Start with the sentence you are being asked to believe.
“This tool can draft an email” and “this tool can run your sales process” are very different claims. Write the exact claim in one sentence. Then separate the source’s words from your own interpretation and any suggested next step.
Five questions before you act
- Who is making the claim?
Find the original publisher. A company announcement is useful evidence of what a company announced, not an independent test of its product.
- What actually happened?
Was this a recording, a written example, a live demonstration, a benchmark, or a workflow tested repeatedly? Check whether unsuccessful attempts, human corrections and setup work are described.
- Under what conditions?
Look for the model or product version, account access, inputs, tools, review effort and evaluation criteria. Missing details limit what you can conclude.
- What matters for your task?
A writing example does not establish accuracy with your records, permission to access them, or reliability when sending something to a customer. Keep a small trial separate from live business action.
- What would change your mind?
Before testing, name a failure that would make you stop: an invented fact, a missed constraint, exposed sensitive information, or an action without approval. Define success in terms of your task, not the excitement of the demonstration.
NIST’s voluntary AI Risk Management Framework addresses trustworthiness in the design, use and evaluation of AI. Use it as a further reference when a task needs more structured risk evaluation; completing this guide is not a NIST assessment.
Explore the NIST frameworkFictional example
“Five posts in seconds. Your marketing, automated.”
A fictional vendor video shows five social posts appearing after one prompt. The video does not show fact-checking, image rights, approvals, scheduling, customer response or failed attempts.
Your exercise: Write one sentence stating what is shown, one stating what is unknown, and one small test you would run without publishing anything.
Compare with a worked answer
Shown: The demonstration presents five generated drafts from a prompt. It does not show the whole marketing process.
Unknown: We do not know whether the claims are accurate, whether drafts respect the brand’s constraints, or how much editing they require.
A bounded test: Use a fictional product description to request one draft. Check every factual claim against that description and record the revisions needed. Do not schedule or publish it.
A more defensible conclusion: This may be worth testing as a drafting aid. The demonstration alone does not establish end-to-end marketing automation.
Turn the lesson into a review habit.
Keep a short note with the claim, source, date, conditions, limitations and your next check. Avoid entering private customer information just to try a tool. Stop when an important condition cannot be checked.
This is a reading and reasoning exercise. It does not connect to an AI model, collect your answers, or certify the safety of a tool.