How to Actually Use Claude in a Small Business

Most advice about AI for small business is written at a level of abstraction that leaves you no better off — transform your operations, unlock productivity, and so on. This is the practical version: specific tasks, honest limitations, and the handful of rules worth setting before your team starts using it, whether or not that's already happening.

One note on scope. We're writing about Claude because it's what we use and know well, but nearly all of this applies to the major assistants generally. The capabilities move fast enough that any specific claim about pricing or model names would be wrong within months, so we've deliberately left those out — check the vendor for current details.

What it's genuinely good at

The pattern worth internalising: it's strong where the work is language-shaped and you can check the output, and weak where being subtly wrong is expensive and hard to notice.

• Turning rough notes into something presentable — a quote, a scope of work, a follow-up email. You know what you want to say; the drafting is the slow part.
• Reading long documents and answering questions about them. A supplier contract, an insurance policy, a compliance document you've been avoiding.
• Rewriting for a different audience. The same explanation for a technical client and a non-technical one.
• First drafts of documentation nobody has had time to write — processes, onboarding notes, how a system actually works.
• Working through a spreadsheet formula or a data question without needing to already know the syntax.
• Being a thinking partner on a decision. Not for the answer, but for the questions you hadn't considered.

What to keep it away from

Being specific about the weaknesses matters more than the strengths, because that's where people get burned.

• Anything where a plausible-sounding wrong answer costs you. It can state incorrect things with complete confidence — the failure mode isn't hedging, it's fluency.
• Current facts it can't look up. Prices, availability, someone's current role, what a regulation says this month.
• Final legal, tax, or medical judgement. It's genuinely useful for understanding a document and preparing questions; it's not a substitute for the person who carries the liability.
• Anything customer-facing that goes out unread. A person checks it before it leaves the building — the same rule you'd apply to a new junior's drafts.
• Arithmetic you're relying on. Have it show the working, and check the working.

The data rule, before anything else

Set this before people start, because they very likely already have. Surveys consistently find most employees use AI tools their employer never approved — usually to do their job better, not to be reckless.

Write down what must not be pasted into a public AI tool: customer personal data, anything covered by a confidentiality agreement, credentials, financial account details, health or employment records, and source code for systems you depend on. Then write down what's fine — public marketing copy, general questions, drafting help with no identifying detail. Specific categories work; "use good judgement" does not.

How to get better results

Most disappointing output traces back to the same few things.

• Give it the context you'd give a new employee. Who it's for, what it's for, what tone, what constraints. "Write a quote" produces generic output because it's a generic request.
• Paste in the actual material rather than describing it. The real email, the real notes, the real document.
• Show it a good example of what you want. One previous quote you were happy with does more than three paragraphs of instruction.
• Ask it to ask you questions before it answers, on anything complicated. It'll surface what you left out.
• Iterate rather than restarting. "Make it shorter and less formal" is faster than rewriting the request.
• Tell it to say when it doesn't know. It'll do this if asked and often won't volunteer it.

A first week that isn't a project

Don't start with a strategy. Pick the single task you most dislike doing that's mostly writing — for a lot of owners that's quotes, follow-ups, or the documentation nobody's written. Use it for that one task for a week and see whether it actually saved time.

Then ask your team what they're already using and what for. You'll get a far more accurate picture from a conversation framed as "we want to make this properly available" than from any monitoring exercise, and the answers usually surface a couple of genuinely valuable uses you'd otherwise have banned by accident.

When the answer is software instead

Worth saying, since it cuts against the grain of most AI advice: sometimes the reason a task is painful is that it's a repetitive process that should have been automated properly years ago. If the same forty-minute job happens every week in exactly the same shape, a small piece of software will do it faster, cheaper, and identically every time — and won't need checking.

AI is the right tool when the work varies, involves judgement, or is language-shaped. Plain automation is the right tool when it doesn't. Reaching for the fashionable option regardless is how people end up paying more for a worse version of something a script could do.

The short version

Use it where a wrong answer is cheap and obvious. Keep a person between it and your customers. Set the data rules before people improvise them. Start with one task, not a strategy. And be honest about when what you actually need is a piece of software rather than a chatbot.

If you want to work out where this genuinely fits in your business — including hearing that it doesn't, in places — that's a conversation we're happy to have. We work with businesses across Middlesex, Somerset and Mercer County, and you can also read where AI actually helps a small business for the wider view.