What AI actually does for a ten-person business

A de-hyped look at the three things AI is genuinely good at in a small business, the things it is not, and how to trial it in a fortnight.

Everyone in your inbox is suddenly an AI expert. The software reps, the LinkedIn posters, the bloke at the trade show who “used to do IT”. They all want to tell you AI will transform your business. Most of them are selling something, and most of what they’re selling won’t survive contact with a Tuesday morning at a ten-person firm.

So let’s skip the transformation talk. Here is what AI actually changes when you run a small business, described plainly, with the bits that don’t work included.

The three things it’s genuinely good at

Strip away the hype and useful AI in a small business collapses into three jobs. That’s it. Once you can see them clearly, you can spot where they apply in your own week.

1. Turning messy input into structured output

You already produce a lot of unstructured information. A voice note on the way back from a job. A rambling email from a supplier. A photo of a receipt stuffed in a glovebox. None of that is usable until someone — usually you, usually late in the evening — turns it into a job sheet, a CRM entry, or a line in your accounts.

This is the job AI is quietly best at. Feed it a voice note and get back a structured job sheet: customer, address, what was agreed, what’s needed next. Feed it an inbox full of enquiry emails and get back consistent records instead of half-remembered details. Feed it a stack of receipts and get back categorised bookkeeping lines instead of a shoebox.

None of this is glamorous. It’s also the difference between admin eating your evenings and admin taking ten minutes.

2. Drafting that a human approves

Quotes, replies, proposals, job descriptions, social posts — these all start from a blank page, and the blank page is where most people lose the most time. AI is good at getting you from nothing to a first draft. It is not good at deciding the draft is right.

The pattern that works: AI writes the first version, using your own past quotes, your tone, your pricing structure as reference. You read it, fix the bits it got wrong, and send it. You are still the last pair of eyes on everything that leaves the building. The tool just moved you from a blank page to an edit job, which is a much smaller task.

3. Answering questions over your own documents

Every small business has a pile of internal knowledge nobody can find fast enough: the price list from eighteen months ago, the policy document nobody’s read since it was written, the spec sheet for a product you only sell twice a year. AI can sit on top of that pile and answer questions against it — “what’s our returns policy for trade customers?” — instead of you digging through folders or ringing someone who might be on holiday.

This only works if the underlying documents are accurate. Ask it a question your documents don’t answer and it will still try to answer — confidently, and sometimes wrongly. More on that below.

What it’s not good at unsupervised

Here’s the part the sales pitches skip. AI is confident. It is confident when it’s right and exactly as confident when it’s wrong, and it gives you no reliable signal for telling the two apart. That’s fine for a first draft you’re going to check. It’s a genuine liability anywhere being confidently wrong costs you money or trust.

Keep AI away from unsupervised control of:

• Final pricing. It doesn’t know your margins, your capacity this month, or why you’re cutting this particular customer a bit of slack.
• Legal or tax advice. It will produce something that sounds authoritative regardless of whether it’s correct, and you won’t know which one you got.
• Anything sent to a customer without a human reading it first. An email, a quote, a social reply — once it’s sent, it’s said. A drafting tool that skips the approval step isn’t saving you time, it’s removing your last chance to catch a mistake.

The rule of thumb: the more it costs you to be wrong, the more a human needs to be in the loop before anything goes out.

Where the time actually goes

Here’s the honest bit nobody selling AI tools wants to say out loud: the time it saves only counts if you do something with it.

If AI turns a forty-minute admin job into a ten-minute one and you spend the other thirty minutes scrolling your phone, nothing in your business has changed. The saving is real, but it’s evaporated. It only shows up on your bottom line, or in your evenings back, if that time gets redeployed — onto the next job, onto actually speaking to customers, onto going home on time.

Before you adopt anything, it’s worth deciding in advance what you’ll do with the time it frees up. That decision matters more than the tool.

How to trial it in two weeks, without buying anything big

You don’t need a platform, a contract, or a consultant to find out whether any of this is useful to you. You need two weeks and one habit.

1. Pick one task from the list above — not three, one. The messiest, most repetitive one. Voice notes into job sheets is a good place to start for most trades and service businesses.
2. Use an off-the-shelf AI tool you can access today to do that one task, manually, for two weeks. Copy and paste is fine. You’re testing the workflow, not building infrastructure.
3. Keep the human check in place every single time. Don’t let anything skip the review step during the trial — the point is to learn where it goes wrong, not to find out the hard way with a customer.
4. At the end of two weeks, ask honestly: did this save real time, and did that time go anywhere useful? If yes, it’s worth setting up properly. If no, you’ve lost nothing and you know something true about your business that you didn’t know before.

That’s the whole test. No new software licence, no commitment, no transformation. Just a clear look at whether one piece of your week gets easier — and proof, one way or the other, before you spend anything on making it permanent.