We rebuilt the course builder around AI — here is what changed

For the first two years, Learnify’s course builder was a classic WYSIWYG editor: blocks, drag handles, a media library. It worked — and our customers’ median time to build a one-hour course was still 9 hours. This year we rebuilt the authoring flow around AI. The median is now 47 minutes. Here is what we changed, what we deliberately kept manual, and what we learned shipping it.

Why we rebuilt it

When we watched session recordings of L&D teams authoring courses, almost none of the 9 hours was spent writing. It went into converting: turning a 60-minute SOP video, a 40-page policy PDF or a trainer’s slide deck into modules, quizzes, summaries and completion criteria. The source material already existed. The editor was making people re-type it.

That reframed the problem. We didn’t need a better editor — we needed a better ingestion pipeline, with the editor as the review surface.

What actually changed

The new flow starts from your material, not from a blank page:

1. Drop in anything. Video, PDF, slides, a webinar recording or a plain document. The Copilot transcribes, segments and extracts the structure.
2. Get a draft course, not suggestions. It proposes a module split (aiming for 6–9 minutes per module), writes summaries, and generates a quiz bank with distractors drawn from the source material itself.
3. Review and edit in place. Every generated element shows its source passage, so reviewers verify against the original instead of fact-checking from memory.
4. Translate on publish. The same pipeline drafts vernacular versions for review — the step that used to gate every multi-language roll-out.

What we kept human

Generated content is never published directly. Quizzes, summaries and translations all land in a review queue, and the reviewer’s approval is recorded — which matters in regulated deployments where someone has to own what learners are taught. We also kept passing scores, completion criteria and assignment rules fully manual. AI is good at producing material; deciding what counts as competent is a policy decision, not a text problem.

Two more constraints we held. The Copilot calls a large-language-model API from our servers; customer content is used only to generate the draft and is not used to train models. And every quiz question it writes is checked against the source text — questions that fail that check are regenerated before an author ever sees them.

The results so far

• Median build time for a one-hour course: 9 hours → 47 minutes.
• Courses shipped per author per month roughly tripled across active workspaces.
• Vernacular versions per course went from 1.4 to 4.2 — translation stopped being the bottleneck.

The honest caveat: the Copilot is only as good as the source material. Teams that fed it outdated SOPs got beautifully structured outdated courses. Garbage in, faster garbage out — the review step exists for a reason.

See it with your own content

The best way to evaluate the builder is to start free and upload a real SOP or policy PDF yourself — or bring it to a demo and watch it become a course during the call.