LMS: AI Academy
AI Academy: Turning a 3-Week Course Build Into a 30-Minute Pipeline
Mentors at our LMS spend 2–3 weeks building a single course — writing outlines, producing slides, drafting assessments, then reconciling feedback across five stakeholders. The bottleneck isn’t creativity. It’s the mechanical assembly.
AI could generate all of it in minutes. But “generate a course” as a single button is more of a trust problem, not a speed problem.
How do you make AI generation fast enough to matter, but reviewable enough to trust?
Constraints
1. Errors compound
2. Review fatigue
3. Approval is organisational, not personal (audit trail)
So we split generation into five checkpoints, each with a human gate. AI proposes, human disposes.
The brief screen was deliberately kept to a single scroll. Every field we considered adding, we asked: can AI infer this? Most could be. Only Course Name and Level are required — the two things AI genuinely cannot guess.
Generation takes 30 seconds for an outline. Up to 20 minutes for full content. Traditional spinners break down at that scale.
Approval Has Three Levels
Early prototypes had one approve button per module. Testing showed reviewers wanted to approve some attachments and reject others.
So we built a nested approval hierarchy: Attachment → Module → Course
Each attachment carries its own status: Pending (amber), Approved (green), Changes Requested (red).
A module can only be approved when its attachments resolve. The course can only proceed when all modules resolve.
That amber banner solved a problem we didn’t anticipate: cross-course content reuse.
The button was disabled the entire time until this moment. No confirmation dialog, no error toast. The interface simply doesn’t allow an invalid state.
The Audit and Delivarable
Human review catches subjective issues. It’s bad at catching systemic ones, inconsistent terminology across 7 modules, missing examples, outdated practices.
So Step 4 runs an AI audit on the AI’s own output.
The output is a ZIP — not a proprietary format, not a locked platform.
Takeaways
Split AI generation into checkpoints. One-shot generation optimises for demo, not adoption.
Match approval granularity to how people actually review. We guessed module-level. Users needed attachment-level.
Disabled beats error. Preventing invalid states is cheaper than explaining them.
Surface prior decisions. The “approved earlier” banner was a small component with outsized impact.
Show the box before you ship the box. Trust in AI output is built by making it inspectable, not by claiming accuracy.
