1. What specific work are we trying to improve?
The answer should describe a workflow, user and measurable problem. “Increase productivity” is too broad. “Reduce the time required for staff to locate approved policy information” is actionable.
2. What information will the system use?
Leaders should know whether the system will use public information, internal documents, personal information, client data, technical knowledge or regulated records.
The information type influences architecture, access, governance and whether a public AI service is appropriate.
3. Who remains responsible for the result?
AI can draft, classify, summarize or retrieve. It does not remove accountability from the person or organization using the output.
Before launch, define who reviews output, who can approve it, who handles errors and which decisions require direct human judgment.
4. How will employees learn to use it?
A licence is not an adoption plan. Employees need examples, boundaries, practice, feedback and a way to ask questions.
Leaders should budget for training, role-specific guidance and follow-up—not only software or infrastructure.
5. How will we decide whether to continue?
A pilot needs success measures and a review date. Measures can include time saved, error reduction, user adoption, information quality, service improvement or reduced search effort.
The review should also examine unexpected risk, support burden and whether the system changed behaviour in useful or harmful ways.
A sixth question: do we need AI at all?
Some problems are better solved with cleaner processes, better documents, simpler software or clearer accountability.
A credible AI partner should be willing to recommend a non-AI solution when it is the better fit.