AI personal tutor

Open the tutor only after the learner's current assignment is loaded as the source of truth. The session should know which module, version, and passage set it may use. Without that bind, the chat becomes a second curriculum L&D never approved.

Load the assigned module first, then start the dialogue. Just-in-time answers and Socratic questions help when they stay inside that material. They harm when they pull in a neighboring policy, last year's deck, or a general HR rule the learner was never assigned.

HR learning stacks already hold the assignment. Systems in the class of Workday, Lattice, Degreed, and Cornerstone typically own who is enrolled, which version they see, and when the item is due. The tutor should consume that assignment, not invent a parallel catalog. Pair the session with role-based learning auto-assignment so the person in the chat is the person who was given the module.

Related is not assigned. If the overtime module is loaded, the tutor may unpack overtime. It may not add travel-time rules from a different handbook chapter because the question sounded similar. In the first tutor turn, name the module, the version date L&D published, and the section the learner is in. If any of those are unknown, stop and reload. Do not guess the pack from the job title.

A usable tutor turn is an explanation that points at the assigned content passage. The cite is the quality check. Without it, the reply is an ungrounded essay, even if the tone is careful.

Require the model to quote or locate the passage before it interprets. "According to section 4.2 of this module" is the start of an answer. "In most companies, overtime usually works like this" is not. The second pattern is the failure mode of an answer with no passage cite.

If the passage is missing, leave the field empty. Do not paraphrase from memory. Do not fill the gap with a neighboring paragraph that "covers the spirit." Empty is the correct output when the assigned text does not contain the fact. L&D can then see the blank, fix the module, and republish. A fluent guess hides the gap until someone acts on it.

Socratic dialogue follows the same rule. Ask questions that send the learner back into the loaded text: which sentence defines eligibility, which exception is listed, which step comes first. Do not ask questions that only make sense if the learner already knows an outside fact. If the module never names a threshold, the tutor must not imply one in a leading question.

Just-in-time answers should be short and local. Pull the sentence the learner is stuck on, explain the term in that sentence, and stop. Expanding into a mini-lecture is how outside material sneaks in. If the learner needs practice after they understand the passage, send them to a separate practice surface such as ai roleplay skill coach rather than stretching the tutor into a rehearsal.

When a cite cannot be produced, the tutor should say that the assigned module does not cover the question and should leave the explanation blank. That refusal is a quality outcome. It is not a product failure.

A people manager opens the assigned "Overtime and rest periods" module, version dated for this year, and asks the tutor: "Does a salaried team lead get overtime after a late customer call?"

The tutor already has the module loaded. It searches the assigned passages. Section 3 defines which roles are overtime-eligible. Section 5 lists after-hours customer work as ordinary working time when the person is eligible, and as unpaid extra duty when they are not. The tutor answers by pointing at those two passages, then asks the manager to identify which eligibility line matches the team lead in front of them. That is Socratic dialogue grounded in the text, not a second opinion from the model.

Suppose instead the manager asks whether travel time to the customer site counts. The loaded module never mentions travel. The tutor does not invent a fact from outside the assigned material. It does not borrow a sentence from the expenses policy, a public labor FAQ, or "what we did last year." It leaves the explanation empty and states that travel is not in this module. L&D can add a passage or point the manager at a different assignment.

The wrong version of the same session looks polished and fails the quality bar. The tutor answers without a cite, tells the manager that salaried leads "generally" are exempt, and offers a confident travel-time rule from general knowledge. The manager treats the chat as the decision. Nobody can show which approved sentence was used, because none was. That transcript must not be stored as evidence that the manager understood the module.

The second question is the test. Empty is the pass.

L&D still owns the curriculum. The tutor explains what is already assigned. It does not add modules, rewrite learning objectives, or promote a chat summary into a new official policy.

Do not auto-certify. Completing a dialogue is not completing the course. A fluent session can hide that the learner never opened the passage. Completion belongs to the assessment, attestation, or observed practice L&D already defined for that item. Treating the chat as a certificate is the second failure mode. It collapses tutoring into compliance theater.

If the organization wants a record, store the cites the tutor used, the blanks it left, and the module version, not a grade inferred from conversational tone. Those records help L&D see where the material is thin. They are not a passing score.

When skills evidence is needed later, collect it from work, not from tutor chat. Skills inference from work artifacts looks at what the person produced. The tutor makes the assigned text usable in the moment. Mixing the two invites a loop where the model teaches a fact, then treats its own teaching as proof the person can do the work.

The same split applies at review time. A performance review draft generator should not treat tutor transcripts as demonstrated competence. At most, note that the person opened the assigned module. Whether they can apply it belongs to evidence L&D and the manager already trust.

Curriculum change has a path: a repeated blank on the same question is a signal to edit the module. The edit is a human publishing act. The tutor may surface the gap. It may not silently patch the gap with generated text and keep teaching from the patch.

Put the tutor in the develop stage as a reading aid, not as the operating system for HR learning. Assignment decides who gets which pack. Tutoring helps them through that pack. Practice lets them try a conversation or a task. Review judges performance with human-owned criteria.

Keep those jobs in separate artifacts so failure modes stay visible. If assignment and tutoring share a blob of chat, you cannot tell whether the person was given the right module. If tutoring and certification share a blob of chat, you cannot tell whether they passed. If tutoring and review share a blob of chat, you score fluency instead of work.

Operational checks before you turn the tutor on:

• The session cannot start until the assigned module and version are loaded.
• Every non-empty explanation includes a cite to a passage in that version.
• A missing passage produces a blank, not a paraphrase, not an outside fact.
• Completion, badge, and legal attestation stay on the L&D-owned assessment.
• Transcripts are coaching aids. They are not certificates and not review evidence.

Vendors in the Workday, Lattice, Degreed, and Cornerstone class can remain the system of record for enrollment and completion. The tutor sits beside that record. It should not replace the catalog, the quiz, or the attester.

If a leader asks whether the tutor "covers" a topic the module omits, the honest answer is no. Coverage is a curriculum decision. The tutor's quality bar is narrower and stricter: explain what is in the assigned passage, point to it, and stay quiet when it is not there.