Claude can build a complete e-learning course. Script, quiz questions, pass mark, plus a SCORM package that tells the LMS who passed. Until recently we would not have expected it to go that far ourselves, until the L&D lead of a retail company walked us through her results in a meeting.
So whether a general-purpose AI can create training is settled. A different question is still open: who will be running all of it two years from now.
What a word processor and your finance system have to do with it
Claude writes good text. Nobody has abandoned their word processor. Claude calculates reliably. No company books its invoices through a chat window.
It is not that the AI could not do it. It is that a dedicated system comes with things no language model answer contains: someone who maintains it and stands behind it, a process that runs the same way whoever uses it, costs you can plan a year ahead, and a review by data protection and IT that happens once rather than for every new use case. Nobody introducing an accounting system asks whether it can add up. They ask who runs it, who installs the updates, what it costs after three years and who is accountable when a figure is wrong. Training raises exactly the same questions. They just get asked less often, because the first impression is so strong.
Which of them matter for you depends on your situation. The sections below are forks in the road. Read the ones that apply to you.
If you need only a few courses a year
Then Claude is a good choice, and there is no point pretending otherwise.
Three or four courses a year, built by one person who enjoys working with AI and has time for rework: no platform pays off here. The same goes for click-through walkthroughs of a new system, where speed matters more than instructional design. Claude handles those surprisingly well too.
If several people are meant to create training
This is where the road forks. A chat window is empty until someone knows what to put in it. The quality of the output depends on the prompt, and the prompt depends on the person.
So after a few weeks a small ecosystem tends to grow: skills for the structure, templates for the corporate identity, instructions for the quiz logic, a repository where all of it is versioned. Someone has to maintain it. Someone has to retest it after every model update, because a new model can read the same instruction differently. And when that person leaves, much of the knowledge of how the machine actually works leaves with them.
A training platform takes on exactly that part as a product. The workflow is guided, templates are held centrally, and subject experts without AI experience still reach a result with the same structure.
If a factual error is expensive
In our conversations with L&D teams, this is the fork where the decision most often falls.
Both routes deliver a first draft quickly. Then the real work begins: a technical term that is wrong, a passage that runs too long, a sentence after which the voice should pause. Built in-house, the correction goes through the script and the course is regenerated, and every pass can shift something new somewhere else. If you answer for the factual accuracy of a course, you do not check once but after every generation.
On a platform you see the slide, the narration and the timing, and change the one spot that is wrong. Everything else stays as it was.
If procurement asks what it will cost in two years
Claude bills through subscriptions with usage limits or, via the API, per token (Anthropic). For a single project that is inexpensive. In routine use the model has a property worth knowing: the better it works, the more it gets used, and the higher consumption climbs. A SCORM package with video, quiz and several revision rounds is not a small job.
Then there are the items no token bill shows. Time for maintaining and testing skills, extra services for voice and avatar, the back and forth with IT over every new connector. The full cost breakdown for training videos lists what else belongs in the calculation.
A platform bills through an agreed package. That is not always cheaper. It is predictable.
If data protection and IT have a say
And they will, as soon as company content flows into an AI. Built in-house, the review starts over for every use case and every connector: where is the data processed, which processing agreement applies, which service gets to see which content? Many IT departments therefore do not approve connectors at first, and the full power of the in-house build stays on paper.
With Omnora the review happens once: a processing agreement under Art. 28 GDPR, application, database and backups in data centres in Frankfurt am Main, AI services in EU regions of Microsoft Azure, and no use of customer data to train AI models. On top of that comes the TISAX label for confidential information, high availability and data protection. Optional premium avatars are the one exception: they are processed in the USA under the EU standard contractual clauses, and only once activated. How this carries through to the LMS is covered in SCORM explained.
And who keeps up with new models
This fork applies to everyone, including readers who have not turned off anywhere so far.
Models change month by month. Built in-house, that means watching, evaluating, adjusting instructions and testing whether the old courses still come out the same way. It is a task that appears in no L&D job description, and someone still has to do it. With a platform it is part of what you pay for: the vendor evaluates new models and methods and ships them in regular releases.
In fairness, there is one question we cannot answer: nobody knows how well general-purpose models will build training in two years. Perhaps the line will blur. The questions about operation, accountability, cost and data protection will not go away, though. They have not gone away for word processors and finance systems either.
The two routes are not mutually exclusive. What Claude produces as a concept or subject text can go into Omnora as a document and become a finished course there. Chaining several tools has its own price at every handover, and the calculation of those seams shows how it adds up.
The L&D lead from the opening, by the way, has not decided yet. She is building the same course in both tools right now. That is probably the most sensible answer to the question this article asks.



