There is a label on nearly every piece of software sold today, and it has quietly stopped carrying any information. "AI-powered" now appears on products containing a genuine machine learning model trained on relevant data, on products that make a single call to someone else's language model when you click a button, and on products that are the same rules engine they were three years ago with a new word on the box. All three say "AI-powered." All three are telling the truth by some reading. And a buyer who cannot tell them apart is negotiating in the dark.

This matters because the label is doing real work in purchasing decisions. It commands a price premium, it shapes which vendor looks more advanced, and it influences a board that has been told to prioritize AI. When a word that spans that much range is treated as if it meant one thing, buyers systematically overpay for the weak version and cannot recognize the strong one. The remedy is not cynicism about AI. It is learning to decode the claim, which is a skill, and a short one.

Regulators have already noticed

Before the decoding, it is worth knowing that this is not a fringe concern or a matter of marketing taste. It has become an enforcement category.

The practice even has an official name. Regulators call it "AI washing," modeled on greenwashing, and they have started acting on it. The SEC brought its first cases in March 2024 against two investment advisers, Delphia and Global Predictions, which settled for a combined $400,000 over false and misleading statements about their use of AI. The FTC launched Operation AI Comply in September 2024, an enforcement initiative aimed at deceptive AI claims, and the SEC extended the theory to a public company, Presto Automation, in January 2025. The Department of Justice has pursued criminal fraud charges in the space as well.

The detail most relevant to a buyer is what these cases had in common. As one analysis put it, these were generally not companies with no AI at all. They had some technology, some data processing, sometimes even machine learning models, and they oversold what those systems could actually do. That is the exact gap this piece is about. The problem is rarely a pure fabrication that a demo would expose. It is the space between what "AI-powered" implies and what the system actually does, and that space is where both the regulatory risk and the buyer's overpayment live.

The four things the label can mean

In practice, "AI-powered" in business software refers to at least four arrangements, and they differ enormously in what they can deliver for you.

• A genuine model trained on relevant data: The product contains a machine learning system built for a specific task, trained on data that resembles yours, that learns and produces outputs no rules engine could. This is the strong version, the one the label is supposed to imply, and it is the least common.

• A wrapper around a general model: The product calls a general-purpose language model through an API when you take an action, passing your text in and returning the response. This can be genuinely useful for drafting, summarizing, and classifying, but it is not proprietary, not specific to your domain, and not something the vendor built. You are paying a markup for a call you could nearly make yourself, and every competitor has access to the identical capability.

• Statistics with a new name: Techniques that existed for decades, regression, forecasting, thresholds, anomaly rules, relabeled as AI because the term now sells. Often perfectly good techniques, and often the right tool, but not what a buyer imagines when they hear the word, and not deserving of a premium the word commands.

• A rules engine wearing a costume: Deterministic if-then logic, hand-coded, that does exactly what it was programmed to do and nothing more, described as AI because the output looks smart. This is the weakest version and, in the enforcement cases, the one closest to the line.

A vendor can say "AI-powered" while meaning any of these, which is precisely why the label conveys nothing on its own. The word is the same. The thing behind it is not.

The questions that decode it

The useful move is not to become technical. It is to ask a small number of questions whose answers separate the four, and to notice how the vendor responds, because vagueness is itself an answer.

• Is the model trained on data like mine, or is it general?
  A model trained on property operations data will behave differently on your problem than a general model prompted about property. If the answer is that it uses a leading general model, you have learned it is a wrapper, which is fine for some tasks and should be priced as a commodity rather than a differentiator.

• What specifically does the AI do that a rules engine could not?

  This question separates genuine learning from relabeled logic instantly. If the honest answer, once you strip the adjectives, is "it applies rules we defined," you have found statistics or a rules engine, and the word "AI" is decoration.

• Does it improve with use, and how would I see that?

  Learning systems get better as they see more data; rules engines and single API calls do not. Asking how improvement would show up, and how you would measure it, tends to end the conversation quickly where there is no learning involved.

• When it produces an output, can you show me why?
  A vendor with a real, understood system can describe how it reaches a result at some level. A vendor who cannot, or who retreats to "it's proprietary" for everything, is either hiding that there is less there than claimed or has a genuine explainability problem, and both are things a buyer needs to surface before signing rather than after. This connects to a requirement that matters independently for anything touching your numbers, covered in why an unexplainable AI is a problem.

The pattern to watch across all four is not the specific answer but the precision of it. A vendor with a real capability can answer these cleanly, because they know what they built. A vendor relying on the label gets vaguer as you get more specific, and that gradient is the most reliable signal available.

The honest part

Several qualifications keep this from tipping into reflexive suspicion, which would cost you real value.

