What AI is actually good at in a small business

You're writing a welcome pack for guests at your holiday let, so you ask AI: "Recommend a good independent café near Market Street." It obliges (an example written for this post, not a real model run):

"Tucked just off Market Street, The Copper Kettle is a local favourite, known for its sourdough toasties and cosy armchairs."

It's warm, specific and a confident review of a café that closed in the spring.

So what is AI good at? It's strong at reshaping material you hand it: first drafts, summaries, rewording and turning notes into structure. It lets you down when it must supply what it can't see, such as local or recent facts and judgement calls about your clients.

What is AI good at in a small business?

The jobs AI does well have something in common. You give it the raw material, and it reshapes it into something more useful.

• First drafts: getting from a blank page to something you can react to, such as a service description, a welcome email or an outline for a talk
• Summaries: pulling the actions out of call notes, a long email thread or a report you don't have time to read twice
• Rewording: making something shorter, plainer, warmer or more formal, or rewriting it for a different reader
• Structure: turning scribbled notes into a tidy agenda, a checklist or a set of headings

None of these needs the AI to know anything about your business that you haven't told it. The facts are in front of it, and its job is to rearrange them well.

What is AI bad at?

The trouble starts when the AI has to fill a gap from its own general knowledge. A model learns from a large body of writing up to a point in time, which is its knowledge cutoff. Unless your tool can search the web, anything after that is guesswork.

So it struggles with facts that are recent, niche or local. It knows plenty about cafés in general and very little about the one on your corner. It also can't see your inbox, your diary or your client history unless you paste them in.

And it can't make judgement calls for you. Which client to chase first, whether to raise a difficult point, how a long-standing customer will take a change: those depend on things only you know.

Why it lets you down in such a confident voice

The café review didn't come with a warning, and that is the part that catches people out. AI writes by producing the most likely next words, and the most likely words for a café review are confident ones.

There is no separate signal in the writing for "I'm not sure about this". A real café and an invented one come out in the same tone, so the tone tells you nothing. That's why choosing the right job matters more than spotting the wrong answer afterwards.

The job test: is the material in the chat?

Before you hand AI a task, ask whether the facts it needs are already in front of it. If they are, it's reshaping, and you can expect a useful first pass. If they aren't, it's supplying, and you should expect gaps filled with plausible guesses.

Often you can move a task from supplying to reshaping by giving it the material. Instead of asking for a café recommendation, write two lines about the place you actually like and ask AI to make them read well. Your notes might say "Bea's on Mill Lane, good coffee, dogs welcome, five minutes' walk", and the reply (an example written for this post) might be:

"For good coffee, try Bea's on Mill Lane, a five-minute walk away. Dogs are welcome too."

The facts stay yours and the polish comes from the tool.

The job test: whose judgement is it?

The second question is about consequences. If the task involves a decision about a real person, such as a client, a supplier or someone on your team, the judgement stays with you.

That doesn't mean AI is no use there. It can lay out options, list what you might be missing or help you word a tricky message. But it can't weigh what it doesn't know, so treat what it gives you as input to your decision rather than the decision itself.

A job that works, and a job that fails

Here is a reshaping job. You paste in rough notes from a discovery call: "wants help with onboarding, current process is email back and forth, loses track, wants something simple, starts in January, worried about cost, partner handles invoices". You ask for a short follow-up email confirming what you heard. Here is the kind of reply you'd get (an example written for this post):

"Thanks for the call today. To check I've understood: your onboarding currently runs on back and forth emails, and it's easy to lose track. You'd like something simpler in place for January, cost is a concern, and your partner looks after invoices."

Every fact in that reply came from your notes. It needs a quick read, but it's ready to send after a few small edits.

Now a supplying job: "Which of my clients is most likely to renew this year?" With nothing pasted in, the AI has nothing to go on.

It may give you a sensible-sounding list of signs to look for, or it may simply guess. Either way, the answer is about clients in general, not about yours.

Before you ask, check whether you're asking AI to reshape what you have or to supply what you don't.

What the job test won't fix

Giving AI the right job doesn't make it infallible. A reshaping task can still go wrong if it misreads your notes, drops an important detail or smooths over something you meant to keep. Read the result against what you gave it.

The test also won't tell you how to word the request itself, or what context about your business to give it. Those are separate skills, and they matter once you've picked the right job.

One thing to do today

Look at the next task you were going to give AI and ask the job test questions. If it's a supplying job, find the material it needs and paste that in first, or keep the task for yourself.

To see where AI could help your business most, take the AI Readiness Check: it's a short check with no account needed.