What business owners actually fear about AI isn’t job losses

When someone tells me they are worried about AI, I used to assume I knew what came next. Jobs. It usually was not.

Anthropic ran what I think is the largest study of its kind — reported as around 81,000 people interviewed one to one, across well over a hundred countries. The leading concern was not redundancy. It was unreliability. People do not trust that they know how to use these tools well enough to trust what comes back out of them.

Two honesty notes before I build anything on that. Firstly, I came to this study through Allie K Miller’s newsletter rather than the source, so treat the precise figures as directional until you read it yourself. Secondly, Anthropic makes these tools, which does not invalidate the research but is worth knowing when you weigh it.

Why that finding rings true

For a small business, a wrong answer is not an inconvenience. It is a price quoted that you cannot honour, an availability promised that you do not have, a claim made to a customer that turns out not to be true. You carry that personally, because your name is on the door. Of course the worry is reliability. The stakes of being wrong are higher for you than for a company with a legal department.

The same research reported something I found more interesting still: the people getting the most out of AI were also considerably more likely to worry about its risks. That reads like a contradiction and is not. It is what knowing something looks like. The person who has watched a model state a wrong figure with total confidence is exactly the person who now checks.

The useful question

So I would swap the question. Not “will this replace me”, which for most small firms is not the live issue — you probably cannot hire the people you need as it is. The better question is: can I check it?

That question has practical answers, and they are all design decisions rather than acts of faith.

• Narrow the scope. A system answering from your price list is checkable. A system answering from everything it has ever read is not.
• Name the sources. Every answer should be traceable to a document you control and could correct.
• Define what happens when it is unsure. A system that says “I do not know” is worth more than one that never does, because the confident one is guessing somewhere you cannot see.
• Keep a human on anything irreversible. Sending, promising, pricing, publishing. The draft is the machine’s job; the send is yours.
• Log it. Not so you read every line, but so you can pull ten at random on a Friday and satisfy yourself it is behaving.

That list is essentially our build philosophy written out, and it is why we hand the AI as little as we can get away with. The parts that must never wobble are ordinary code you can inspect.

Overall: trust is not a feeling you have to talk yourself into. It is a property you build in, or leave out. If you would rather see that demonstrated than described, our free audit shows its working for every number it gives you — including saying “unavailable” when it could not measure something.