A wrapper around a general model is not a scam, and dismissing it as one is a mistake. For a large class of tasks, drafting, summarizing, classifying, a well-built product on top of a general model is exactly the right tool, delivers real value, and there is nothing wrong with a vendor charging for the engineering, integration, and interface they built around it. The point is not that wrappers are bad. It is that you should know you are buying one, so you can judge the price against a capability your competitors can also buy, rather than paying a proprietary-technology premium for a commodity.

Older techniques relabeled as AI are frequently the correct choice, too. A well-understood forecasting method is often more reliable, more explainable, and more appropriate than a fashionable model, and a vendor using it is not deceiving you unless they are charging for sophistication that is not there. Simple and proven beats complex and trendy more often than the AI conversation admits, and the goal of decoding the label is not to demand the fanciest option but to pay the right price for whatever the thing actually is.

And most vendors are not committing fraud. Marketing enthusiasm, imprecise language, and genuine belief account for the vast majority of inflated AI claims, not deliberate deception. The decoding questions are not an accusation. They are how a buyer establishes what they are actually purchasing, which is their job regardless of the vendor's intent.

Buy the capability, not the word

The discipline reduces to refusing to let the label carry weight it cannot support. "AI-powered" tells you nothing, so treat it as a prompt to ask what the AI actually is, what it does that a simpler system could not, whether it is specific to your domain or a general capability everyone can buy, and whether the vendor can explain it when pressed. Then price it as what it turns out to be: a genuine trained model is worth a premium, a wrapper is worth the commodity it wraps plus the engineering around it, and a relabeled rules engine is worth exactly what that rules engine was worth before it was renamed.

This is the same decoding discipline that applies to real-time and to every other term that gets bought on faith, examined more generally in real-time is the most abused word in property software. The word on the slide is a claim, not a specification, and the gap between the two is where money is lost. Regulators are now policing the extreme end of that gap. A buyer who asks four precise questions is protecting themselves from the rest of it, which is far larger and entirely their own responsibility to catch.

FAQs

Q1. Doesn't "AI-powered" have a clear meaning?
Not anymore. It is applied to genuine machine learning models trained on relevant data, to products that make a single call to a general language model, to decades-old statistical techniques relabeled as AI, and to conventional rules engines described as AI because the output looks intelligent. All four use the same phrase truthfully by some reading, which is why the label conveys no information on its own and has to be decoded.

Q2. What is "AI washing"?
It is the practice, named by regulators after greenwashing, of making false, misleading, or exaggerated claims about a product's AI capabilities. The SEC, FTC, and DOJ have all pursued enforcement actions since 2024. Notably, most charged companies were not fabricating AI entirely, they had some technology and oversold what it could do, which is the same gap buyers need to detect between the label's implication and the system's reality.

Q3. What is the difference between a real model and a wrapper?
A genuine model is a machine learning system trained for a specific task on data resembling yours, producing outputs a rules engine could not. A wrapper calls a general-purpose language model through an API when you act, which can be useful but is not proprietary, not domain-specific, and available identically to every competitor. Both can say "AI-powered," but they warrant very different prices.

Q4. Is a wrapper around a general model a bad thing?
No. For drafting, summarizing, and classifying, a well-built product on top of a general model is often exactly right and delivers real value, and charging for the surrounding engineering is legitimate. The issue is only knowing that is what you are buying, so you price it against a capability competitors can also purchase rather than paying a proprietary-technology premium for a commodity.

Q5. How can I tell what a vendor actually has?
Ask whether the model is trained on data like yours or is general, what specifically the AI does that a rules engine could not, whether it improves with use and how you would see that, and whether the vendor can explain how it reaches an output. The specific answers matter, but the more reliable signal is precision: real capability answers cleanly, while reliance on the label gets vaguer as your questions get more specific.

Q6. Are older techniques relabeled as AI a problem?
Only if the vendor charges for sophistication that is not there. A well-understood forecasting or statistical method is frequently more reliable, more explainable, and more appropriate than a trendy model, and using it is not deceptive. Simple and proven often beats complex and fashionable, so the goal is to pay the correct price for what the technique actually is, not to demand the fanciest option.

Q7. Does this mean most AI vendors are being dishonest?
No. The large majority of inflated claims come from marketing enthusiasm, imprecise language, and genuine belief rather than deliberate fraud. The decoding questions are not an accusation, they are how a buyer establishes what they are actually purchasing. That responsibility exists regardless of the vendor's intent, and asking clearly tends to produce clearer answers from honest vendors too.

Q8. What should drive the price I pay?
What the capability actually is, once decoded, rather than the label. A genuine trained model specific to your domain justifies a premium. A wrapper is worth the commodity capability it provides plus the engineering built around it. A relabeled rules engine is worth what that rules engine was worth before the rename. Matching price to the real thing is the entire point of asking.