Add an AI Tutor to the LMS You Already Have

You Switched It On, and Nobody Used It

The sequence is almost always the same.

Your administrator enables the AI features in your learning platform. There is a brief flurry of interest. The assistant tidily summarises the page a learner is already reading, which is pleasant but solves nothing. Then someone asks a real question, the kind people actually have: which module covers the fire safety rules for our warehouse sites? The assistant produces a confident, fluent, entirely invented answer. Word gets round, and by week three the panel is closed.

Most organisations then reach one of two wrong conclusions. Either the plugin was bad and a better one exists, or the platform is outdated and needs replacing.

Neither is true. The AI in your LMS is doing exactly what it was built to do. It was never built to be a tutor. Once that is clear, the decision in front of you gets simpler, and far cheaper than a migration.

A One Minute Check on Your Own Installation

Open your Moodle site administration and go to General, then AI placements. Everything is disabled until someone switches it on, access is granted per role, and users have to accept an AI usage policy before their first use.

Now look at what the actions actually are. On Moodle 4.5 LTS there are three, spread across two placements, and 5.2 adds a fourth (MoodleDocs):

Placement	Actions	What they are for
Text editor	Generate text, Generate image	Drafting and illustrating, mostly for course authors
Course assistance	Summarise text, plus Explain text in 5.2	Shortening or unpacking content the reader already has open

Read that list again with a learner’s question in mind. Every action works on text that is already in front of the user, or writes new text from the model’s general knowledge. Not one of them goes and looks something up in your course library.

That is the whole story. The assistant did not know your content because nothing in the feature set was designed to find your content. It is a genuinely good authoring aid that has been widely mistaken for a tutor. Your administrator did not misconfigure anything, and there is no setting you missed.

The Five Things a Tutor Does That Summarising Cannot

Each one answers a complaint you have probably already heard.

1. It answers from your whole library, and shows its source. It makes things up. A tutor finds the passage that actually answers the question, writes the answer from it, and links to it so the learner can check. This single capability turns a novelty into something trainers will let learners use unsupervised.

2. It remembers the learner. I have to explain my situation every time. A tutor knows this person is in their second year, struggled with the same topic a fortnight ago, and has an assessment in three weeks.

3. It marks against your syllabus. It is encouraging but useless. Generic models are relentlessly positive. A tutor holds an answer against the criteria you are actually assessed on, which is the only feedback that moves a pass rate.

4. It changes route when someone struggles. Everyone gets the same thing. If a learner misses the same concept three times, repeating the explanation is not teaching. A tutor switches approach or hands the case to a human with a note about what went wrong.

5. It refuses to guess. We cannot trust it. The most valuable sentence a tutor can produce is that the answer is not in your material and here is who to ask. In regulated training, a model with no boundary is not a quirk. It is a liability.

Why No Setting Fixes This

A language model is like an exceptionally well read graduate on their first morning at your organisation. Enormous general knowledge, and no idea what is on your shared drive. Ask about your warehouse fire safety module and they will reason from what such modules usually contain. That is not lying. It is the only thing they can do.

What fixes it is retrieval, often called RAG. Your courses, handbooks, past papers and internal notes are indexed. When a question arrives, the system searches that index first, pulls the passages that are genuinely relevant, and hands them to the model with an instruction to answer from those and cite which one it used.

The same model that invented an answer now says: according to module 4.2, section 3, the requirement for your warehouse sites is X, and here is the link.

Retrieval is a layer between your content and the model, which is precisely why no checkbox in the LMS gets you there. For the architecture in depth, including models you host yourself, see our guide to custom AI on open-source models.

Three Ways to Close the Gap

These routes are not competing for the same organisation. For many readers the honest answer is the first one.

