"AI-powered" is the most inflated three words in software marketing right now. Every vendor with a text box in their app has slapped it on the homepage. The label costs nothing to add and means whatever the sales page needs it to mean that week.
So what does AI-powered mean, in practice, when a vendor prints it under their logo. Usually one of three things, and none of them require the vendor to have built anything new.
What does AI-powered mean, usually
First: an autocomplete field. The system watches what you type and finishes the sentence, the way a phone keyboard suggests the next word. That's a real feature. It is not intelligence, and it was possible with basic language models years before anyone printed "AI" on the button.
Second: a chatbot wrapper around a help center. The vendor takes their existing FAQ pages, feeds them into an off-the-shelf model, and puts a chat window in the corner of the app. The bot answers by paraphrasing the same articles the help center already had. Useful, sometimes. A new capability, no.
Third: a rebrand of last year's feature. Whatever the product already did with a rules engine or a lookup table gets a fresh name at the next release. "Smart scheduling" becomes "AI Scheduling." The logic underneath didn't change. Only the word next to it did.
None of these three is fraud exactly. It's marketing doing what marketing does: reaching for the word that sells this quarter. This is the pattern people mean when they call it AI washing. The problem is the label tells you nothing about whether the thing solves your problem, or whether it works on your data at all.
Questions that expose the label
You don't need to be technical to test the claim. Three questions do most of the work, and they're the same ones you'd use to evaluate AI software vendors in any category.
Which model. Ask what's actually running under the "AI" label — a specific model name, or a vague "our proprietary technology." A vendor using GPT-4, Claude, or a named open-source model can tell you which one and roughly what it costs them to run. A vendor who can't name it is usually running a basic script with an AI label stapled on, or reselling someone else's API without knowing what's inside it.
Trained or prompted on what data. There's a real difference between a model trained on your industry's data and a general model handed a prompt at runtime. A vendor should be able to say which one, plainly. "Trained on years of data from businesses like yours" is a claim that needs a source behind it. If they can't name what the model actually learned from, assume it learned nothing specific to your business and is a general model wearing your logo.
What happens when it's wrong. This is the question vendors dodge hardest. Every model is wrong sometimes. Ask what the failure looks like: does it flag its own uncertainty, does a human review before anything reaches a customer, is there a log you can audit afterward. If the answer is "it just works," that's not an answer. That's a company that hasn't shipped enough volume to see the failure happen yet.
Ask these three questions on a sales call and watch the pace of the answer slow down. A vendor who built something real answers fast, because they've been asked before and know the tradeoffs. A vendor riding the label stalls, because there's nothing underneath the sentence they haven't said out loud yet.
What genuinely useful AI actually looks like inside an SMB tool
Strip the marketing and there's a plain test: does it do a task you can name, and can you check its work.
Task you can name means you can finish the sentence "it takes ___ and turns it into ___." Say a tool takes a customer's incoming email and drafts a reply using the last ten tickets from that account. That's a task. "It makes your business smarter" is not a task, it's a slogan.
Check its work means there's a way to verify the output without taking the vendor's word for it. Say the tool drafts the reply above. You should be able to open the ticket, see the source tickets it pulled from, and see the draft before it sends. If you can't see the sources and you can't stop it before it acts, you're not checking anything, you're hoping.
That combination — a named task plus a visible way to check it — is a low bar. Most software that deserves the "AI-powered" tag clears it without needing the phrase at all. The tools that need the label to sell are usually the ones that would fail this test if you asked them to demonstrate it live, on your own data, in the room.
The cheap fix before you buy anything
Before paying for an "AI-powered" upgrade, run the vendor's demo on a task from your own business, live, not on their canned example. If it's a drafting tool, feed it a real email from last week and see what it produces. If it's a scheduling tool, feed it your actual calendar conflicts, not their clean demo data. Say the whole test takes twenty minutes end to end. That's the entire cost of finding out whether the label matches the product, before you sign anything.
If the demo holds up on your own data, and the vendor answers the three questions above without stalling, the label was probably earned. If it doesn't hold up, you've saved yourself a subscription and a rollout that would have gone nowhere either way.
That's the whole method. Not a checklist that needs software of its own, just three questions and one live demo run on your own numbers. The vendors who built something real will not mind you asking.
