The three sources
AI-generated. Built from the position's skills, the seniority, and the job description. Generated ahead of the interview, not improvised during it, so every candidate for a stage faces the same planned set.
Your question bank. Questions your organisation has written or approved. Reusable across positions, and the natural home for anything specific to how you work.
Curated picks. Individual questions chosen by hand for this stage.
Mixing them
A stage can draw from more than one. The usual shape is a majority AI-generated for coverage, plus a handful of your own for the things only you would ask.
Where you set a mix, the platform respects the proportions and tells you if it cannot — asking for twenty bank questions when the bank holds eight is refused with the number, rather than quietly padding with generated ones.
Skill coverage
Questions are allocated per skill according to the weights on the position. A skill you marked mandatory gets more than one marked nice-to-have. If a skill would get fewer questions than can meaningfully assess it, you are told before you publish rather than finding out from a report that scored it on one answer.
Reviewing before candidates see them
Generated questions can be reviewed, edited and replaced before the flow is published. For a stage you will run many times this is worth doing once — every candidate sees the result.
Language
Questions are generated in the interview's language. Where a bank question was authored in another language it is translated before use, and a question that cannot be translated faithfully is left out rather than served half-converted.