI agents for role-play, game-based and retrieval-practice activities.

How to design pedagogically driven AI agents

To support different forms of active learning, we categorised our agents into three distinct pedagogical categories:

1. Game-based learning agents

The first category of agents used game-based learning to increase engagement. For example, they generated escape-room environments in which students needed to resolve scenario-based dilemmas to unlock clues and progress through the game. 

• Your students use GenAI. Let’s redesign learning so thinking still happens
• Banning AI won’t make students care about learning
• Campus Talks: What does ‘AI across the curriculum’ look like in practice?

These scenarios often involved professional and ethical conflicts, requiring students to apply theoretical principles and justify their decisions. 

Another agent generated risk-assessment simulations. Students were required to identify risks and evaluate their significance, before receiving iterative performance feedback. By shifting from passive reading to active application, these activities encouraged learning through challenge and exploration.

2. Role-play and simulation agents

The second category immersed students in realistic workplace scenarios, where they interacted with AI-generated characters to gather information and make professional judgements. One agent simulated interactions with a managing director, requiring students to obtain information for risk assessment by asking relevant, purposeful questions. 

Another agent simulated investigative interviews. Students questioned both cooperative and more resistant interviewees, before receiving structured feedback on their questioning techniques and professional judgement. 

These role-play agents supported experiential learning by allowing students to practise communication, enquiry and professional judgement in a safe environment, before they enter the workforce.

3. Retrieval practice agents

The third category focused on retrieval practice through scenario-based questioning and multiple-choice activities. Rather than simply providing answers, these agents supported repeated practice and explanation-based learning. 

For example, one agent generated realistic scenarios, multiple-choice questions and explanations to reinforce conceptual understanding, while another produced customised questions on specific concepts to support revision and retrieval practice. 

This approach positioned revision as an active process of building conceptual understanding rather than passive review.

Implementation and findings

To understand whether pedagogically designed AI agents influenced student engagement, we worked with two cohorts of undergraduate accountancy students.

In the first cohort, students had access to an off-the-shelf chatbot that was not integrated into the broader learning design. More than half of the students (52 per cent) chose not to use the chatbot and many perceived it as no more useful than the existing AI tools already available to them.

The second cohort interacted with a series of pedagogically designed AI agents, embedded directly within course activities and learning tasks. Non-engagement dropped significantly to 15 per cent, while 73 per cent of students reported that the agents changed the way they learned.

Students described the agents as follows:

“The agent acts like a second professor; it helps explain uncertainties I have.”

“Using it improved my ability to ask questions and generate better responses.”

More important than increased usage was the way students interacted with the agents. Students who asked follow-up questions, revisited their reasoning and engaged actively with the prompts generally appeared to engage more deeply with the learning process than those who primarily used the agents for answer retrieval.

These findings suggest that access to AI alone does not improve thinking. Instead, learning depends on both how AI is designed and how students choose to interact with it.

What does this mean for educators?

These experiences suggest that, rather than simply providing students with access to AI tools, educators may need to consider how AI interactions are structured to encourage reasoning, participation and reflection.

Our implementation suggests several practical considerations:

• Prioritise cognitive challenge: Design interactions that encourage reasoning rather than simple answer delivery. 
• Encourage multi-turn dialogue: Structured and iterative engagement is often more valuable than open-ended chat.
• Support AI literacy: Students need guidance on how to interact critically and effectively with AI systems.
• Ensure curriculum integration: AI agents are more meaningful when embedded within learning activities rather than positioned as optional add-ons.

Ultimately, the value of AI lies not in what it can generate, but in how it is designed to support learning. The challenge for educators is to create learning experiences where AI supports deeper human thinking, rather than replacing it.