Strategy question

You're launching a Text2SQL capability for intelligence analysts inside a regulated, air-gapped environment. How would you run discovery, validation, red-team testing, and phased rollout, and what launch gates would you require before broad deployment?

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Whether you can run a rigorous, safety-conscious launch process for a high-stakes AI capability in a regulated, isolated environment.

Run discovery inside the constraints of the environment: structured interviews with analysts about query patterns and current manual SQL workflows, without relying on data that cannot leave the air-gapped system.
Validate the approach against real analyst tasks using a representative, appropriately classified test set, checking both query accuracy and whether generated SQL respects data-access boundaries.
Run red-team testing specifically for this context: adversarial prompts attempting to access restricted data, generate misleading queries, or bypass intended query scope.
Design a phased rollout: start with a small group of trusted analysts on non-sensitive query types, expanding scope only as accuracy and safety hold up.
Set launch gates before broader deployment: minimum query accuracy on validated test sets, zero tolerance for access-boundary violations in red-team testing, and a clear human-review step for any query touching sensitive data.
Maintain an audit log of every generated query and its data access scope, since accountability is critical in this environment even after rollout.

Treats red-team testing for access-boundary violations as a hard launch gate, not an optional nice-to-have, appropriate for a regulated intelligence environment.
Designs the phased rollout around query sensitivity, not just user count, starting with non-sensitive query types.
Requires zero tolerance on access-boundary violations specifically, distinguishing that from a softer accuracy bar on general query correctness.
Plans for an audit log as a standing requirement, not a launch-only check, since ongoing accountability matters in this domain.

Treating this like a standard enterprise AI launch without addressing the air-gapped and classified constraints.
No red-team testing specifically targeting data-access violations, as opposed to general accuracy testing.
No audit or logging requirement for ongoing accountability after launch.

How would you red-team a system without exposing real classified data during testing?
What would you do if red-team testing finds a boundary violation after the phased rollout has started?
How would you decide when the pilot group is ready to expand?