AI features: signal vs. sticker
Every SaaS product now claims AI. A short field guide to telling a working capability from a label — the questions we ask when we review one.
Reviewing software in 2026 means wading through AI claims on nearly every homepage. Some describe real capabilities. Many describe a text box wired to someone else's model. The gap matters, because the sticker is priced like the signal.
Here's the checklist we use when a listing claims AI — the same questions worth asking before you pay for one.
Does the feature have a job?
A real AI capability is attached to a workflow verb: it drafts the reply, extracts the invoice fields, flags the anomaly. A sticker is attached to a noun — "AI-powered platform" — with no sentence anywhere that says what it does to which input. If the marketing page can't name the job, the feature usually can't do one.
Is there an input you control and an output you can judge?
Working AI features let you see the transformation: here's what went in, here's what came out, here's where to fix it. Be suspicious of features that only ever show results in a demo video, and of outputs you can't correct — a summarizer with no way to see the source isn't a tool, it's a slot machine.
What happens when it's wrong?
Every model is wrong sometimes; the product's job is to make that survivable. Look for the unglamorous furniture: confidence signals, review steps, undo, an audit trail of what the AI changed. A product that pretends its AI is never wrong has designed for the demo, not for your Tuesday afternoon.
Where does your data go?
An AI feature is also a data flow. The questions are old ones, sharpened: which third-party model sees your content, is it retained, is it trained on, can you turn the feature off and keep the product? A vendor that answers these in its docs is taking the feature seriously. A vendor that answers them only in a sales call is not answering them.
Does the price make sense?
Inference costs money, so genuinely AI-heavy features usually meter something — seats, credits, volume. "Unlimited AI" at a flat low price tends to mean either a very small model or a very short-lived pricing page. Neither is disqualifying; both are worth knowing before you build a process on top.
None of this requires machine-learning expertise. It's the same discipline as any software evaluation: ignore the adjective, find the verb, test the failure case. When we score a product's AI claims in a review, this is the lens — and when a product passes it, we say so with the evidence attached.