Route	Best when	What you give up
Community plugin	You want to test the idea, your content is small, no compliance pressure	Grounding in your library, learner memory, assessment, and any real support
SaaS tutor via LTI	Your needs are standard, speed matters more than fit, per-user pricing works at your headcount	Control of the model and the data, and the tutor’s behaviour is the vendor’s decision
Your own layer	Your material is specific, data cannot leave the EU, you train large groups, or an examination framework applies	A one-time build instead of a subscription, and you need a partner who does both instructional design and engineering

The useful question is not which is best in the abstract. It is this: if the tutor gets an answer wrong, whose problem is it? If the answer is “ours, and it matters”, you need control over the grounding and the guardrails, which points to the third row. If a wrong answer is merely annoying, buy the finished product and get on with your week.

For the full build, buy and white-label comparison with budgets, see our guide to custom AI tutor development.

How an External Tutor Plugs In

This is the part people expect to be difficult, and it is the part that is standardised.

An external tutor connects through LTI 1.3, the interoperability standard maintained by 1EdTech and supported by Moodle, ILIAS and the other major platforms. Three services do the work:

• Deep Linking 2.0 lets a trainer drop the tutor into a course as an activity, scoped to the right material.
• Names and Role Provisioning Services 2.0 passes who the learner is and what role they hold, so a trainer is treated differently from a first-year apprentice.
• Assignment and Grade Services creates a gradebook column and posts grades to it, so the work shows up in the reporting you already run.

Single sign-on comes with the launch. The learner clicks an activity in the course they are already in and lands in the tutor, already signed in.

Everything else stays put: enrolments, course structure, certificates, and the years of material your team has built. You are adding a layer, not migrating a platform.

Run This Before You Buy Anything

Copy this and use it on your current installation and on anything you are considering. Most of it can be answered in an afternoon.

Where the Data Goes

Every question a learner types is personal data in context. It says what they do not understand, what they are preparing for, and often where they work. Hosting is therefore a governance decision, not a technical preference.

The choice is already built into the platform. Moodle 4.5 LTS ships dedicated OpenAI and Azure providers and reaches anything else through the OpenAI-compatible layer, Ollama for locally hosted models included. Moodle 5.2 adds dedicated providers for Ollama, DeepSeek, Amazon Bedrock and Google Gemini (MoodleDocs). Running the model on your own EU infrastructure is a documented path, and the tutoring layer above it does not care which you pick, so you can prototype on a hosted API and move later.

Put three things in writing: the hosting location for the model and the search index, a data processing agreement covering both, and an explicit opt-out from your material training anyone’s models. For the wider compliance picture, see our guide to GDPR compliant AI tools for trainers.

One point that is easy to miss: the conversation logs are an asset. They are a running record of exactly what your learners do not understand, which your curriculum team has never had before.

What Drives the Price

Anyone who quotes a number before seeing your content is guessing. The connection is standardised and small. The cost lives in four places:

1. The state of your content. Clean, current, structured material indexes quickly. Twelve years of mixed PDFs, slide decks and scanned handouts is the real project.
2. How much is in scope. One course is a pilot. A whole catalogue in three languages is a programme.
3. How much the tutor has to judge. Answering questions is one thing. Marking against an examination framework and reporting a competency level is considerably more.
4. How strict the guardrails must be. In regulated training, the effort spent on what the tutor must never do can match the effort spent on what it should do.

The sensible way to find your number is to scope one course, put it in front of real learners, and measure whether questions to your trainers drop. That is also the right way to build it.

What to Do Next

Run the one minute check on your own installation. Take the readiness check into any vendor conversation and score it on your own content, not on a demo. Ask whether a wrong answer is your problem or merely an annoyance, because that single question decides between buying a product and building a layer you control. Then pilot one course before you touch the rest.

Let’s scope your first course. Tell us which platform you run and what your learners keep asking, and you get a no obligation quote with a concrete plan for a first working version. Or book a free strategy call: no pitch, and a realistic roadmap plus a ballpark figure at the end.

To see how we build AI tutors and learning platforms with EU hosting and ownership of the result, visit our page on custom e-learning development. Weighing up a full platform instead of a layer? Our guide to what custom e-learning costs covers that decision.