Somewhere in your inbox right now there is a sales email promising that a tool is "powered by AI." Maybe it is a scheduling app, a quoting system, a CRM add-on, or a chatbot for your website. The pitch is familiar: smarter, faster, learns from your business. The price is usually 30 to 60 per cent above the version that was sitting on the same website eighteen months ago under a different name.
Some of those claims are real. A good number are not. The practice has a name now — AI washing — and it is the software industry's version of greenwashing: slapping an AI label on a product whose underlying mechanics have not changed. For a business owner trying to spend a limited budget well, the skill worth learning is not "which AI tool is best." It is how to tell, in about twenty minutes, whether there is any AI in the box at all.
What AI washing actually is
AI washing sits on a spectrum, and it helps to know which part of it you are dealing with.
At the mild end is relabelling. A product that has used if-then rules for a decade — "if the invoice is 30 days overdue, send reminder template B" — gets a new badge reading "AI-powered automation." Nothing about the product got worse. It just got renamed, and often repriced.
In the middle is thin wrapping. The vendor has added an API call to a large language model, usually OpenAI's or Anthropic's, and put a text box in front of it. This is genuine AI, but it is AI you could access yourself for €20 a month, now resold at €149 with a logo on it. Sometimes the wrapper adds real value — your data, your templates, your workflow, proper permissions. Often it adds a chat window and nothing else.
At the far end is fabrication: claiming capabilities that do not exist. This is the one regulators have started prosecuting, and it is rarer than the first two — but it is the reason the other two now carry legal risk for the vendors doing them.
The VOC pattern we hear from owners over and over is exactly this confusion: "Is this actually AI, or just rebranded automation?" It is the right question. It just needs a repeatable way to answer it.
Why small businesses are the easiest target
Enterprise buyers have procurement teams, security reviews, and a technical architect who will ask the vendor which model they are calling. A ten-person company has none of that. The owner evaluates the tool between two client meetings, watches a three-minute demo video, and signs up for the annual plan because it saves 20 per cent.
That asymmetry is well understood by the people writing the sales pages. It shows up in three predictable ways:
- Demos built on cherry-picked data. The demo dataset is clean, small, and chosen because the tool handles it well. Your data is messy, large, and full of edge cases.
- Benefit language with no mechanism. "Our AI learns your business" is a benefit. "We fine-tune a classifier on your last 12 months of tickets" is a mechanism. Only one of those can be checked.
- Annual-only pricing on new AI tiers. If the AI feature were as transformative as claimed, the vendor would be happy to sell it monthly and let the results retain you.
None of this means the tool is bad. It means the burden of proof has quietly been moved onto you.
Seven questions that expose a hollow AI claim
Send these to the vendor by email, not on a call — you want the answers in writing, and you want to see how long they take to produce them. A vendor with real AI answers all seven in a day. A vendor without one tends to answer three and redirect.
- Which model or models does the feature use, and are they yours or a third party's? "We use GPT-4.1 via the OpenAI API" is a perfectly good answer. "Proprietary algorithms" is not an answer.
- What happens to my data? Is it used to train anything? A real AI vendor has a crisp, documented answer here because they have had to think about it. Vague reassurance is a red flag in itself.
- What is the accuracy rate, measured how, on what sample? You are not expecting a research paper. You are checking whether anyone has ever measured it.
- What does the tool do when it is not confident? Real systems have a fallback — escalate to a human, flag for review, return "unknown." Rule-based systems dressed as AI usually just produce a wrong answer confidently.
- Can I see it run on my data, not your demo data? This is the single most revealing question. Ask for a trial with 50 of your own records.
- What did this product cost before the AI tier existed? Compare against an archived version of their pricing page. A 300 per cent increase for a chat box is worth knowing about.
- Who is accountable if the output is wrong and it costs me money? Read the clause. Most AI terms disclaim everything, which is fine — but it should change how much you rely on the output.
