Begin with a bounded candidate need

Choose one audience or journey and collect the questions people genuinely need answered. A graduate programme, a role family or a repeated eligibility question gives the team a clearer test than "put the whole careers site into AI".

Create a source register

List the pages, policies, FAQs and structured role data proposed for use. For each source, record an owner, review date, audience, jurisdiction and official status. Conflicting sources need resolution before they become conversational answers.

Separate facts, guidance and routes

Mark which statements are employer facts, which are explanatory guidance and which direct the candidate into an official process. The assistant should not turn a general example into a promise or replace the binding wording on an application page.

Write the unknown-answer behaviour

Decide what happens when the information is absent, stale or outside scope. A useful boundary names the limitation and points to a reviewed source or support route where one exists. It should not invent a contact or pretend every question has an answer.

Test for answerability

Use representative questions, paraphrases, edge cases and deliberately unsupported questions. Review the answer against the source, not merely for fluency. Record missing content separately from assistant behaviour so the right owner can act.

A readiness checklist

  • Every included source has an accountable owner.
  • Time-sensitive information has a review date or withdrawal route.
  • Official application and support routes are named in the source material.
  • Conflicts and unsupported questions have documented handling.
  • The exact compiled revision is available for evaluation and approval.

Further reading