AI learning assistant rollout in Canvas LMS
Operationalising an AI learning assistant across five university-partnered programmes and a career development track in Canvas — from configuration through live deployment, with the flows, escalation paths, and trust decisions that decide whether learners actually use it.
Project overview
LUMA is FourthRev's AI-powered learning assistant, embedded inside Canvas LMS and available to students across multiple university-partnered programmes and FourthRev's career development track. It provides on-demand academic support, course navigation, concept clarification, and escalation routing to human support — drawing answers from a walled-garden knowledge base of approved course materials.
I led the operational rollout from November 2025 through April 2026, working with LearnWise's implementation specialist, a supporting learning technologist, programme delivery managers, the Student Success Manager, and the Career Manager. LUMA is now live across five academic programmes and the career development courses.
Key features & outcomes
- Led the end-to-end operational rollout of LUMA across five academic programmes and FourthRev's career development courses
- Built and configured the bot in Canvas in collaboration with LearnWise's implementation specialist
- Designed conversational flows, escalation paths, and tone behaviour to match FourthRev's learner-support model
- Created the test process, knowledge base structure, and operational playbook used for each new programme deployment
- Coordinated across delivery managers, the Student Success Manager, and the Career Manager to align LUMA with each programme's existing learner-support workflows
- Used LearnWise monthly engagement reports to identify knowledge-base gaps, refine flows, and support ongoing tuning after launch
The original brief came from the delivery team: repetitive student support questions — assignment deadlines, submission processes, conceptual clarifications, navigation queries — were saturating facilitator inboxes across multiple programmes. The questions were predictable but persistent, and the volume scaled with student numbers, not with the complexity of the work.
An earlier prototype with a bespoke development vendor helped clarify what the team actually needed: stronger reporting, clearer bug-tracking processes, scalable configuration across multiple programmes, and a smoother learner experience. Those lessons shaped the second pilot — the one I led.
The deeper challenge: this was the first time FourthRev had deployed an AI learning assistant in production. There was no internal precedent for how to configure flows, how to handle escalations, how to integrate the assistant with existing programme-specific support workflows, or how to make it consistent across five different academic programmes plus a career development track. The pilot had to build the operational playbook alongside the product.
Configured the platform, programme by programme. Working with LearnWise's implementation specialist, I set up LUMA in Canvas across each programme — KCL Product Management, KCL UX & UI, LSE Data Analytics, Cambridge Data Science, Cambridge Cybersecurity, and FourthRev's Career Development track. Each programme had its own course content, its own learner support model, and its own escalation routes. The configuration had to be consistent enough to scale, and specific enough to fit each programme's actual learner workflow.
Designed conversational flows and escalation paths. The bot's behaviour isn't just a model and a knowledge base. It's a set of explicit flows that decide when LUMA answers, when it escalates, and where it routes. I defined the trigger phrases, the negative triggers (questions LUMA should answer from the knowledge base rather than escalate), and the escalation messages.
A worked example from the Career Development track: for career-related questions, LUMA distinguishes between content questions ("What is an elevator pitch?" — answer from the knowledge base) and personalised coaching requests ("I'm stuck on my elevator pitch and need feedback" — escalate to the learner's career coach). For administrative queries ("Who is my career coach?"), LUMA routes to a central career coaching inbox. For distress signals, LUMA confirms the programme and routes to the relevant student support email.
Built the operational scaffolding. The first deployment is the hardest because there's no template. I created the test process, the knowledge-base structure, the deployment checklist, and the documentation that the team now reuses for each new programme.
Coordinated across the stakeholders who own learner support. My role in the rollout was as much about cross-team alignment as it was about platform configuration. The bot only works if the escalation paths it triggers match the human support model already in place.
Worked with vendor reporting. LearnWise supplies monthly engagement and quality reports — conversation volumes, content-grounded rates, escalation rates, top topics. I use these to identify gaps in the knowledge base, refine flows that are mis-triggering, and tune the bot's tone for each programme.
The operational layer is the product. An AI learning assistant isn't just the model and the UI. It's the test process, the escalation paths, the bug-tracking, the integration with existing support workflows, and the reporting cadence. Most of the rollout work happened in the operational scaffolding around the bot, not in the bot itself.
Flows are instructional design. Defining when LUMA answers vs. when it escalates, what counts as a "personalised" request vs. a content question, how to phrase the escalation message — all of this is instructional design work, just applied to an AI conversation rather than a course page.
Trust is the thing that scales. AI learning assistants in Canvas only work if learners trust them. Trust comes from grounded answers, clear escalation paths, and a tone that meets students where they are. Most operational decisions came back to one question: does this strengthen learner trust, or weaken it?
Cross-team alignment is the hardest part. The technology was the easy bit. The hard bit was working with five delivery managers, a career manager, and the student success team to design escalation paths that matched each programme's actual support model. The rollout only worked because the operational playbook was negotiated, not imposed.