New hires can read many documents and still not know what to deliver today. Asking AI to “make an onboarding plan” from a large document dump often produces a polished but untestable checklist. Start instead with the first real tasks a person must complete, then organise approved material, practice and review around those tasks.
Use NotebookLM for constrained sources and ChatGPT or Claude for practice design. Managers still confirm access, customer communication and quality.
Define tasks, not knowledge topics
Choose three to five tasks for the first two weeks. For each, specify input, acceptable output, common errors and reviewer. “Use the CRM” is vague; creating a lead from an email, marking uncertain fields and setting the next action is reviewable.
Use an approved source list
Include current procedures, approved templates, help material, redacted cases and role permissions with an owner and effective date. Exclude old rules, customer data, credentials and unapproved roadmap material. Mark conflicts for review rather than letting AI select a version.
Find gaps before generating steps
Ask AI to list facts needed for each task, corresponding sources, missing material, misused rules and manager-confirmation boundaries. “Evidence missing” is a useful result; it should become a documentation task.
Use demonstrate, practise, review
A skilled colleague demonstrates once with redacted data; the hire practises in a safe environment; a reviewer uses a fixed checklist. AI can generate simulated cases only from approved flow and must not invent customers, orders, prices or policy exceptions.
Review the knowledge base with the cohort
Track where new hires stop, which outdated sources they cite and what must be repeatedly asked. Update materials and practice cases by problem cluster. Speed alone is not success: correctly escalating an uncertain case is better than confidently making a wrong commitment.