Six questions HR should bring to an AI conversation

1. What work are we trying to improve?

“Become AI-ready” is too broad to guide an operating decision. Ask which task or workflow is changing, who uses the output and what a better outcome would look like. The answer should be concrete enough to evaluate after a pilot.

For example, drafting internal guidance and deciding an employment outcome are different use cases. They should not inherit the same assumptions simply because both involve an AI tool. This guide is a discussion framework, not an approval process for sensitive use cases.

2. What is inside—and outside—the boundary?

Agree which inputs, tools and outputs are approved for the experiment. Include where staff should go when they are unsure. Do not ask people to discover the rules through trial and error.

IMDA and the AI Verify Foundation’s 2024 generative-AI governance framework takes a whole-ecosystem approach, including accountability, data, testing and incident reporting. It is a useful starting reference, not a substitute for your organisation’s current policies or specialist review.

Source context: IMDA — Model AI Governance Framework for Generative AI, finalised May 2024

3. Who owns the decision after the tool produces an answer?

Map responsibility for the output, its review and its use. “A human is in the loop” is not enough if that person lacks time, relevant knowledge or authority to reject the output.

Use the table below to make the discussion concrete. The roles are illustrative; organisations should assign their own accountable people.

4. What support makes responsible use realistic?

Check access, time to learn, relevant examples and manager expectations. An employee cannot responsibly review outputs if performance targets leave no time to do so.

Discuss differences between roles rather than assuming a single training session suits everyone. A general introduction may build awareness; task-specific practice is a different need. Keep capability development connected to real work and the boundaries already agreed.

5. How will people report problems or challenge an answer?

Provide a route for incorrect outputs, confusing guidance and unexpected effects. Explain who responds, how urgent cases are handled and how lessons reach other teams.

Invite examples from employees who are hesitant as well as enthusiastic early adopters. Low usage may reflect a legitimate constraint. Treat that evidence as useful information before treating it as resistance.

6. What would count as progress?

Agree a baseline, a review date and the conditions for stopping or adapting. Look at the whole workflow and the experience of the people doing it, not only usage statistics.

Close the meeting with one owner per unresolved question. A list of open issues with accountable next steps is more useful than a confident declaration that the organisation is now AI-ready.

• Capture the decision, the owner and the evidence still missing.
• Separate a hypothesis from an observed result.
• Schedule a review before scaling to another team.