Begin with a candidate problem

Define the audience, the questions they struggle to answer and the official next step. "Add an AI chatbot" is a technology instruction, not a service outcome. A narrow use case makes content ownership, evaluation and accountability easier to see.

Ask what supports each answer

Find out which employer pages, policies, FAQs and structured role data are included; who approves them; how conflicts are resolved; and how stale information is withdrawn. Ask whether the assistant can use the wider web or model memory for employer facts, and what happens when the reviewed material is silent.

Write down the product boundary

Separate careers guidance from application handling and employment decisions. Confirm whether vacancy data is a reviewed snapshot or a live ATS connection. Ask explicitly whether the system submits applications, reports application status, scores suitability, ranks candidates or runs an assessment. Do not let a general capability label stand in for release-specific evidence.

Review the whole data path

Ask what the browser sends, where transient conversation state lives, what the model provider logs, which reporting fields are retained and how deletion works. Aggregate metadata in one store does not prove that every other part of the path is content-free or privacy-safe.

Test accessibility and alternatives

Keyboard operation, visible focus, readable contrast, clear disclosure and an accessible non-chat route are part of the candidate service. Check the embedded experience in the real host-page context, not only in an isolated vendor demo.

Demand revision-bound evidence

Evaluation should identify the exact content and configuration revision, question set, result and remaining risks. Client approval should attach to that revision. It should not silently authorise a later draft or a production deployment.

Understand integration and exit

Record the hostname, embed boundary, content format, release owner, operational dependencies and recovery route. A contained pilot should not require an ATS migration. Ask how the service is removed or rolled back without losing the careers site or application journey.

A concise procurement checklist

  • Which candidate problem and audience are in scope?
  • Which reviewed sources support the answers, and who owns freshness?
  • Which capabilities are present in this exact release?
  • What happens when the information is absent or sensitive?
  • Does any part of the system rank, score or decide about candidates?
  • What data is sent, logged, retained, reported and deleted?
  • Can every essential task be completed without chat?
  • Which revision was evaluated, and who controls publication and deployment?
  • How can the experience be rolled back or removed?

Further reading