AI for small business: where it actually pays

Most AI advice for small businesses is written by people selling AI. The useful version is narrower and considerably less exciting: a handful of tasks where it genuinely helps, several where it creates work, and one rule that sorts them.

The rule

It is good at tasks where you can check the answer faster than you could produce it. It is bad at everything else.

That single sentence predicts nearly every success and failure we have watched. Editing a mediocre draft is faster than facing a blank page, so drafting works. Verifying a price you did not already know is slower than looking it up, so asking for prices does not.

Where it pays

Drafting. Not final copy — first drafts of product descriptions, service pages, follow-up emails, job posts. Text you were going to write anyway and will edit regardless.

Summarising. Long email threads, meeting notes, a stack of customer reviews. Anywhere you need the gist of something long and can check the result against the source in seconds.

Triage. Sorting incoming enquiries by subject and apparent urgency so the right person sees them first. The cost of a mistake is low and self-correcting: a misrouted email gets forwarded.

Translation, which matters more here than in most cities. Good enough to communicate, and worth a human read before anything goes on a contract or a sign.

Where it does not

Anything where being confidently wrong is expensive. Prices, availability, medical or legal specifics, promises to customers. These systems produce fluent text whether or not they know the answer, and fluency reads as authority — to you as much as to the customer.

Anything requiring current facts about your business it has no way to know. If it has not been given today's stock or this week's hours, it will invent something plausible rather than say it cannot tell.

Anything you will not check. The entire saving comes from reviewing output faster than writing it from scratch. Remove the review and you have not saved time — you have automated the production of errors and moved the discovery of them to your customers.

What it costs, honestly

Subscriptions are cheap — usually less than a phone line per person. The real costs are elsewhere.

The learning curve: a few weeks before anyone is faster rather than slower, and some people never get there, which is a fact about tools and not about them.

The review time, which is permanent and should be counted as part of the task rather than treated as overhead.

And the temptation to use it where it does not belong, which is the one that actually costs money. The failure is rarely the tool. It is a person under time pressure skipping the check.

What a month of this actually looks like

Take product descriptions, since almost every business has a backlog of them.

Week one is slower than doing it by hand. You are learning what to ask for, and the first outputs are generic in a way that takes as long to fix as writing would have. Most people who give up, give up here.

Week two changes when you stop asking for "a product description" and start giving it the raw material — the spec sheet, two descriptions you already like, and the thing that makes this product different. The output stops being generic because you stopped withholding the specifics.

By week four you have a shape that works and you are editing rather than rewriting. The saving is real but smaller than the marketing suggests: perhaps half the time on a task you were doing anyway, not the elimination of it.

That is the honest number, and it is still worth having. A business with two hundred products and no descriptions can now have two hundred descriptions.

Where your information goes

Worth settling before it becomes a question rather than after.

Business and paid tiers generally do not train on what you send; free consumer tiers sometimes do. That distinction is in the terms, it changes, and it is the one thing to check rather than assume — particularly if anyone on your team is using a personal account for work.

The practical rule is simpler than the policy. Do not paste anything you would not email to a supplier: customer records, anything covered by a confidentiality agreement, card or health details. Not because a leak is likely, but because there is rarely a reason to take the risk for the sake of a draft.

Tell your team the rule out loud. Most exposure is not a decision anyone made; it is somebody helpful pasting a spreadsheet.

How to tell whether it is actually saving time

The feeling of speed and the fact of it come apart easily here, and the gap is worth measuring rather than assuming.

Pick one recurring task and time it honestly for a week without the tool — not the writing alone, but the whole thing, including the thinking beforehand. Then time it with the tool, and count the checking and correcting as part of the total, because it is. A first draft produced in ten seconds that takes fifteen minutes to make usable has not saved anything on a task you could have written in twelve.

What this usually reveals is that the wins are lopsided rather than general. It is dramatic on long, formulaic, low-stakes writing, and roughly neutral on short things you already know how to say. Businesses that get real value have found the two or three tasks where it is dramatic and left the rest alone.

The trap is the opposite: using it everywhere, feeling faster, and quietly spending the saved time on review.

A reasonable place to start

Pick one repetitive writing task. Use it there for a month. Keep a note of what you had to fix each time.

That list tells you more about where this fits in your business than any vendor demonstration, and it is specific to you. If the fixes are small and shrinking, expand. If you are rewriting from scratch every time, stop — that task is not a fit, and no amount of prompt advice will change it.

The businesses getting real value are not the ones using it for the most things. They are the ones who found two or three tasks it does well and left the rest alone.