96. What additional requirements matter when an AI assistant can take actions rather than only suggest text?
HardChange & outcomes
Technical
Show model answer
I'd define the exact authorized actions, data scope, approval boundaries and evidence required before execution. A helpful suggestion and a message, purchase or account change have different consequences, so permissions must not be inferred from a broad goal or untrusted content.
I'd define specialist-reviewed failures and external reconciliation. The workflow needs accountable permissions and tests; human-like language, a safety prompt or a reversible screen does not make consequential actions authorized.
What interviewers listen for
- Specify exact data and action scope
- Require consequential approval evidence
- Test failures and external-state recovery