Practical guide
Questions to ask about AI at work
Plain-language answers for diagnosing AI adoption, capability, verification and work-quality problems without assuming specialist vocabulary.
The best question about AI at work usually begins with the decision or symptom: What changed, what evidence is missing, where can an error travel, and who remains accountable? Plain-language questions help teams distinguish adoption from capability, a weak output from a weak process, and a training gap from a workflow or governance problem.
Turn a vague concern into a testable question
Statements such as “people are not using AI well” combine several possible problems. The issue might be access, task selection, weak context, poor source use, inadequate checking, unclear policy or a final deliverable that does not meet the audience's needs.
Begin by locating the symptom in a real workflow. Ask which task was attempted, what good work would look like, what evidence was available, how AI was used and where a human decision remained necessary. This creates a question that evidence can actually answer.
NIST's map, measure and manage functions follow the same broad logic: understand context, choose appropriate methods and act on the resulting evidence. A diagnostic question should narrow uncertainty before a team jumps to a course, policy or new tool.
Sources: [1]
Separate adoption, capability, quality and risk questions
Adoption asks whether and where AI is used. Capability asks what people can understand and do. Quality asks whether a particular output meets its criteria. Risk asks what can go wrong, who may be affected and which controls are proportionate.
A dashboard may answer an adoption question while leaving capability open. A polished document may answer part of a quality question while hiding unsupported claims or omitted evidence. A responsible-use course may improve knowledge without showing whether participants transfer it to a realistic task.
Skills England's benchmark groups workplace AI foundations across technical, non-technical, responsible and ethical skills. That breadth is a reminder to avoid treating one familiar signal as the whole problem. Match the question to the evidence and keep adjacent questions distinct only when they change the decision.
Sources: [2] [1]
Ask what evidence would change the next action
A useful question has a consequence. Before collecting more data, state how different findings would change training, workflow design, review, access, policy or follow-up. If every possible answer leads to the same action, the question may be decorative rather than diagnostic.
For a verification concern, compare deliverables and workflow evidence from a bounded task, then look for checks tied to material claims. For an adoption concern, examine access and task fit before assuming resistance. For a training concern, test whether learners can transfer the behavior after instruction rather than relying only on completion.
The answer should remain proportionate. One failed check may justify a targeted workflow intervention, not a permanent conclusion about a person. One strong simulation may support a development discussion, not a claim about every future task.
- What decision are we trying to make?
- Which observable evidence would distinguish the competing explanations?
- What remains unknown even after that evidence is collected?
- What action would each plausible finding support?
Sources: [1] [3]
Sources
- 1.Artificial Intelligence Risk Management Framework (AI RMF 1.0) · National Institute of Standards and Technology
- 2.AI foundation skills for work benchmark · Skills England
- 3.AI and skills: What we know so far · OECD
Explore this topic
- Questions That Reveal the Quality of AI Adoption
“How many people use AI?” is a useful operational question, but it cannot show whether use is appropriate, verified, integrated into real work, or producing acceptable artifacts. Adoption-quality questions should follow the evidence chain from opportunity and behavior to output and consequence. They should help a team locate uncertainty, not invite a single reassuring percentage.
- Questions to Ask Before Designing an AI Work Assessment
Assessment-design questions should force clarity about purpose, opportunity, evidence, and interpretation before a scenario is built. The most important question is not “Which role should this simulate?” It is “What bounded decision will the evidence inform, and what must a participant be able to do for that evidence to exist?” The question set below prevents a polished task from measuring an.
- Questions to Ask Before Approving an AI Business Case
An AI business case should explain which problem is worth solving, which options were considered, what evidence supports costs and benefits, what risks and dependencies exist, and how the organization will decide whether to continue. The right questions prevent a compelling model demo from becoming an unsupported financial promise.
- Evidence Questions for AI-Enabled Work Programmes
Strong evidence questions prevent an AI programme from treating activity as capability or a polished output as proof of a sound process. Ask what decision the evidence must inform, what construct is being observed, where the opportunity to demonstrate it occurs, which records support the interpretation, what is missing, and what alternative explanation remains. Answers should end in an evidence.
- Questions to Ask Before Measuring AI-Enabled Work
The best measurement question is not “What can we count?” It is “What decision will this evidence change, and what would make the conclusion wrong?” A question-to-decision map helps programme owners distinguish adoption, activity, capability, work quality, risk, and business outcomes before choosing data.
- Questions That Clarify Ownership of an AI Programme
An AI programme often has many interested leaders but no single person accountable for each decision. A programme ownership map turns vague sponsorship into named rights, evidence duties, and escalation paths. The goal is not to force one executive to own everything; it is to make each decision traceable to the person or forum authorized to take it.
- Questions to ask about AI risk and governance at work
Useful AI governance questions turn broad concern into decisions that can be evidenced. Begin with seven questions: what action is proposed, whose interests are affected, which sources and data are authoritative, what can fail, who has authority to proceed or stop, what evidence demonstrates the controls operated, and when must the decision be reopened? The working artifact is a governance.
- Questions to ask about AI training impact
The central training-impact question is not whether people completed a course; it is whether relevant behavior changed in the work the programme intended to improve. A useful artifact is an impact evidence chain that connects the target task, learning activity, opportunity to apply, observed behavior, work outcome, comparison, and interpretation limit. The chain keeps participation, confidence,.
- Questions to ask when verifying AI-assisted work
Verification starts with the decision the work must support, not a generic instruction to “check the AI.” Ask: what must be true, which sources are authoritative, what transformations occurred, where could a material error hide, how will each failure be tested, who can resolve uncertainty, and what evidence proves the final output changed? The working artifact is a verification question ledger.
- Questions for designing AI-enabled work
Good AI work design starts by deciding how a task, evidence, judgment, and accountability should fit together. It does not start with “Where can we add AI?” The working artifact is a work-design question canvas that tests task boundaries, source authority, human contribution, exception handling, workload, learning, and release rights before a workflow is expanded.