Here is a prompt that does the first pass for you. Paste the vendor's marketing page and their terms of service into Claude or ChatGPT:
"Below is a software vendor's marketing page and terms of service. Identify every specific claim they make about artificial intelligence or machine learning. For each claim, classify it as (a) verifiable and specific, (b) vague marketing language, or (c) contradicted elsewhere in the document. Then list the five questions I should ask this vendor before buying. Be sceptical and do not give them the benefit of the doubt."
What the rules now say about AI claims
This is no longer only a buyer-beware problem. Regulators on both sides of the Atlantic have started treating exaggerated AI claims as straightforward misrepresentation, and enforcement actions against companies overstating their AI capabilities have been building since 2024.
In the EU, the AI Act adds a second layer: transparency obligations mean that where you are interacting with an AI system, that generally has to be disclosed. A vendor who is vague with you about whether there is AI inside may also be vague with your customers about it — and if you are the one deploying the tool to your customers, a share of that obligation lands on you, not on the vendor.
There is a practical consequence for SMBs. If you resell, embed, or build a client-facing service on top of a tool, you inherit its claims. "Our supplier told us it was AI" is a weak position if a client asks you to substantiate what you have been billing them for. Our guide to the EU AI Act for small businesses covers where those obligations start and stop.
A 20-minute test you can run before you buy
You do not need technical skills for this. You need twenty minutes and a willingness to be slightly awkward.
Minutes 1–5: the archive check. Look up the vendor's pricing page from 18 months ago on the Wayback Machine. Note what changed: the price, the feature names, or genuinely the product.
Minutes 5–10: the variation test. In the trial, give the tool two inputs that mean the same thing but are worded completely differently. A rules engine keys off specific words and will treat them as unrelated. A language model will handle both. This one test separates the mild end of the spectrum from the real thing faster than any sales call.
Minutes 10–15: the nonsense test. Feed it something it should not be able to handle — a record in another language, a deliberately contradictory entry, an empty field. Watch whether it degrades gracefully or produces confident nonsense.
Minutes 15–20: the cost-per-outcome sum. Divide the annual price by the number of times you realistically use the feature per year. A €1,200 tool used twice a month costs €50 per use. If a generalist assistant you already pay for can do the same job in four minutes, you have your answer. Our breakdown of what AI actually costs a small business has the full comparison, and if you suspect this is not the only subscription in question, auditing your AI tool stack is the natural next step.
When "it is just rules" is perfectly fine
Here is the part most sceptical articles get wrong. Discovering that a tool is rules-based rather than AI-driven is not, by itself, a reason to reject it.
Rules are predictable, cheap to run, easy to audit, and they do not hallucinate. For invoice reminders, appointment confirmations, stock alerts, and most of what a small business automates day to day, a well-built rules engine is the better technology. It will do the same thing on Tuesday that it did on Monday, which is exactly what you want from your billing.
The problem is never that a product uses rules. The problem is paying an AI premium for them, and building expectations — yours and your customers' — on a capability that is not there. Judge the tool on what it does for your workflow at the price you are paying. Make the label irrelevant.
That reframing also makes vendor selection much less anxious, because you stop asking "is this real AI?" and start asking "does this do the job, and is the price fair for what is actually inside?" If you want a structured version of that process, our walkthrough on how to choose an AI vendor turns it into a scorecard.
The bottom line
AI washing works because most buyers feel they are not technical enough to challenge the claim. You do not need to be. You need seven questions in an email, one archived pricing page, two deliberately awkward test inputs, and the confidence to divide the annual cost by the number of times you will actually use the thing.
The owners getting the most from AI in 2026 are not the ones who bought the most tools with AI in the name. They are the ones who worked out what job they needed doing, bought the cheapest thing that did it reliably, and spent the difference on getting their team to actually use it. Scepticism is not a cost here. It is most of the return.
Where does your business stand on AI?
Take the free 3-minute AI Readiness Quiz and get a personalised score with your next steps.
Take the Free Quiz →