Before turning on a vendor's AI feature, confirm eight things: a named job, a named owner, a data path you understand, approvals set to manual, a tested off switch, a baseline to compare against, a sandbox trial result, and a review date. If any is missing, wait.
The eight line checklist
- Job. Write one sentence on what the feature is for. If you cannot, you cannot measure it.
- Owner. One named person is accountable, plus a backup.
- Data path. Write down what data the feature receives, where it goes, and how long it is kept. Confirm it against the vendor's written terms.
- Approvals. Set every category to manual approval. Relax later, with evidence.
- Off switch. Test it. Turn the feature off and verify that it stops.
- Baseline. Record how the work is done now, with timing and error rates, so you can compare.
- Trial. Run it in a sandbox or a low risk project and grade the outputs against criteria you wrote first.
- Review date. Put a date on the calendar, 30 to 90 days out, to decide whether to keep, expand, or switch off.
How to use the list
Print it, or paste it into the change ticket. Tick each line with a name and date. A feature that goes live without a ticked list is a feature nobody agreed to be accountable for.
What the list does not cover
It does not replace your vendor review, your legal review of contract terms, or your privacy assessment. Treat it as the last operational gate before a switch is flipped, after the larger reviews are done. This post is not legal advice.
Why manual first
The OWASP Top 10 for LLM Applications 2025 lists excessive agency among its risks. Starting with manual approvals limits what a mistake can cost. You can always loosen controls once you have seen the feature behave.
Why a review date
NIST's voluntary AI Risk Management Framework describes managing AI risk as continuing work. A review date is the smallest version of that. At the review, ask whether the feature did its job, what went wrong, and whether it should still be on.
After launch
Tell the people affected. Engineers who learn about an AI feature from a surprise pull request lose trust in the whole security program. A short note on what it does, what it cannot do, and how to report a bad output goes a long way.
Try it yourself
Click through it at your own pace, or open it full screen.
See how a gate holds a change until it passes checks, and think of your checklist as the gate for turning on a feature.
Where Safeguard fits
Safeguard keeps autonomous remediation off by default with every category on Manual, which matches line four. You can also try the free scan at gold.safeguard.sh/scan.
Quick check
Who signs off?
The feature owner, with the security lead for anything that can change code.
What if we skip the trial?
You will find out how it behaves in production instead. That is the costlier test.
Sources
- OWASP Top 10 for LLM Applications 2025: https://genai.owasp.org/llm-top-10/
- NIST, AI Risk Management Framework: https://www.nist.gov/itl/ai-risk-management-framework
Last reviewed: 19 October 2026.
