AI inside the LMS, under your admin’s control

One provider, one key, one master switch

Everything starts on Settings, AI Integration. You choose Anthropic (Claude) or OpenAI (GPT), pick a current model from that provider’s list (for example Claude Sonnet 5.5 or GPT-6), set the reasoning effort the model should use, and paste the API key, which is stored encrypted. The “Enable AI Features” switch sits above everything else: turn it off and every AI tool stops, tutor and reporting agent included. Below it are separate switches for the reporting agent, the AI email writer, the course recommender and the assignment grader, and the tutor button has its own toggle in lesson settings, so you can approve features one at a time.

Watching cost

You enter a cost per 1,000 tokens, and the AI Usage page turns logged calls into an estimate. It shows the current month by feature, daily calls for the last 30 days, the ten heaviest users, and a list of individual calls you can filter by feature, with model and token counts. The settings tab also holds a monthly token budget, and the usage page shows the month’s total against it.

What the AI Tutor will and will not answer

In lesson settings you set the tutor’s display name and avatar: an animated character that speaks its answers, or an instructor photo with no animation. The tutor draws on the course’s lesson text, video captions and notes, and you can let it use your knowledge base too. It is instructed to answer only about that course and decline anything else in one sentence. It remembers the last ten messages of each learner’s conversation on that lesson, accepts voice input, and replies in the language the learner uses the platform in. Admins can read tutor conversations in a log, which is the quickest way to see what learners keep asking and which lesson needs a clearer explanation. Tutor conversations can be purged automatically under data retention.

Captions, translation and AI video

AI captions are off by default. Switch them on in Settings, Lessons, and an uploaded video of up to 25 MB is sent to OpenAI Whisper. The captions come back as a draft, and an admin publishes them once they read correctly, which matters for product names and jargon. AI-translate sits with the other lesson tools for academies that teach the same lesson in more than one language.

The AI video block renders a script of 3 to 4,000 characters as an MP4 in the background. Pick narration, where one of 6 OpenAI text-to-speech voices reads over a still image, or a lip-synced talking avatar made from an image through D-ID (English only, 8 voices). Captions and narration need an OpenAI key even if your main provider is Anthropic, and the avatar needs a D-ID key.

Where a person stays in the loop

Generated courses are saved as drafts. The interactive module assistant drafts 6 to 12 blocks, including at least one knowledge check, into the builder for you to edit. Question Bank tools only propose: new questions appear in a preview you pick from, audit findings are sorted from high to low severity with a suggested fix, and auto-tagging only fills fields that are empty. The grader writes a suggested score and feedback for an instructor to release. There is an auto-approve switch that skips that review. It is off by default, and we suggest keeping it off for any grade that counts.

The reporting agent is read-only

The agent answers questions like “which quiz in Fire Safety has the lowest pass rate?” by calling a fixed set of queries: course lists, course stats, lessons where learners drop off, quiz performance, enrollment trends, top students and a platform overview. It is told never to state a number a query did not return, and it cannot create, edit or delete anything.