Before You Launch AI: A Responsible Delivery Checklist
Responsible AI delivery is a product and operational discipline, not just a model setting. The right safeguards depend on what a system does, who uses it and what happens when it is wrong.
Define ownership and boundaries
NIST offers a voluntary AI Risk Management Framework to help organisations address AI risks throughout a system’s life cycle. For a practical project kickoff, our recommendation is to name an accountable owner, document the intended use and list actions the system must never take without approval.
Review the data path
Map what information enters the system, which services receive it, how long it is retained and who can access it. Use the minimum information needed. Do not assume that a prototype configuration is appropriate for sensitive production data.
Evaluate realistic failure cases
Create tests for incorrect answers, missing context, inappropriate requests and malicious instructions embedded in external content. Review the user experience when the system cannot complete a task. Make limitations visible and provide an accessible path to a person.
Launch with monitoring and a fallback
Our suggested release checklist includes a small pilot audience, documented approval points, a way to report errors, and a rollback plan. Monitor outcomes after launch, not just uptime. Re-run evaluations when prompts, models, tools or source documents change.
Keep the checklist alive
A review is not a one-time certificate of safety. Revisit controls as the product’s scope changes, and record what you learn from real incidents. Start with a system whose behaviour your team can explain and whose mistakes your team can correct.
Further reading: NIST: AI Risk Management Framework.