Henry Ford reportedly said that if he had asked people what they wanted, they would have asked for faster horses. The quote gets used a lot to argue for radical innovation. But the more useful reading is gentler: people answer with the tools they already know, and that's a reasonable starting point.
LEARNTEC 2026 in Karlsruhe gave us a clean look at where the AI in learning market actually stands. We were there as speakers, presented the Sika pilot on stage together with Daniel Shavit, took questions from the audience afterwards, and watched the rest of the floor. We saw three patterns sitting next to each other, each solving a different problem.
A large group of vendors is using AI to speed up content production. You bring your storyboard, your subject matter expert outputs, your existing material. The tool drafts SCORM modules faster than a human team could.
This is real value, especially for organizations producing high volumes of compliance or product training. Localization, updates, version management get less painful. If your bottleneck is "we need to produce more courses, sooner," this pattern is built for you.
It doesn't change the shape of the course. A SCORM module made with AI assistance still plays back the same way as one built by hand. The faster part is upstream.
A second group is building chatbots directly into the LMS. The learner is in a course, has a question, opens the chat next to the player, and the system answers using content indexed at the platform level.
This is the pattern most large LMS vendors are converging on, and it makes sense from where they sit. If you already commit to one LMS and you want a single answer to the "where's the AI" question, embedding it in the LMS gives you a coherent story for IT, security, and procurement.
There are two trade-offs to be aware of. The first is portability. The chatbot is tied to that LMS, so if the same course needs to play through a different system, in a partner portal, or on a learning network for a specific business unit, the chatbot doesn't come along. For organizations with multiple LMS or external distribution, that becomes a real constraint.
The second trade-off is course context, and it applies even with a single LMS. Most LMS-embedded chatbots act more like a course-catalog navigator than a tutor inside the course. They answer from a platform-level index of all content, not from the specific slide the learner is currently looking at. That works well for "find me the right module" questions. It works less well when a learner is mid-course and wants the diagram on the current slide explained in their own language, where the value of the AI depends on knowing exactly where the learner is and what's on the screen. Pattern 3 below addresses that gap.
A smaller group, including us, takes a different approach. The tutor lives inside the SCORM package. It travels with the course. Whatever LMS plays the package, the tutor is there.
This pattern fits a different set of constraints. If you've already invested in a SCORM library, if you distribute courses across more than one LMS, if you want to layer AI onto existing content without waiting for a platform vendor to add the feature, the in-course tutor is the path with the fewest moving parts.
It's also the pattern we ran with Sika, which we presented on stage at LEARNTEC together with Daniel Shavit, a Corporate Training Manager from Sika. Sika has roughly one hundred technical e-learnings in English for a workforce of around 34,000 people spread across more than one hundred countries. The pilot started with a proof of concept on the three most demanding modules, evaluated by sixteen domain experts who were asked to push the tutor with the hardest questions they could think of. The next phase took twenty-three learners through the same modules in their regular training process. The tutor sat inside Sika's SCORM modules on adhesion and elasticity, answered questions during the course in the learner's language, grounded every answer in the slide content. Reported response times during the pilot phase: 3.7 seconds on average, ninety percent under seven seconds.
We've written separately about the production-side trade-offs in The Faster-Horse Trap in E-Learning. This post is about where the AI sits once the course is ready to play.
After our session, three questions came up that are worth writing down, because they tell you what the market is weighing right now.
"How do you keep the AI from making things up?" An audience member from a large LMS-using organization described the hallucinations they'd seen with an LMS-embedded AI feature. Our answer, from Jonas: we have a restrictive system prompt that grounds every response in the extracted course content, and makes it visible whenever the tutor steps outside the source material. Daniel added the customer side: through the entire pilot phase, his team had not forwarded a single quality complaint to us about the tutor inventing things. The architecture choice and the lived experience point the same way.
"Our courses aren't slides. They're lessons with branching and scenario-based learning." A fair pushback on simplified pitch language. The real challenge behind this question is how the tutor knows where the learner currently is in the course, so that an answer stays anchored to whatever is on the screen at that moment. For dynamic courses, JavaScript tags signal which lesson, chapter, or activity the learner is on. Another option is for the tutor to take a screenshot of the current screen state and read it directly. The exact method varies with the SCORM format, but the problem stays solvable even with more complex course structures. Either way, the tutor doesn't assume a flat slide deck, and the answer stays anchored to where the learner actually is.
"Can I add additional knowledge sources to the AI tutor inside the course?" Yes, and that's the path many pilots take once they move past the first couple of modules. The tutor's context can be extended beyond the SCORM package with additional documents, databases, or other reference material a domain expert wants in scope. Retrieval-Augmented Generation lets the tutor decide on its own when to reach for the supplementary material to answer a question more completely. The course stays the anchor, with the supporting material available whenever the tutor judges it useful.
Each of the three patterns answers a different starting point. Treating them as competitors misses the point.
If your bottleneck is content production volume, look at the authoring-side AI tools. They will save your team real time.
If your bottleneck is catalog navigation, helping learners find the right module across a large library with one consolidated IT story, the LMS-embedded chatbot covers that. It acts as a guide through the course catalog, not as a tutor inside the course itself.
If your bottleneck is what happens once the learner is inside a course, explaining a specific slide, answering a mid-course question in the learner's own language, quizzing on the topic just covered, then the in-course tutor pattern is what fits. The same applies if you have an existing SCORM library that already represents serious investment, if you distribute courses across more than one LMS or through partner networks, or if you need the AI capability without waiting for a platform vendor roadmap.
The two patterns also coexist well: a navigator at the LMS layer, a tutor inside each course. Most organizations end up with a mix over time, and we expect to see more of that, not less.
We put together a downloadable deck with our LEARNTEC sessions, the Sika pilot setup, and the architectural reasoning behind the in-course tutor. Get the LEARNTEC 2026 deck (PDF).
If you want to see how the in-course tutor pattern applies to a SCORM module you already use, we're happy to walk through a 30-minute demo with your team. Book a demo.
We're a Leipzig-based tech company, in business since 1998, with more than 25 years of experience integrating systems that have to coexist with what's already there. ScormIQ is the most recent expression of that habit: making the courses you already invested in smarter, without asking you to rebuild the rest of your stack.