Every self-paced course has the same missing person: the tutor. The one who rephrases the concept, answers the question a learner is too embarrassed to ask twice, and nudges reflection along. An AI assistant can fill that seat — but only if it's designed as scaffolding, not a shortcut machine.
Ground it or don't bother
A generic chatbot bolted onto a course answers from the open internet, which means it answers slightly differently from your content — and learners notice the seams. The assistants that work are grounded: fed the course materials, the organisation's own documentation and the domain terminology, so every answer matches what the course actually teaches. That's information architecture as much as prompt design, and it's where most implementations fall over.
Design for the question behind the question
Learners rarely ask "explain constructive alignment." They ask "why did I get question four wrong?" A well-scoped tutor answers the second question by teaching the first — clarifying the concept, offering a worked example, then handing the thinking back with a follow-up. Laurillard's Conversational Framework is the useful blueprint here: cycles of acquisition, discussion, practice and reflection. The AI supplies the discussion cycle on demand, in private, at the learner's pace.
The accountability trap
The honest risk: learners can lean on assistance instead of learning from it. Two design moves counter it. First, make the assistant Socratic by default — it asks before it answers. Second, aim it at reflection tasks, where the AI's job is prompting the learner's own thinking rather than replacing it. In one recent build, the assistant's most valuable output wasn't answers at all — it was the question log, which told us exactly where the course was unclear.
AI won't rescue weak instructional design. But wrapped around a well-aligned course, it turns self-paced from "alone" into "supported" — and that's the difference between finishing and abandoning.