88. How would you evaluate an AI assistant intended to summarize board papers before live adoption?
HardDigital records & AI governance
Case
Show model answer
Define permitted tasks, information boundaries, authoritative sources, prohibited actions, and required human review. Use fictional or properly authorized evaluation cases containing conflicting figures, stale policies, missing conditions, and material risks.
Measure consequential omissions, unsupported statements, provenance, confidentiality failures, and reviewer workload against a named baseline. Start with bounded assistance and clear stop criteria; fluent summaries or faster reading do not demonstrate a reliable governance decision process.
What interviewers listen for
- Define task, data, and action boundaries
- Evaluate material omissions and unsupported claims
- Compare reviewed outcomes against a named baseline