Designing Conversational Agents for Adaptive Instructional Support in Business Simulation Gaming

Synthesis: Wenzel, Geiger, and Liening use Action Design Research to design, build, and evaluate "Lara," an AI-enhanced conversational agent that provides adaptive instructional support within business simulation games (BSGs) for entrepreneurial learning. Grounded in educational psychology and information systems research — spanning cognitive load, self-determination, emotion, and social presence theories — the study produces the CAIS-GBL framework (four design principles, fifteen design features) instantiated in Lara and evaluated across two build–intervention–evaluation cycles. The design is deliberately framed by an equity-by-design stance and the Universal Design for Learning (UDL) principles, aiming to provide adaptive support that removes learning barriers rather than reinforcing educational disparities.

Key Findings

• Consistently positive perceptions across all five dimensions. In the five-week field study (n = 112 users), students rated Lara's cognitive presence, support for self-regulated learning, social presence, perceived ease of use, and behavioral intention all significantly above the theoretical midpoint.
• Perceptions evolved over time. Social presence and perceived support for self-regulated learning increased significantly across the five BSG rounds (e.g., self-regulated learning rose from β = 0.254 in round 2 to β = 0.470 in round 5), while cognitive presence stayed stably high, ease of use declined slightly, and engagement peaked during the simulated crisis in week 3.
• Dispositions, not demographics, shape acceptance. Gender, age, and prior knowledge did not significantly predict CA perceptions, whereas AI attitude and the "Socialiser" player type positively predicted social presence, cognitive presence, and support for self-regulated learning; the "Achiever" type showed no significant effects.
• A needs analysis revealed mostly cognitive hurdles. Of 61 reported hurdles across the BSG, 68.9% were cognitive — shifting from information selection early on to organization/integration for strategic decisions as the game progressed — highlighting demand for formative feedback and on-demand guidance.
• Two iterative evaluation cycles refined the artifact. An alpha workshop with graduate student teachers surfaced improvement areas in cognitive and social presence (e.g., reducing artificial follow-up questions), which informed the beta version while the CAIS-GBL framework was retained.
• Equity-by-design is integral, not incidental. The authors caution that the absence of demographic effects does not guarantee equitable benefit, since attitudinal and motivational profiles still shaped perceptions and may warrant targeted support to prevent disengagement.

What this means for practice

• Instructional designers. Adopt the CAIS-GBL framework (four design principles, fifteen design features) as a reusable template for conversational agents in digital game-based learning, with metric-based formative feedback, frictionless onboarding, empathic interaction, and sociability and equity by design.
• Instructors. Target support at learners' AI attitudes and game-related motivations rather than at demographics: gender, age, and prior knowledge did not predict perceptions, while AI attitude and the "Socialiser" player type did.
• Designers. Ground agent feedback exclusively in in-game performance metrics, minimize personal data, and make support transparent and contestable so the agent removes barriers instead of reinforcing disparities.
• Instructional designers. Plan for acceptance to build over the term: perceived social presence and self-regulated learning support rose across the five simulation rounds (β = 0.254 in round 2 to β = 0.470 in round 5).
• Instructors. Use in-game metrics to deliver timely support at scale where one-to-one human support is infeasible, since 68.9% of the 61 reported learning hurdles were cognitive rather than technical.

Limitations

• No comparison group: the authors state they cannot isolate Lara's added value relative to using the business simulation game without the agent, so incremental impact on learning processes remains unestimated.
• All five perception dimensions are self-reported, and the beta analysis rests on the 83 of 122 invited students who completed all five questionnaires; attrition across the multi-week study reduced statistical power and constrained model complexity.
• Adaptive support was driven mainly by group-level performance, so it may not accurately reflect individual knowledge or behavior.
• The evaluation sits in a single institution and two courses with a relatively homogeneous age range (M = 23.0, SD = 2.59), and the needs analysis drew on n = 100 students in one course, limiting generalizability.

Citation

Wenzel, A., Geiger, J.-M., & Liening, A. (2026). Designing Conversational Agents for Adaptive Instructional Support in Business Simulation Gaming. Computers and Education: Artificial Intelligence, 10, 100576.