Strategy question

OpenAI is considering a zero-data-retention deployment for a government customer using agentic workflows. How would you define launch-readiness criteria and a go/no-go process that balances residual misuse risk, privacy constraints, and customer value?

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 define launch-readiness criteria and a decision process for a genuinely conflicting set of privacy, safety, and value requirements.

Define launch-readiness dimensions explicitly: technical readiness of zero-data-retention infrastructure, coverage of misuse-detection mechanisms that work without stored content, and validated value for the customer's specific agentic workflows.
Set a residual-risk threshold for misuse: since raw content cannot be reviewed after the fact, detection and prevention must happen in real time, so readiness requires demonstrating real-time detection works at an acceptable rate before launch, not as a post-launch improvement.
Set a privacy-verification step: independent confirmation (technical audit) that no raw content is retained anywhere in the pipeline, not just a policy statement.
Set a customer-value validation step: confirm through a scoped pilot or early access that the government customer's actual use cases work reliably under these constraints, since privacy and safety controls should not be validated in the abstract only.
Run the go or no-go decision through a defined cross-functional forum (privacy, security, policy, product) with documented decision rights, not an informal sign-off.
Set the go decision as conditional and reversible: launch with a defined rollback trigger if real-world misuse or reliability metrics diverge from pre-launch expectations.

Requires real-time detection to be proven pre-launch specifically because zero retention removes the ability to review after the fact, a precise and correct reading of the constraint.
Separates privacy verification (a technical audit) from a policy statement, since a written privacy commitment alone is not suffficient evidence of compliance.
Includes a scoped pilot to validate real customer value under the constraints, not just theoretical risk analysis.
Frames the go decision as conditional and reversible with a rollback trigger, appropriate for a first-of-its-kind government deployment.

Treating written privacy policy as sufficient evidence without a technical audit.
Assuming misuse can be caught and reviewed after launch, which zero-data-retention makes impossible.
No rollback plan, treating the go decision as final and irreversible.

How would you technically audit that no raw content is retained anywhere in the pipeline?
What would you do if the pilot reveals real customer value but detection reliability is borderline?
Who has final decision rights if privacy and product value disagree on readiness?