DISCUSSION CASES

Practice the decision.
Not the perfect answer.

Three fictional scenarios for a leadership working session. No company, result or incident below is claimed to be real.

Assign one person to describe the business benefit and another to identify missing evidence. A third person records unresolved questions. Rotate those roles. This is a facilitation suggestion, not a tested training outcome or a promise of improved performance.

Case 1. A reply assistant gets a refund button.

Fictional situation. A service team has been using an assistant to draft responses that employees review. A proposed integration would let it issue refunds directly. The sponsor argues that the language quality is already good enough and the team is overloaded.

The hidden change. The system is no longer only suggesting text. It would be taking a financial action. The old writing evaluation does not answer whether refund conditions, authorization limits and exceptions are handled correctly.

Which actions may be delegated, and which require an authorized person?
What cases would test mistaken eligibility, repeated requests and unavailable account information?
Who can suspend the integration, and how would an incorrect action be corrected?

A useful output. Write two separate proposals: assisted drafting and bounded transaction authority. List the additional evidence required for the second. The exercise does not prescribe whether the organization should approve it.

Case 2. “A human checks it” is the whole control.

Fictional situation. A manager proposes using an AI-generated summary to prioritize applications for an internal opportunity. The system will not make the final selection, so the sponsor describes the risk as minimal. Reviewers have little time and usually see the summary before the underlying materials.

The hidden question. A person being present does not establish that review is meaningful. The sequence, workload and information available can influence how much the summary shapes the decision. Consequential employment-related use may also require specialist review under applicable rules; this scenario is not legal advice.

Could omissions in a summary disadvantage an applicant before a person reads the original?
Can reviewers question the summary and inspect the supporting material?
How can an affected person seek a correction or human explanation?

A useful output. Define what a reviewer must inspect independently and what information is prohibited from the workflow. Record missing evidence rather than assuming that a human signature resolves every concern.

Case 3. An internal tool becomes a customer channel.

Fictional situation. A knowledge assistant used by trained internal staff receives positive feedback. A commercial team wants to expose the same interface to customers. The internal pilot did not measure how unfamiliar users interpret uncertain answers or whether retrieved documents are appropriate for external disclosure.

The hidden change. The audience, information boundary and consequences have changed even if the underlying model has not. A successful internal demonstration is not evidence for a materially different deployment.

Which assumptions from the internal use case no longer hold?
What information may appear in a customer answer, and who owns that decision?
What explanation, escalation path and fallback would a customer need?

A useful output. Create a new decision record for the external use. Identify which prior evidence remains relevant and which must be collected again. A proposed launch date is not evidence of readiness.

Source context

These original fictional cases draw on the distinction between purpose, context, measurement and governance in the NIST AI RMF Core, and the accountability and transparency considerations in the OECD AI Principles. Neither organization endorses these exercises.