The AI teacher in CoreSkills: practice and answers, no grades

In brief

In CoreSkills, AI stands beside the learner in two places: as the teacher, who goes through a task in conversation, and as the Ask the content conversation, which answers from the organisation's own material, courses and statutes with sources named. Both have the same limit: they help learning, and the practice assessment is advisory, but neither records competence. A task's result is decided by the learner's answer, and a certificate rests on a separate exam whose questions a person has approved. The limit is a founding decision of the product, and it follows from reliability and from the AI Act alike.

A teacher that asks first

The day's task can be gone through with the teacher. The teacher is an AI that discusses the subject, explains and asks. It always asks first whether the learner already knows the matter, and does not explain what is already mastered. A learner who wants to do the task directly chooses Without the teacher.

The way the teacher teaches can be adjusted in settings. The learner chooses whether the teacher starts from a practical example or from the rule and its reasoning, proceeds one point at a time or from the big picture, and gives a hint or a direct answer when something is missing from the answer. Each can also be answered "no preference". The choices affect only the conversation, not the task's result, the exam or certificates.

The teacher also comes to the learner. The day's tasks can arrive as an app notification and be gone through with the teacher in the chat in the bottom corner, and once the organisation has connected its Microsoft 365 environment, the CoreSkills teacher brings the day's tasks into Teams: teach starts them, today shows them, stop ends and help says what the bot can do.

Practice that records nothing

A course can be practised by talking with the AI. Practice is voluntary, and it gives an advisory view of how it went. It does not affect skills, the exam or the certificate. A skill is recorded when a course's last step is done, and the certificate comes from the exam.

The same applies to the skills-mapping interview. The AI asks about the person's work and proposes levels, but only the learner's own confirmation saves them, and they are saved as self-assessment, which steers recommendations but never a certificate.

Ask the content: an answer with a source

The chat in the bottom corner has a second tab, Ask the content. A question is written in one's own words: whom to report a security breach to, what the onboarding guide says about remote work, what the Cybersecurity Act says about management's liability. The AI answers from what the organisation's own material, the courses in use and the statute library say, and numbered sources sit under the answer, from which a course can be opened directly.

The answer is built only from that content. If the material says nothing on the matter, the AI says so rather than guess. A statute's wording is quoted as it is. Of the organisation's documents only those the administrator has permitted for the AI are read, and the statute library can be left out by an organisation setting. The conversation never sees another organisation's material or an unpublished course.

The conversation records no competence and affects no courses or certificates. It is the fastest way to check what a guideline or a law says, but it does not replace the responsible person when a decision has consequences. Earlier questions are in a list of their own, and a conversation can be deleted.

Search without AI

The search box in the header searches the same content as Ask the content: own material, the courses in use and the statute library. Results come in three groups, and under every hit is the passage where the search term was found. The search uses no AI, so it is ready at once and finds only what the material says. When the list is not yet an answer, the button Ask the AI about this opens the conversation with the same question.

This is a deliberate division: search when you know what you are looking for, conversation when the question is in your own words. Both see the same documents and only those.

Why an AI assessment records no competence

The line is drawn for two reasons, and they are the same ones for which a certificate rests on a multiple-choice exam.

The first is reliability. An AI's assessment can be gamed with AI: the same model that grades an answer writes it. A teacher that granted skills on the strength of a conversation would in the end only tell who can use a language model. So the teacher teaches and the exam verifies, and the exam's questions have been approved by a person.

The second is the AI Act. Evaluating the performance and behaviour of employees with AI belongs to the high-risk uses in the Act's Annex III, and it would bring the employer the obligations of a deployer of a high-risk system. CoreSkills does not do it: AI discusses, proposes and answers, the person answers and decides, and a deterministic exam verifies. How the exam and the certificate work is covered in Exam and certificate.

Where the messages go

The learner's own messages go to the AI as written, without the learner's name and email; the text is not anonymised. By default the text AI is the OpenAI API, whose terms say the data is not used to train models; an organisation can take its own Azure OpenAI into use, and then conversations run through its own tenant. Learners write into conversations themselves, so staff are instructed not to write customers' or other people's data into them. How the data is protected is covered in Where your data is.