Somebody is pitching you AI. Here is how to tell if it is real.

The room you are sitting in

Someone has a deck. There is a slide with a diagram on it, and the diagram has your business in the middle. They are describing something that will handle your enquiries, or your invoices, or your stock, and it sounds plausible, and you cannot tell whether it is.

That is not a gap in your knowledge. It is the normal condition of buying something technical, and it is the same feeling you would have buying a commercial oven or a delivery van if you had never bought one before. The difference is that with an oven you know to ask what it costs to run. With AI, nobody has told you what the equivalent question is.

Here is the useful starting point, and it will sound like it is against our own interests. Most small businesses that have not adopted AI have not adopted it because they see no use for it in their business: 77% of non-adopters give that as the reason, in a U.S. Chamber of Commerce survey of 3,870 small businesses fielded in June 2025 and reported by the JPMorganChase Institute on 14 April 2026. Not cost. Not fear of the technology. They looked and could not see where it fits.

A great many of them are right. If that is you, the honest answer is that you should not buy anything, and no amount of pitching should change that. What follows is how to find out which one you are.

The four questions

None of these require you to know anything about how the technology works. They are all questions about your business, which you are already the world expert on.

1. Which specific job does this do, and who does it today?

Not “it improves customer service”. Which job. Whose desk it comes off. If the person selling to you cannot name a task and a person, they have not looked at your business, and everything after that is a guess dressed as a plan.

2. What happens the first time it is wrong?

It will be wrong. Everything is wrong sometimes, including the person doing the job now. The question is what the wrongness costs and who catches it. A tool that drafts a reply for someone to approve is a different risk from one that sends the reply. A tool that suggests a price is different from one that sets it. If the answer is that it will not be wrong, stop, because that answer is never true.

3. What does it cost when it works, and what does it cost when we stop?

The monthly fee is the easy half. The other half is what you have to keep doing to keep it working, and what happens to your data and your process if you switch it off in a year. If nobody has told you what leaving looks like, you have not been told the price.

4. Show me it running on my own numbers.

Not a demo with sample data. Your enquiries, your invoices, your stock list. If that is not possible yet, the honest version is “we cannot do that yet and here is what we would need”. It is the request that separates people who have built the thing from people who have described it.

Four questions, and not one of them is about the technology. That is deliberate. You are not buying a technology, you are buying a change to how a job gets done, and you already know how the job gets done.

What a good answer sounds like

A good answer is specific, and it is usually a bit disappointing. It names one job rather than a category. It admits a failure mode without being pushed. It has a number attached that came from somewhere you could check.

That last part is not a preference of ours, it is the top-ranked thing owners say they need. 87% of small-business respondents said being able to verify the accuracy of data sources was very to extremely important before they would trust an AI tool, in a Paychex survey of more than 1,100 owners and HR professionals published on 11 March 2025. Verification came out ahead of everything else. So when a number appears in a pitch, the correct response is not to be impressed or unimpressed. It is to ask where it came from, and to notice how the room changes when you do.

A bad answer is expansive. It widens when questioned instead of narrowing. It moves from your business to the industry, from the industry to the technology, from the technology to the future. Watch for the direction of travel: good answers get smaller and more concrete under pressure, bad ones get bigger and vaguer.

A note on jargon

You will be told that you need to understand some terms. You do not, and there is evidence that trying is actively counterproductive.

In an experiment with 650 adults reading science and technology passages, jargon reduced how fluently readers processed the material and how much they engaged with it — even when every term was defined for them (Shulman, Dixon, Bullock and Colon Amill, Journal of Language and Social Psychology, 2020). The definition does not repair the damage. The word itself is the damage.

Which means a pitch that needs a glossary has a writing problem, not an audience problem. Anything genuinely worth doing in your business can be described in the words you already use for that business. If it cannot, that is information about the pitch.

The one that actually decides it

Of the four, question one does most of the work. Almost every AI project that goes wrong goes wrong because it was bought as a capability rather than aimed at a job. Somebody buys a thing that can do many things, and then nobody in the business is responsible for making it do any of them, and eleven months later it is a line on a card statement.

So before you compare tools, do the smaller and more useful thing: pick the one job. The one that eats a person’s week, or the one that keeps producing the mistake you are tired of apologising for. Write it on one line. Then everything else becomes a question you can actually answer, including whether to spend anything at all.

The next chapter puts a number on what that one job is costing you, using your figures rather than ours. It runs in your browser and it sends us nothing.