A recent MIT report put a number on something many leaders in higher education already suspect: 95% of generative AI pilots are failing to deliver measurable return on investment or move beyond the pilot gate. The question is why, and what the exceptions have in common.

‍

Eric's explanation starts with design. Generic, large language models are built to answer everything. For a student trying to understand a specific concept in a specific course, that breadth becomes noise. A 25-page response to a simple question is not useful. It is overwhelming.

‍

The University of Luxembourg took a different path. Rather than deploying a general-purpose AI tool and hoping for adoption, Eric rebuilt his courses from the ground up, defining learning objectives in precise detail and using them to scope and guardrail the LLM to exactly what students needed to learn.

‍

Key points from the conversation:

• Generic AI tools generate broad, often irrelevant responses that are poorly suited to the precision learning requires. The alternative is not a bigger model. It is a better-framed one.
• Guardrails are not a constraint. They are what makes the tool useful. When a student knows the LLM will only surface content relevant to their course, the noise disappears and the learning signal becomes clear.
• The distinction that matters most is philosophical: AI as a companion for knowledge acquisition, not a tool for generating answers to assignments. These are fundamentally different design orientations and they produce fundamentally different outcomes.
• Intentionality in design is what separates implementations that scale from pilots that stall. As Laurie notes, the research on technology in classrooms is consistent: tools introduced without purpose, scaffolding, or pedagogical intent rarely improve learning. The same applies to AI.
• Lecture, EDT&Partners' open-source generative AI framework, gave the University of Luxembourg the infrastructure to implement AI with that level of intentionality: scoped, measurable, and designed for teaching and learning from the ground up.

‍

"It's not a replacement. It's something which accompanies the student at whatever level he or she is, so that we are enhancing the capability of knowledge acquisition." — Professor Eric Tschirhart

‍

The conversation points to something the sector has been slow to say clearly: most AI pilots fail not because the technology is wrong, but because the framing is wrong. Deploying a general-purpose tool into an educational context and measuring whether students use it is not the same as designing an AI solution around what students need to learn and then measuring whether they learn it.

‍

That shift in perspective, from deployment to design, is where the real transformation begins.

‍

For EDT&Partners, this is a question we work through with every institution we partner with. The University of Luxembourg conversation is one data point in a wider pattern: the institutions seeing the most meaningful outcomes from AI are those that have invested in the design layer first.

‍

That means defining learning objectives before selecting tools. It means building guardrails that serve students rather than restrict them. And it means treating AI as part of a pedagogical approach, not a standalone solution.

‍

Stay tuned for more from the full conversation between Laurie and Eric, where they explore faculty resistance, student behaviour in the lead-up to exams, and what the next phase of AI in higher education might look like.

‍