Where AI actually helps a small business — and where it doesn't

There are two unhelpful positions on AI and most people are standing on one of them.

One says it changes everything and you're finished if you're not using it. The other says it's a toy that makes things up. Neither helps you decide whether to use it on a Tuesday morning.

We use it every day building things. Here's the honest split.

The question that decides it

Before anything else, one test:

What does it cost me if this is confidently wrong?

AI is very good at producing plausible work. It is not reliable at producing correct work. So the question isn't "can it do this" — it usually can. It's what happens when it does it badly and you don't notice.

Draft an email badly? You read it and fix it. Cost: nothing.

Get a figure wrong in a quote? You've underpriced a job for six weeks.

Everything below is that one question applied.

Where it genuinely helps

Starting things. The blank page is the expensive bit. "Write me a first draft of a letter chasing an unpaid invoice" gets you 80% of something in ten seconds, and you fix the rest. You were always going to rewrite it anyway.

Turning your words into tidy words. Say what you do into your phone, hand it the transcript, ask for a clean paragraph. This is genuinely useful for anyone who knows their trade but hates writing — and it's your content, just tidied.

Summarising and finding. A long document, a pile of emails, a thread you weren't on. "What did we agree about the delivery date?" Fast, and easy to verify because the source is right there.

Explaining things. "What does this clause mean in plain English?" is a legitimate use — as long as you treat the answer as a starting point rather than advice you'd act on.

Boring repetitive text. Reformatting, renaming, turning a mess into a list. Dull work it does without getting bored.

Where it quietly costs you

Anything with a number that matters. Prices, tax, measurements, deadlines. It will produce a figure that looks right. It has no idea whether it is right. Never take a number from it and put it in front of a customer.

Anything specific to Scotland. This one bites locally. Ask about business rates or small business relief and you'll often get the English answer delivered with total confidence, because most of what it learned from is English. Scotland sets its own rules. We built a Scottish business rates calculator precisely because the generic answers are wrong here.

Anything where it doesn't know what it doesn't know. It won't tell you that your process is broken, that the real problem is elsewhere, or that the question you asked isn't the question you needed answered. It answers what you typed.

Anything that needs to sound like you. AI has a register — smooth, slightly corporate, faintly American. Customers can feel it even when they can't name it. If your whole appeal is being the local one who talks straight, don't hand your voice to something that writes like a brochure.

Anything where being wrong is expensive and hard to spot. That's the dangerous combination. Wrong and obvious is fine. Wrong and plausible is what costs money.

What it hasn't changed

It hasn't made judgement cheap.

It'll write you a website. It won't tell you that your phone number should be tappable, that "joiner Perth" matters more than "bespoke joinery solutions", or that your real problem is a Google profile you never verified. We wrote about that in what a website should cost in the age of AI — the code got cheap, the knowing-what-to-build didn't.

Same in bigger businesses. AI will happily automate a process nobody ever mapped. Now the mess runs faster.

A sensible way to start

1. Pick something where being wrong is cheap — a first draft, a summary, tidying your own words
2. Always read it before it leaves your hands
3. Never let a number through unchecked
4. Never let it near anything Scotland-specific without verifying
5. Keep your own voice — use it to tidy what you said, not to replace what you'd say

Used that way it's a decent assistant who's fast, tireless, occasionally confidently wrong, and never says "I'm not sure".

Which is fine, as long as you remember the last two.