Chile trial separates adoption from learning: tutor-use guidance improves final-exam performance

Teaching leaders and student-success teams may have broad AI access without a clear learning benefit. Include course faculty, tutors, accessibility staff and procurement. Ask how students use the tool, whether support prompts encourage independent reasoning and how exams assess that reasoning. A credible value hypothesis is that explicit guidance improves the quality of study. Offer one coordinated course pilot with an agreed comparison and an evaluation plan. The trial motivates testing a guidance package, not purchasing a particular platform. Do not promise its effect size locally, infer durable retention, equate increased usage with learning or propose guaranteed reductions in tutoring staff.

Fit a course assistant to approved instructional materials and an existing learning platform. Version content and guidance, keep source references available and provide an instructor escalation path. Prerequisites include a maintained content set, shared assessment rubric and an approved provider. Check answer quality, inappropriate answer disclosure and response behavior after model updates. Minimize student data and verify cloud processing terms; on-premises or hybrid designs require separate feasibility work.

Proposed proof of value

compare guidance with access alone using independently scored reasoning tasks and delayed reassessment, while recording participation and spillovers. The evidence concerns a teaching workflow; autonomous agents and student-information-system writes are unnecessary to this validation.

Course coordination should own the pilot calendar, faculty should approve prompts and teaching assistants should handle escalation. Prepare accessible guidance, orient students to verification, and preserve existing support channels. Dependencies include content approval, privacy review, assessment capacity and a process for correcting misleading responses. Proposed acceptance criteria include delivery of all planned guidance, complete reporting of participation and missing outcomes, independently scored learning comparisons and no unresolved critical privacy or accessibility defects. These are proposed gates rather than trial findings. Review delayed performance and support workload before expansion. Risks include peer contamination, self-selection in reporting and overgeneralizing from one course.