Verify AI answers before you act.

What will change if you use this output? A brainstorm for a meeting has different consequences from instructions that alter a production system. State the intended use, the people affected, and the cost of an error. That tells you how much independent checking the task needs.

Do not paste confidential, personal, or customer information into a service until you have confirmed that the use is allowed by your organization and the service’s applicable terms. A tool available to you is not necessarily approved for every kind of information.

Separate facts, assumptions, recommendations, and unknowns. A draft may mix them into a single smooth paragraph. Rewrite the decisive statements as a short list. For each, ask what would make it true and where you can check it.

For an illustrative vendor comparison, decisive claims might include whether a feature exists on the plan you would buy, whether it supports your required region, and what happens to uploaded data. A general description of the vendor is not enough to support those decisions.

Ask the model to identify uncertainties and cite primary sources, but verify those sources yourself. Asking for citations is a request for more output, not proof that the output is grounded.

Follow the link. Confirm that the page exists, that it belongs to the relevant publisher, and that its text supports the specific claim. Check the date and whether the information applies to the exact product, version, region, or situation you are evaluating.

Prefer the source closest to the claim: official product documentation for a feature, the actual research paper for a research finding, or your own approved operating procedure for a company process. A source can be authentic and still be outdated or irrelevant.

Keep a short record: the claim, the source, when you checked it, and any limit that changes its use. If you cannot verify a claim, mark it unresolved rather than polishing it into a confident sentence.

If the output is code or a procedure, review it before trying it in a controlled environment appropriate to the task. Check ordinary cases, a failure case, and how you would recover. One successful example does not demonstrate reliability across every situation.

Name the person authorized to accept the result. Record which parts were checked, which remain uncertain, and what should trigger escalation. “The AI said so” does not tell a colleague who owns the decision.

If the task requires legal, medical, financial, or other specialist judgment, use the qualified reviewer your situation requires. The checklist helps organize evidence; it does not supply the expertise or authority to approve every decision.