A short list of questions that separate a real AI capability from an \"AI-powered\" label

"AI-powered" has become a label you can put on almost anything — a dropdown that used to say 'auto-detect' now says 'AI-detected,' a rules-based alert threshold gets rebranded as a 'smart' one, a template library becomes 'AI-generated content.' None of that is necessarily dishonest, but very little of it tells you anything about whether a model is actually reasoning over your specific data or whether the word 'AI' was added to a feature that would function identically without it. As a buyer, you don't need to become a machine learning expert to tell the difference. You need a short list of questions that a real capability answers easily and a relabeled one doesn't.

Questions worth asking any vendor claiming an AI feature

• What specific input does the model see, and can you show me one real example of an output next to the input that produced it? Vague answers here are the biggest tell.
• What happens when the model is wrong — is there a human review step before anything consequential happens, or does the output act directly on your systems?
• Is the AI's action logged separately from any human approval, or is there just one combined record that makes it impossible to tell afterward who actually decided?
• Can the vendor point to what's live in their own production use today versus what's roadmap language dressed as a current feature?
• If you removed the AI branding, would the underlying mechanism still make sense as a real capability, or does the sentence only work because 'AI' is doing the load-bearing work?

What the answers tell you

A vendor with a real capability will show you the mechanism without flinching, because the mechanism is the interesting part and they built it on purpose. A vendor relabeling an existing feature will answer in adjectives — 'intelligent,' 'enterprise-grade,' 'next-generation' — because there's no specific input-to-output chain to point to. The gap between those two conversations is usually obvious within the first two questions, and it's worth pushing on it even when it's socially awkward to keep asking, because the awkwardness is cheap compared to buying a rebranded feature at an AI premium.

If a vendor can't show you one real input, one real output, and where the human checkpoint sits between them, you haven't found an AI capability — you've found a word.

We'd rather you hold us to exactly this standard than take our word for anything. Every AI claim on this site is written to distinguish what's running in our own practice today from what's design direction, because that distinction is the whole point of writing about it honestly in the first place — and if you can't tell which is which from how we've described something, that's a gap in our writing worth calling out, not a gap in your understanding.