What should you check before shipping an AI feature?

Shipping an AI feature is not the same as shipping ordinary software. The model can be talked into doing things you did not plan for, and once it can take actions in the real world, a mistake can send an email, change a record, or charge a card. A short pre-launch checklist keeps the damage small when something goes wrong.

Two public references shape the items below. The OWASP AI Agent Security Cheat Sheet is a practical list for teams putting agents into production, covering least privilege, human approval, treating outside data as untrusted, output checks, and logging.^[1] The NIST AI Risk Management Framework is a voluntary framework released in January 2023, organized around four functions named Govern, Map, Measure, and Manage; its Manage function is where post-launch monitoring and the ability to turn a system off live.^[2]

The checklist is six items, split across the next two pages. First, lock down what the feature can do: give it the least access it needs, keep a human in the loop for anything irreversible, and filter what goes in and out. Second, make it observable and reversible: log its actions, test it before release, and keep a way to monitor and roll it back. None of these stop a model from being wrong. They make a wrong model survivable.