Start with an honest disclosure

Say that the visitor is interacting with an AI assistant. Explain its intended purpose and source boundary: it helps interpret reviewed careers content rather than deciding who should be hired. Do not imply that it is a recruiter, has live knowledge of every vacancy or can commit the employer to an outcome.

Show provenance and freshness

Answers need an accountable route back to their source. Link the reviewed careers page where practical, record when a source was reviewed and establish a process to update or withdraw outdated content. If the assistant is uncertain, it should say so plainly.

Review privacy for the actual deployment

Collect only what is necessary for the stated task and avoid inviting sensitive information merely to provide general guidance. Review transient conversation state, provider logs, retention, suppression, reporting and deletion for the chosen deployment. A central aggregate-report design does not prove every part of the data path is privacy-safe.

Provide human and non-chat routes

Not every question is suitable for a generated answer. Official application information, an accessible non-chat route and a named human-support path are important for adjustments, complaints, sensitive circumstances and out-of-scope questions. Those routes should come from the reviewed material.

Evaluate in the open

Set a representative evaluation set, test source support and refusal or handoff behaviour, and review failure patterns against one exact revision. A dated evaluation and a visible limitation are more meaningful than a broad accuracy claim.

Sources and further reading